oom: badness heuristic rewrite
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
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.
17  *
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.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.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>
44 #include <linux/fs.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 "internal.h"
51
52 #include <asm/uaccess.h>
53
54 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
55 #define MEM_CGROUP_RECLAIM_RETRIES      5
56 struct mem_cgroup *root_mem_cgroup __read_mostly;
57
58 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
59 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
60 int do_swap_account __read_mostly;
61 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
62 #else
63 #define do_swap_account         (0)
64 #endif
65
66 /*
67  * Per memcg event counter is incremented at every pagein/pageout. This counter
68  * is used for trigger some periodic events. This is straightforward and better
69  * than using jiffies etc. to handle periodic memcg event.
70  *
71  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
72  */
73 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
74 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
75
76 /*
77  * Statistics for memory cgroup.
78  */
79 enum mem_cgroup_stat_index {
80         /*
81          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
82          */
83         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
84         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
85         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
86         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
87         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
88         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89         MEM_CGROUP_EVENTS,      /* incremented at every  pagein/pageout */
90
91         MEM_CGROUP_STAT_NSTATS,
92 };
93
94 struct mem_cgroup_stat_cpu {
95         s64 count[MEM_CGROUP_STAT_NSTATS];
96 };
97
98 /*
99  * per-zone information in memory controller.
100  */
101 struct mem_cgroup_per_zone {
102         /*
103          * spin_lock to protect the per cgroup LRU
104          */
105         struct list_head        lists[NR_LRU_LISTS];
106         unsigned long           count[NR_LRU_LISTS];
107
108         struct zone_reclaim_stat reclaim_stat;
109         struct rb_node          tree_node;      /* RB tree node */
110         unsigned long long      usage_in_excess;/* Set to the value by which */
111                                                 /* the soft limit is exceeded*/
112         bool                    on_tree;
113         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
114                                                 /* use container_of        */
115 };
116 /* Macro for accessing counter */
117 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
118
119 struct mem_cgroup_per_node {
120         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
121 };
122
123 struct mem_cgroup_lru_info {
124         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
125 };
126
127 /*
128  * Cgroups above their limits are maintained in a RB-Tree, independent of
129  * their hierarchy representation
130  */
131
132 struct mem_cgroup_tree_per_zone {
133         struct rb_root rb_root;
134         spinlock_t lock;
135 };
136
137 struct mem_cgroup_tree_per_node {
138         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
139 };
140
141 struct mem_cgroup_tree {
142         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
143 };
144
145 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
146
147 struct mem_cgroup_threshold {
148         struct eventfd_ctx *eventfd;
149         u64 threshold;
150 };
151
152 /* For threshold */
153 struct mem_cgroup_threshold_ary {
154         /* An array index points to threshold just below usage. */
155         int current_threshold;
156         /* Size of entries[] */
157         unsigned int size;
158         /* Array of thresholds */
159         struct mem_cgroup_threshold entries[0];
160 };
161
162 struct mem_cgroup_thresholds {
163         /* Primary thresholds array */
164         struct mem_cgroup_threshold_ary *primary;
165         /*
166          * Spare threshold array.
167          * This is needed to make mem_cgroup_unregister_event() "never fail".
168          * It must be able to store at least primary->size - 1 entries.
169          */
170         struct mem_cgroup_threshold_ary *spare;
171 };
172
173 /* for OOM */
174 struct mem_cgroup_eventfd_list {
175         struct list_head list;
176         struct eventfd_ctx *eventfd;
177 };
178
179 static void mem_cgroup_threshold(struct mem_cgroup *mem);
180 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
181
182 /*
183  * The memory controller data structure. The memory controller controls both
184  * page cache and RSS per cgroup. We would eventually like to provide
185  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
186  * to help the administrator determine what knobs to tune.
187  *
188  * TODO: Add a water mark for the memory controller. Reclaim will begin when
189  * we hit the water mark. May be even add a low water mark, such that
190  * no reclaim occurs from a cgroup at it's low water mark, this is
191  * a feature that will be implemented much later in the future.
192  */
193 struct mem_cgroup {
194         struct cgroup_subsys_state css;
195         /*
196          * the counter to account for memory usage
197          */
198         struct res_counter res;
199         /*
200          * the counter to account for mem+swap usage.
201          */
202         struct res_counter memsw;
203         /*
204          * Per cgroup active and inactive list, similar to the
205          * per zone LRU lists.
206          */
207         struct mem_cgroup_lru_info info;
208
209         /*
210           protect against reclaim related member.
211         */
212         spinlock_t reclaim_param_lock;
213
214         /*
215          * While reclaiming in a hierarchy, we cache the last child we
216          * reclaimed from.
217          */
218         int last_scanned_child;
219         /*
220          * Should the accounting and control be hierarchical, per subtree?
221          */
222         bool use_hierarchy;
223         atomic_t        oom_lock;
224         atomic_t        refcnt;
225
226         unsigned int    swappiness;
227         /* OOM-Killer disable */
228         int             oom_kill_disable;
229
230         /* set when res.limit == memsw.limit */
231         bool            memsw_is_minimum;
232
233         /* protect arrays of thresholds */
234         struct mutex thresholds_lock;
235
236         /* thresholds for memory usage. RCU-protected */
237         struct mem_cgroup_thresholds thresholds;
238
239         /* thresholds for mem+swap usage. RCU-protected */
240         struct mem_cgroup_thresholds memsw_thresholds;
241
242         /* For oom notifier event fd */
243         struct list_head oom_notify;
244
245         /*
246          * Should we move charges of a task when a task is moved into this
247          * mem_cgroup ? And what type of charges should we move ?
248          */
249         unsigned long   move_charge_at_immigrate;
250         /*
251          * percpu counter.
252          */
253         struct mem_cgroup_stat_cpu *stat;
254 };
255
256 /* Stuffs for move charges at task migration. */
257 /*
258  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
259  * left-shifted bitmap of these types.
260  */
261 enum move_type {
262         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
263         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
264         NR_MOVE_TYPE,
265 };
266
267 /* "mc" and its members are protected by cgroup_mutex */
268 static struct move_charge_struct {
269         struct mem_cgroup *from;
270         struct mem_cgroup *to;
271         unsigned long precharge;
272         unsigned long moved_charge;
273         unsigned long moved_swap;
274         struct task_struct *moving_task;        /* a task moving charges */
275         wait_queue_head_t waitq;                /* a waitq for other context */
276 } mc = {
277         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
278 };
279
280 static bool move_anon(void)
281 {
282         return test_bit(MOVE_CHARGE_TYPE_ANON,
283                                         &mc.to->move_charge_at_immigrate);
284 }
285
286 static bool move_file(void)
287 {
288         return test_bit(MOVE_CHARGE_TYPE_FILE,
289                                         &mc.to->move_charge_at_immigrate);
290 }
291
292 /*
293  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
294  * limit reclaim to prevent infinite loops, if they ever occur.
295  */
296 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
297 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
298
299 enum charge_type {
300         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
301         MEM_CGROUP_CHARGE_TYPE_MAPPED,
302         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
303         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
304         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
305         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
306         NR_CHARGE_TYPE,
307 };
308
309 /* only for here (for easy reading.) */
310 #define PCGF_CACHE      (1UL << PCG_CACHE)
311 #define PCGF_USED       (1UL << PCG_USED)
312 #define PCGF_LOCK       (1UL << PCG_LOCK)
313 /* Not used, but added here for completeness */
314 #define PCGF_ACCT       (1UL << PCG_ACCT)
315
316 /* for encoding cft->private value on file */
317 #define _MEM                    (0)
318 #define _MEMSWAP                (1)
319 #define _OOM_TYPE               (2)
320 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
321 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
322 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
323 /* Used for OOM nofiier */
324 #define OOM_CONTROL             (0)
325
326 /*
327  * Reclaim flags for mem_cgroup_hierarchical_reclaim
328  */
329 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
330 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
331 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
332 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
333 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
334 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
335
336 static void mem_cgroup_get(struct mem_cgroup *mem);
337 static void mem_cgroup_put(struct mem_cgroup *mem);
338 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
339 static void drain_all_stock_async(void);
340
341 static struct mem_cgroup_per_zone *
342 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
343 {
344         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
345 }
346
347 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
348 {
349         return &mem->css;
350 }
351
352 static struct mem_cgroup_per_zone *
353 page_cgroup_zoneinfo(struct page_cgroup *pc)
354 {
355         struct mem_cgroup *mem = pc->mem_cgroup;
356         int nid = page_cgroup_nid(pc);
357         int zid = page_cgroup_zid(pc);
358
359         if (!mem)
360                 return NULL;
361
362         return mem_cgroup_zoneinfo(mem, nid, zid);
363 }
364
365 static struct mem_cgroup_tree_per_zone *
366 soft_limit_tree_node_zone(int nid, int zid)
367 {
368         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
369 }
370
371 static struct mem_cgroup_tree_per_zone *
372 soft_limit_tree_from_page(struct page *page)
373 {
374         int nid = page_to_nid(page);
375         int zid = page_zonenum(page);
376
377         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
378 }
379
380 static void
381 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
382                                 struct mem_cgroup_per_zone *mz,
383                                 struct mem_cgroup_tree_per_zone *mctz,
384                                 unsigned long long new_usage_in_excess)
385 {
386         struct rb_node **p = &mctz->rb_root.rb_node;
387         struct rb_node *parent = NULL;
388         struct mem_cgroup_per_zone *mz_node;
389
390         if (mz->on_tree)
391                 return;
392
393         mz->usage_in_excess = new_usage_in_excess;
394         if (!mz->usage_in_excess)
395                 return;
396         while (*p) {
397                 parent = *p;
398                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
399                                         tree_node);
400                 if (mz->usage_in_excess < mz_node->usage_in_excess)
401                         p = &(*p)->rb_left;
402                 /*
403                  * We can't avoid mem cgroups that are over their soft
404                  * limit by the same amount
405                  */
406                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
407                         p = &(*p)->rb_right;
408         }
409         rb_link_node(&mz->tree_node, parent, p);
410         rb_insert_color(&mz->tree_node, &mctz->rb_root);
411         mz->on_tree = true;
412 }
413
414 static void
415 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
416                                 struct mem_cgroup_per_zone *mz,
417                                 struct mem_cgroup_tree_per_zone *mctz)
418 {
419         if (!mz->on_tree)
420                 return;
421         rb_erase(&mz->tree_node, &mctz->rb_root);
422         mz->on_tree = false;
423 }
424
425 static void
426 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
427                                 struct mem_cgroup_per_zone *mz,
428                                 struct mem_cgroup_tree_per_zone *mctz)
429 {
430         spin_lock(&mctz->lock);
431         __mem_cgroup_remove_exceeded(mem, mz, mctz);
432         spin_unlock(&mctz->lock);
433 }
434
435
436 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
437 {
438         unsigned long long excess;
439         struct mem_cgroup_per_zone *mz;
440         struct mem_cgroup_tree_per_zone *mctz;
441         int nid = page_to_nid(page);
442         int zid = page_zonenum(page);
443         mctz = soft_limit_tree_from_page(page);
444
445         /*
446          * Necessary to update all ancestors when hierarchy is used.
447          * because their event counter is not touched.
448          */
449         for (; mem; mem = parent_mem_cgroup(mem)) {
450                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
451                 excess = res_counter_soft_limit_excess(&mem->res);
452                 /*
453                  * We have to update the tree if mz is on RB-tree or
454                  * mem is over its softlimit.
455                  */
456                 if (excess || mz->on_tree) {
457                         spin_lock(&mctz->lock);
458                         /* if on-tree, remove it */
459                         if (mz->on_tree)
460                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
461                         /*
462                          * Insert again. mz->usage_in_excess will be updated.
463                          * If excess is 0, no tree ops.
464                          */
465                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
466                         spin_unlock(&mctz->lock);
467                 }
468         }
469 }
470
471 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
472 {
473         int node, zone;
474         struct mem_cgroup_per_zone *mz;
475         struct mem_cgroup_tree_per_zone *mctz;
476
477         for_each_node_state(node, N_POSSIBLE) {
478                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
479                         mz = mem_cgroup_zoneinfo(mem, node, zone);
480                         mctz = soft_limit_tree_node_zone(node, zone);
481                         mem_cgroup_remove_exceeded(mem, mz, mctz);
482                 }
483         }
484 }
485
486 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
487 {
488         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
489 }
490
491 static struct mem_cgroup_per_zone *
492 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
493 {
494         struct rb_node *rightmost = NULL;
495         struct mem_cgroup_per_zone *mz;
496
497 retry:
498         mz = NULL;
499         rightmost = rb_last(&mctz->rb_root);
500         if (!rightmost)
501                 goto done;              /* Nothing to reclaim from */
502
503         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
504         /*
505          * Remove the node now but someone else can add it back,
506          * we will to add it back at the end of reclaim to its correct
507          * position in the tree.
508          */
509         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
510         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
511                 !css_tryget(&mz->mem->css))
512                 goto retry;
513 done:
514         return mz;
515 }
516
517 static struct mem_cgroup_per_zone *
518 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
519 {
520         struct mem_cgroup_per_zone *mz;
521
522         spin_lock(&mctz->lock);
523         mz = __mem_cgroup_largest_soft_limit_node(mctz);
524         spin_unlock(&mctz->lock);
525         return mz;
526 }
527
528 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
529                 enum mem_cgroup_stat_index idx)
530 {
531         int cpu;
532         s64 val = 0;
533
534         for_each_possible_cpu(cpu)
535                 val += per_cpu(mem->stat->count[idx], cpu);
536         return val;
537 }
538
539 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
540 {
541         s64 ret;
542
543         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
544         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
545         return ret;
546 }
547
548 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
549                                          bool charge)
550 {
551         int val = (charge) ? 1 : -1;
552         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
553 }
554
555 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
556                                          struct page_cgroup *pc,
557                                          bool charge)
558 {
559         int val = (charge) ? 1 : -1;
560
561         preempt_disable();
562
563         if (PageCgroupCache(pc))
564                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
565         else
566                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
567
568         if (charge)
569                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
570         else
571                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
572         __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
573
574         preempt_enable();
575 }
576
577 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
578                                         enum lru_list idx)
579 {
580         int nid, zid;
581         struct mem_cgroup_per_zone *mz;
582         u64 total = 0;
583
584         for_each_online_node(nid)
585                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
586                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
587                         total += MEM_CGROUP_ZSTAT(mz, idx);
588                 }
589         return total;
590 }
591
592 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
593 {
594         s64 val;
595
596         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
597
598         return !(val & ((1 << event_mask_shift) - 1));
599 }
600
601 /*
602  * Check events in order.
603  *
604  */
605 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
606 {
607         /* threshold event is triggered in finer grain than soft limit */
608         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
609                 mem_cgroup_threshold(mem);
610                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
611                         mem_cgroup_update_tree(mem, page);
612         }
613 }
614
615 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
616 {
617         return container_of(cgroup_subsys_state(cont,
618                                 mem_cgroup_subsys_id), struct mem_cgroup,
619                                 css);
620 }
621
622 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
623 {
624         /*
625          * mm_update_next_owner() may clear mm->owner to NULL
626          * if it races with swapoff, page migration, etc.
627          * So this can be called with p == NULL.
628          */
629         if (unlikely(!p))
630                 return NULL;
631
632         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
633                                 struct mem_cgroup, css);
634 }
635
636 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
637 {
638         struct mem_cgroup *mem = NULL;
639
640         if (!mm)
641                 return NULL;
642         /*
643          * Because we have no locks, mm->owner's may be being moved to other
644          * cgroup. We use css_tryget() here even if this looks
645          * pessimistic (rather than adding locks here).
646          */
647         rcu_read_lock();
648         do {
649                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
650                 if (unlikely(!mem))
651                         break;
652         } while (!css_tryget(&mem->css));
653         rcu_read_unlock();
654         return mem;
655 }
656
657 /*
658  * Call callback function against all cgroup under hierarchy tree.
659  */
660 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
661                           int (*func)(struct mem_cgroup *, void *))
662 {
663         int found, ret, nextid;
664         struct cgroup_subsys_state *css;
665         struct mem_cgroup *mem;
666
667         if (!root->use_hierarchy)
668                 return (*func)(root, data);
669
670         nextid = 1;
671         do {
672                 ret = 0;
673                 mem = NULL;
674
675                 rcu_read_lock();
676                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
677                                    &found);
678                 if (css && css_tryget(css))
679                         mem = container_of(css, struct mem_cgroup, css);
680                 rcu_read_unlock();
681
682                 if (mem) {
683                         ret = (*func)(mem, data);
684                         css_put(&mem->css);
685                 }
686                 nextid = found + 1;
687         } while (!ret && css);
688
689         return ret;
690 }
691
692 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
693 {
694         return (mem == root_mem_cgroup);
695 }
696
697 /*
698  * Following LRU functions are allowed to be used without PCG_LOCK.
699  * Operations are called by routine of global LRU independently from memcg.
700  * What we have to take care of here is validness of pc->mem_cgroup.
701  *
702  * Changes to pc->mem_cgroup happens when
703  * 1. charge
704  * 2. moving account
705  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
706  * It is added to LRU before charge.
707  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
708  * When moving account, the page is not on LRU. It's isolated.
709  */
710
711 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
712 {
713         struct page_cgroup *pc;
714         struct mem_cgroup_per_zone *mz;
715
716         if (mem_cgroup_disabled())
717                 return;
718         pc = lookup_page_cgroup(page);
719         /* can happen while we handle swapcache. */
720         if (!TestClearPageCgroupAcctLRU(pc))
721                 return;
722         VM_BUG_ON(!pc->mem_cgroup);
723         /*
724          * We don't check PCG_USED bit. It's cleared when the "page" is finally
725          * removed from global LRU.
726          */
727         mz = page_cgroup_zoneinfo(pc);
728         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
729         if (mem_cgroup_is_root(pc->mem_cgroup))
730                 return;
731         VM_BUG_ON(list_empty(&pc->lru));
732         list_del_init(&pc->lru);
733         return;
734 }
735
736 void mem_cgroup_del_lru(struct page *page)
737 {
738         mem_cgroup_del_lru_list(page, page_lru(page));
739 }
740
741 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
742 {
743         struct mem_cgroup_per_zone *mz;
744         struct page_cgroup *pc;
745
746         if (mem_cgroup_disabled())
747                 return;
748
749         pc = lookup_page_cgroup(page);
750         /*
751          * Used bit is set without atomic ops but after smp_wmb().
752          * For making pc->mem_cgroup visible, insert smp_rmb() here.
753          */
754         smp_rmb();
755         /* unused or root page is not rotated. */
756         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
757                 return;
758         mz = page_cgroup_zoneinfo(pc);
759         list_move(&pc->lru, &mz->lists[lru]);
760 }
761
762 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
763 {
764         struct page_cgroup *pc;
765         struct mem_cgroup_per_zone *mz;
766
767         if (mem_cgroup_disabled())
768                 return;
769         pc = lookup_page_cgroup(page);
770         VM_BUG_ON(PageCgroupAcctLRU(pc));
771         /*
772          * Used bit is set without atomic ops but after smp_wmb().
773          * For making pc->mem_cgroup visible, insert smp_rmb() here.
774          */
775         smp_rmb();
776         if (!PageCgroupUsed(pc))
777                 return;
778
779         mz = page_cgroup_zoneinfo(pc);
780         MEM_CGROUP_ZSTAT(mz, lru) += 1;
781         SetPageCgroupAcctLRU(pc);
782         if (mem_cgroup_is_root(pc->mem_cgroup))
783                 return;
784         list_add(&pc->lru, &mz->lists[lru]);
785 }
786
787 /*
788  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
789  * lru because the page may.be reused after it's fully uncharged (because of
790  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
791  * it again. This function is only used to charge SwapCache. It's done under
792  * lock_page and expected that zone->lru_lock is never held.
793  */
794 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
795 {
796         unsigned long flags;
797         struct zone *zone = page_zone(page);
798         struct page_cgroup *pc = lookup_page_cgroup(page);
799
800         spin_lock_irqsave(&zone->lru_lock, flags);
801         /*
802          * Forget old LRU when this page_cgroup is *not* used. This Used bit
803          * is guarded by lock_page() because the page is SwapCache.
804          */
805         if (!PageCgroupUsed(pc))
806                 mem_cgroup_del_lru_list(page, page_lru(page));
807         spin_unlock_irqrestore(&zone->lru_lock, flags);
808 }
809
810 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
811 {
812         unsigned long flags;
813         struct zone *zone = page_zone(page);
814         struct page_cgroup *pc = lookup_page_cgroup(page);
815
816         spin_lock_irqsave(&zone->lru_lock, flags);
817         /* link when the page is linked to LRU but page_cgroup isn't */
818         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
819                 mem_cgroup_add_lru_list(page, page_lru(page));
820         spin_unlock_irqrestore(&zone->lru_lock, flags);
821 }
822
823
824 void mem_cgroup_move_lists(struct page *page,
825                            enum lru_list from, enum lru_list to)
826 {
827         if (mem_cgroup_disabled())
828                 return;
829         mem_cgroup_del_lru_list(page, from);
830         mem_cgroup_add_lru_list(page, to);
831 }
832
833 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
834 {
835         int ret;
836         struct mem_cgroup *curr = NULL;
837
838         task_lock(task);
839         rcu_read_lock();
840         curr = try_get_mem_cgroup_from_mm(task->mm);
841         rcu_read_unlock();
842         task_unlock(task);
843         if (!curr)
844                 return 0;
845         /*
846          * We should check use_hierarchy of "mem" not "curr". Because checking
847          * use_hierarchy of "curr" here make this function true if hierarchy is
848          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
849          * hierarchy(even if use_hierarchy is disabled in "mem").
850          */
851         if (mem->use_hierarchy)
852                 ret = css_is_ancestor(&curr->css, &mem->css);
853         else
854                 ret = (curr == mem);
855         css_put(&curr->css);
856         return ret;
857 }
858
859 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
860 {
861         unsigned long active;
862         unsigned long inactive;
863         unsigned long gb;
864         unsigned long inactive_ratio;
865
866         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
867         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
868
869         gb = (inactive + active) >> (30 - PAGE_SHIFT);
870         if (gb)
871                 inactive_ratio = int_sqrt(10 * gb);
872         else
873                 inactive_ratio = 1;
874
875         if (present_pages) {
876                 present_pages[0] = inactive;
877                 present_pages[1] = active;
878         }
879
880         return inactive_ratio;
881 }
882
883 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
884 {
885         unsigned long active;
886         unsigned long inactive;
887         unsigned long present_pages[2];
888         unsigned long inactive_ratio;
889
890         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
891
892         inactive = present_pages[0];
893         active = present_pages[1];
894
895         if (inactive * inactive_ratio < active)
896                 return 1;
897
898         return 0;
899 }
900
901 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
902 {
903         unsigned long active;
904         unsigned long inactive;
905
906         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
907         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
908
909         return (active > inactive);
910 }
911
912 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
913                                        struct zone *zone,
914                                        enum lru_list lru)
915 {
916         int nid = zone->zone_pgdat->node_id;
917         int zid = zone_idx(zone);
918         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
919
920         return MEM_CGROUP_ZSTAT(mz, lru);
921 }
922
923 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
924                                                       struct zone *zone)
925 {
926         int nid = zone->zone_pgdat->node_id;
927         int zid = zone_idx(zone);
928         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
929
930         return &mz->reclaim_stat;
931 }
932
933 struct zone_reclaim_stat *
934 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
935 {
936         struct page_cgroup *pc;
937         struct mem_cgroup_per_zone *mz;
938
939         if (mem_cgroup_disabled())
940                 return NULL;
941
942         pc = lookup_page_cgroup(page);
943         /*
944          * Used bit is set without atomic ops but after smp_wmb().
945          * For making pc->mem_cgroup visible, insert smp_rmb() here.
946          */
947         smp_rmb();
948         if (!PageCgroupUsed(pc))
949                 return NULL;
950
951         mz = page_cgroup_zoneinfo(pc);
952         if (!mz)
953                 return NULL;
954
955         return &mz->reclaim_stat;
956 }
957
958 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
959                                         struct list_head *dst,
960                                         unsigned long *scanned, int order,
961                                         int mode, struct zone *z,
962                                         struct mem_cgroup *mem_cont,
963                                         int active, int file)
964 {
965         unsigned long nr_taken = 0;
966         struct page *page;
967         unsigned long scan;
968         LIST_HEAD(pc_list);
969         struct list_head *src;
970         struct page_cgroup *pc, *tmp;
971         int nid = z->zone_pgdat->node_id;
972         int zid = zone_idx(z);
973         struct mem_cgroup_per_zone *mz;
974         int lru = LRU_FILE * file + active;
975         int ret;
976
977         BUG_ON(!mem_cont);
978         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
979         src = &mz->lists[lru];
980
981         scan = 0;
982         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
983                 if (scan >= nr_to_scan)
984                         break;
985
986                 page = pc->page;
987                 if (unlikely(!PageCgroupUsed(pc)))
988                         continue;
989                 if (unlikely(!PageLRU(page)))
990                         continue;
991
992                 scan++;
993                 ret = __isolate_lru_page(page, mode, file);
994                 switch (ret) {
995                 case 0:
996                         list_move(&page->lru, dst);
997                         mem_cgroup_del_lru(page);
998                         nr_taken++;
999                         break;
1000                 case -EBUSY:
1001                         /* we don't affect global LRU but rotate in our LRU */
1002                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1003                         break;
1004                 default:
1005                         break;
1006                 }
1007         }
1008
1009         *scanned = scan;
1010         return nr_taken;
1011 }
1012
1013 #define mem_cgroup_from_res_counter(counter, member)    \
1014         container_of(counter, struct mem_cgroup, member)
1015
1016 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1017 {
1018         if (do_swap_account) {
1019                 if (res_counter_check_under_limit(&mem->res) &&
1020                         res_counter_check_under_limit(&mem->memsw))
1021                         return true;
1022         } else
1023                 if (res_counter_check_under_limit(&mem->res))
1024                         return true;
1025         return false;
1026 }
1027
1028 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1029 {
1030         struct cgroup *cgrp = memcg->css.cgroup;
1031         unsigned int swappiness;
1032
1033         /* root ? */
1034         if (cgrp->parent == NULL)
1035                 return vm_swappiness;
1036
1037         spin_lock(&memcg->reclaim_param_lock);
1038         swappiness = memcg->swappiness;
1039         spin_unlock(&memcg->reclaim_param_lock);
1040
1041         return swappiness;
1042 }
1043
1044 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1045 {
1046         int *val = data;
1047         (*val)++;
1048         return 0;
1049 }
1050
1051 /**
1052  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1053  * @memcg: The memory cgroup that went over limit
1054  * @p: Task that is going to be killed
1055  *
1056  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1057  * enabled
1058  */
1059 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1060 {
1061         struct cgroup *task_cgrp;
1062         struct cgroup *mem_cgrp;
1063         /*
1064          * Need a buffer in BSS, can't rely on allocations. The code relies
1065          * on the assumption that OOM is serialized for memory controller.
1066          * If this assumption is broken, revisit this code.
1067          */
1068         static char memcg_name[PATH_MAX];
1069         int ret;
1070
1071         if (!memcg || !p)
1072                 return;
1073
1074
1075         rcu_read_lock();
1076
1077         mem_cgrp = memcg->css.cgroup;
1078         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1079
1080         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1081         if (ret < 0) {
1082                 /*
1083                  * Unfortunately, we are unable to convert to a useful name
1084                  * But we'll still print out the usage information
1085                  */
1086                 rcu_read_unlock();
1087                 goto done;
1088         }
1089         rcu_read_unlock();
1090
1091         printk(KERN_INFO "Task in %s killed", memcg_name);
1092
1093         rcu_read_lock();
1094         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1095         if (ret < 0) {
1096                 rcu_read_unlock();
1097                 goto done;
1098         }
1099         rcu_read_unlock();
1100
1101         /*
1102          * Continues from above, so we don't need an KERN_ level
1103          */
1104         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1105 done:
1106
1107         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1108                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1109                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1110                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1111         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1112                 "failcnt %llu\n",
1113                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1114                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1115                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1116 }
1117
1118 /*
1119  * This function returns the number of memcg under hierarchy tree. Returns
1120  * 1(self count) if no children.
1121  */
1122 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1123 {
1124         int num = 0;
1125         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1126         return num;
1127 }
1128
1129 /*
1130  * Return the memory (and swap, if configured) limit for a memcg.
1131  */
1132 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1133 {
1134         u64 limit;
1135         u64 memsw;
1136
1137         limit = res_counter_read_u64(&memcg->res, RES_LIMIT) +
1138                         total_swap_pages;
1139         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1140         /*
1141          * If memsw is finite and limits the amount of swap space available
1142          * to this memcg, return that limit.
1143          */
1144         return min(limit, memsw);
1145 }
1146
1147 /*
1148  * Visit the first child (need not be the first child as per the ordering
1149  * of the cgroup list, since we track last_scanned_child) of @mem and use
1150  * that to reclaim free pages from.
1151  */
1152 static struct mem_cgroup *
1153 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1154 {
1155         struct mem_cgroup *ret = NULL;
1156         struct cgroup_subsys_state *css;
1157         int nextid, found;
1158
1159         if (!root_mem->use_hierarchy) {
1160                 css_get(&root_mem->css);
1161                 ret = root_mem;
1162         }
1163
1164         while (!ret) {
1165                 rcu_read_lock();
1166                 nextid = root_mem->last_scanned_child + 1;
1167                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1168                                    &found);
1169                 if (css && css_tryget(css))
1170                         ret = container_of(css, struct mem_cgroup, css);
1171
1172                 rcu_read_unlock();
1173                 /* Updates scanning parameter */
1174                 spin_lock(&root_mem->reclaim_param_lock);
1175                 if (!css) {
1176                         /* this means start scan from ID:1 */
1177                         root_mem->last_scanned_child = 0;
1178                 } else
1179                         root_mem->last_scanned_child = found;
1180                 spin_unlock(&root_mem->reclaim_param_lock);
1181         }
1182
1183         return ret;
1184 }
1185
1186 /*
1187  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1188  * we reclaimed from, so that we don't end up penalizing one child extensively
1189  * based on its position in the children list.
1190  *
1191  * root_mem is the original ancestor that we've been reclaim from.
1192  *
1193  * We give up and return to the caller when we visit root_mem twice.
1194  * (other groups can be removed while we're walking....)
1195  *
1196  * If shrink==true, for avoiding to free too much, this returns immedieately.
1197  */
1198 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1199                                                 struct zone *zone,
1200                                                 gfp_t gfp_mask,
1201                                                 unsigned long reclaim_options)
1202 {
1203         struct mem_cgroup *victim;
1204         int ret, total = 0;
1205         int loop = 0;
1206         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1207         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1208         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1209         unsigned long excess = mem_cgroup_get_excess(root_mem);
1210
1211         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1212         if (root_mem->memsw_is_minimum)
1213                 noswap = true;
1214
1215         while (1) {
1216                 victim = mem_cgroup_select_victim(root_mem);
1217                 if (victim == root_mem) {
1218                         loop++;
1219                         if (loop >= 1)
1220                                 drain_all_stock_async();
1221                         if (loop >= 2) {
1222                                 /*
1223                                  * If we have not been able to reclaim
1224                                  * anything, it might because there are
1225                                  * no reclaimable pages under this hierarchy
1226                                  */
1227                                 if (!check_soft || !total) {
1228                                         css_put(&victim->css);
1229                                         break;
1230                                 }
1231                                 /*
1232                                  * We want to do more targetted reclaim.
1233                                  * excess >> 2 is not to excessive so as to
1234                                  * reclaim too much, nor too less that we keep
1235                                  * coming back to reclaim from this cgroup
1236                                  */
1237                                 if (total >= (excess >> 2) ||
1238                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1239                                         css_put(&victim->css);
1240                                         break;
1241                                 }
1242                         }
1243                 }
1244                 if (!mem_cgroup_local_usage(victim)) {
1245                         /* this cgroup's local usage == 0 */
1246                         css_put(&victim->css);
1247                         continue;
1248                 }
1249                 /* we use swappiness of local cgroup */
1250                 if (check_soft)
1251                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1252                                 noswap, get_swappiness(victim), zone,
1253                                 zone->zone_pgdat->node_id);
1254                 else
1255                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1256                                                 noswap, get_swappiness(victim));
1257                 css_put(&victim->css);
1258                 /*
1259                  * At shrinking usage, we can't check we should stop here or
1260                  * reclaim more. It's depends on callers. last_scanned_child
1261                  * will work enough for keeping fairness under tree.
1262                  */
1263                 if (shrink)
1264                         return ret;
1265                 total += ret;
1266                 if (check_soft) {
1267                         if (res_counter_check_under_soft_limit(&root_mem->res))
1268                                 return total;
1269                 } else if (mem_cgroup_check_under_limit(root_mem))
1270                         return 1 + total;
1271         }
1272         return total;
1273 }
1274
1275 static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1276 {
1277         int *val = (int *)data;
1278         int x;
1279         /*
1280          * Logically, we can stop scanning immediately when we find
1281          * a memcg is already locked. But condidering unlock ops and
1282          * creation/removal of memcg, scan-all is simple operation.
1283          */
1284         x = atomic_inc_return(&mem->oom_lock);
1285         *val = max(x, *val);
1286         return 0;
1287 }
1288 /*
1289  * Check OOM-Killer is already running under our hierarchy.
1290  * If someone is running, return false.
1291  */
1292 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1293 {
1294         int lock_count = 0;
1295
1296         mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1297
1298         if (lock_count == 1)
1299                 return true;
1300         return false;
1301 }
1302
1303 static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1304 {
1305         /*
1306          * When a new child is created while the hierarchy is under oom,
1307          * mem_cgroup_oom_lock() may not be called. We have to use
1308          * atomic_add_unless() here.
1309          */
1310         atomic_add_unless(&mem->oom_lock, -1, 0);
1311         return 0;
1312 }
1313
1314 static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1315 {
1316         mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1317 }
1318
1319 static DEFINE_MUTEX(memcg_oom_mutex);
1320 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1321
1322 struct oom_wait_info {
1323         struct mem_cgroup *mem;
1324         wait_queue_t    wait;
1325 };
1326
1327 static int memcg_oom_wake_function(wait_queue_t *wait,
1328         unsigned mode, int sync, void *arg)
1329 {
1330         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1331         struct oom_wait_info *oom_wait_info;
1332
1333         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1334
1335         if (oom_wait_info->mem == wake_mem)
1336                 goto wakeup;
1337         /* if no hierarchy, no match */
1338         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1339                 return 0;
1340         /*
1341          * Both of oom_wait_info->mem and wake_mem are stable under us.
1342          * Then we can use css_is_ancestor without taking care of RCU.
1343          */
1344         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1345             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1346                 return 0;
1347
1348 wakeup:
1349         return autoremove_wake_function(wait, mode, sync, arg);
1350 }
1351
1352 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1353 {
1354         /* for filtering, pass "mem" as argument. */
1355         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1356 }
1357
1358 static void memcg_oom_recover(struct mem_cgroup *mem)
1359 {
1360         if (atomic_read(&mem->oom_lock))
1361                 memcg_wakeup_oom(mem);
1362 }
1363
1364 /*
1365  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1366  */
1367 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1368 {
1369         struct oom_wait_info owait;
1370         bool locked, need_to_kill;
1371
1372         owait.mem = mem;
1373         owait.wait.flags = 0;
1374         owait.wait.func = memcg_oom_wake_function;
1375         owait.wait.private = current;
1376         INIT_LIST_HEAD(&owait.wait.task_list);
1377         need_to_kill = true;
1378         /* At first, try to OOM lock hierarchy under mem.*/
1379         mutex_lock(&memcg_oom_mutex);
1380         locked = mem_cgroup_oom_lock(mem);
1381         /*
1382          * Even if signal_pending(), we can't quit charge() loop without
1383          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1384          * under OOM is always welcomed, use TASK_KILLABLE here.
1385          */
1386         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1387         if (!locked || mem->oom_kill_disable)
1388                 need_to_kill = false;
1389         if (locked)
1390                 mem_cgroup_oom_notify(mem);
1391         mutex_unlock(&memcg_oom_mutex);
1392
1393         if (need_to_kill) {
1394                 finish_wait(&memcg_oom_waitq, &owait.wait);
1395                 mem_cgroup_out_of_memory(mem, mask);
1396         } else {
1397                 schedule();
1398                 finish_wait(&memcg_oom_waitq, &owait.wait);
1399         }
1400         mutex_lock(&memcg_oom_mutex);
1401         mem_cgroup_oom_unlock(mem);
1402         memcg_wakeup_oom(mem);
1403         mutex_unlock(&memcg_oom_mutex);
1404
1405         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1406                 return false;
1407         /* Give chance to dying process */
1408         schedule_timeout(1);
1409         return true;
1410 }
1411
1412 /*
1413  * Currently used to update mapped file statistics, but the routine can be
1414  * generalized to update other statistics as well.
1415  */
1416 void mem_cgroup_update_file_mapped(struct page *page, int val)
1417 {
1418         struct mem_cgroup *mem;
1419         struct page_cgroup *pc;
1420
1421         pc = lookup_page_cgroup(page);
1422         if (unlikely(!pc))
1423                 return;
1424
1425         lock_page_cgroup(pc);
1426         mem = pc->mem_cgroup;
1427         if (!mem || !PageCgroupUsed(pc))
1428                 goto done;
1429
1430         /*
1431          * Preemption is already disabled. We can use __this_cpu_xxx
1432          */
1433         if (val > 0) {
1434                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1435                 SetPageCgroupFileMapped(pc);
1436         } else {
1437                 __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1438                 ClearPageCgroupFileMapped(pc);
1439         }
1440
1441 done:
1442         unlock_page_cgroup(pc);
1443 }
1444
1445 /*
1446  * size of first charge trial. "32" comes from vmscan.c's magic value.
1447  * TODO: maybe necessary to use big numbers in big irons.
1448  */
1449 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1450 struct memcg_stock_pcp {
1451         struct mem_cgroup *cached; /* this never be root cgroup */
1452         int charge;
1453         struct work_struct work;
1454 };
1455 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1456 static atomic_t memcg_drain_count;
1457
1458 /*
1459  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1460  * from local stock and true is returned. If the stock is 0 or charges from a
1461  * cgroup which is not current target, returns false. This stock will be
1462  * refilled.
1463  */
1464 static bool consume_stock(struct mem_cgroup *mem)
1465 {
1466         struct memcg_stock_pcp *stock;
1467         bool ret = true;
1468
1469         stock = &get_cpu_var(memcg_stock);
1470         if (mem == stock->cached && stock->charge)
1471                 stock->charge -= PAGE_SIZE;
1472         else /* need to call res_counter_charge */
1473                 ret = false;
1474         put_cpu_var(memcg_stock);
1475         return ret;
1476 }
1477
1478 /*
1479  * Returns stocks cached in percpu to res_counter and reset cached information.
1480  */
1481 static void drain_stock(struct memcg_stock_pcp *stock)
1482 {
1483         struct mem_cgroup *old = stock->cached;
1484
1485         if (stock->charge) {
1486                 res_counter_uncharge(&old->res, stock->charge);
1487                 if (do_swap_account)
1488                         res_counter_uncharge(&old->memsw, stock->charge);
1489         }
1490         stock->cached = NULL;
1491         stock->charge = 0;
1492 }
1493
1494 /*
1495  * This must be called under preempt disabled or must be called by
1496  * a thread which is pinned to local cpu.
1497  */
1498 static void drain_local_stock(struct work_struct *dummy)
1499 {
1500         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1501         drain_stock(stock);
1502 }
1503
1504 /*
1505  * Cache charges(val) which is from res_counter, to local per_cpu area.
1506  * This will be consumed by consume_stock() function, later.
1507  */
1508 static void refill_stock(struct mem_cgroup *mem, int val)
1509 {
1510         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1511
1512         if (stock->cached != mem) { /* reset if necessary */
1513                 drain_stock(stock);
1514                 stock->cached = mem;
1515         }
1516         stock->charge += val;
1517         put_cpu_var(memcg_stock);
1518 }
1519
1520 /*
1521  * Tries to drain stocked charges in other cpus. This function is asynchronous
1522  * and just put a work per cpu for draining localy on each cpu. Caller can
1523  * expects some charges will be back to res_counter later but cannot wait for
1524  * it.
1525  */
1526 static void drain_all_stock_async(void)
1527 {
1528         int cpu;
1529         /* This function is for scheduling "drain" in asynchronous way.
1530          * The result of "drain" is not directly handled by callers. Then,
1531          * if someone is calling drain, we don't have to call drain more.
1532          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1533          * there is a race. We just do loose check here.
1534          */
1535         if (atomic_read(&memcg_drain_count))
1536                 return;
1537         /* Notify other cpus that system-wide "drain" is running */
1538         atomic_inc(&memcg_drain_count);
1539         get_online_cpus();
1540         for_each_online_cpu(cpu) {
1541                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1542                 schedule_work_on(cpu, &stock->work);
1543         }
1544         put_online_cpus();
1545         atomic_dec(&memcg_drain_count);
1546         /* We don't wait for flush_work */
1547 }
1548
1549 /* This is a synchronous drain interface. */
1550 static void drain_all_stock_sync(void)
1551 {
1552         /* called when force_empty is called */
1553         atomic_inc(&memcg_drain_count);
1554         schedule_on_each_cpu(drain_local_stock);
1555         atomic_dec(&memcg_drain_count);
1556 }
1557
1558 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1559                                         unsigned long action,
1560                                         void *hcpu)
1561 {
1562         int cpu = (unsigned long)hcpu;
1563         struct memcg_stock_pcp *stock;
1564
1565         if (action != CPU_DEAD)
1566                 return NOTIFY_OK;
1567         stock = &per_cpu(memcg_stock, cpu);
1568         drain_stock(stock);
1569         return NOTIFY_OK;
1570 }
1571
1572 /*
1573  * Unlike exported interface, "oom" parameter is added. if oom==true,
1574  * oom-killer can be invoked.
1575  */
1576 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1577                         gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1578 {
1579         struct mem_cgroup *mem, *mem_over_limit;
1580         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1581         struct res_counter *fail_res;
1582         int csize = CHARGE_SIZE;
1583
1584         /*
1585          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1586          * in system level. So, allow to go ahead dying process in addition to
1587          * MEMDIE process.
1588          */
1589         if (unlikely(test_thread_flag(TIF_MEMDIE)
1590                      || fatal_signal_pending(current)))
1591                 goto bypass;
1592
1593         /*
1594          * We always charge the cgroup the mm_struct belongs to.
1595          * The mm_struct's mem_cgroup changes on task migration if the
1596          * thread group leader migrates. It's possible that mm is not
1597          * set, if so charge the init_mm (happens for pagecache usage).
1598          */
1599         mem = *memcg;
1600         if (likely(!mem)) {
1601                 mem = try_get_mem_cgroup_from_mm(mm);
1602                 *memcg = mem;
1603         } else {
1604                 css_get(&mem->css);
1605         }
1606         if (unlikely(!mem))
1607                 return 0;
1608
1609         VM_BUG_ON(css_is_removed(&mem->css));
1610         if (mem_cgroup_is_root(mem))
1611                 goto done;
1612
1613         while (1) {
1614                 int ret = 0;
1615                 unsigned long flags = 0;
1616
1617                 if (consume_stock(mem))
1618                         goto done;
1619
1620                 ret = res_counter_charge(&mem->res, csize, &fail_res);
1621                 if (likely(!ret)) {
1622                         if (!do_swap_account)
1623                                 break;
1624                         ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1625                         if (likely(!ret))
1626                                 break;
1627                         /* mem+swap counter fails */
1628                         res_counter_uncharge(&mem->res, csize);
1629                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1630                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1631                                                                         memsw);
1632                 } else
1633                         /* mem counter fails */
1634                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1635                                                                         res);
1636
1637                 /* reduce request size and retry */
1638                 if (csize > PAGE_SIZE) {
1639                         csize = PAGE_SIZE;
1640                         continue;
1641                 }
1642                 if (!(gfp_mask & __GFP_WAIT))
1643                         goto nomem;
1644
1645                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1646                                                 gfp_mask, flags);
1647                 if (ret)
1648                         continue;
1649
1650                 /*
1651                  * try_to_free_mem_cgroup_pages() might not give us a full
1652                  * picture of reclaim. Some pages are reclaimed and might be
1653                  * moved to swap cache or just unmapped from the cgroup.
1654                  * Check the limit again to see if the reclaim reduced the
1655                  * current usage of the cgroup before giving up
1656                  *
1657                  */
1658                 if (mem_cgroup_check_under_limit(mem_over_limit))
1659                         continue;
1660
1661                 /* try to avoid oom while someone is moving charge */
1662                 if (mc.moving_task && current != mc.moving_task) {
1663                         struct mem_cgroup *from, *to;
1664                         bool do_continue = false;
1665                         /*
1666                          * There is a small race that "from" or "to" can be
1667                          * freed by rmdir, so we use css_tryget().
1668                          */
1669                         from = mc.from;
1670                         to = mc.to;
1671                         if (from && css_tryget(&from->css)) {
1672                                 if (mem_over_limit->use_hierarchy)
1673                                         do_continue = css_is_ancestor(
1674                                                         &from->css,
1675                                                         &mem_over_limit->css);
1676                                 else
1677                                         do_continue = (from == mem_over_limit);
1678                                 css_put(&from->css);
1679                         }
1680                         if (!do_continue && to && css_tryget(&to->css)) {
1681                                 if (mem_over_limit->use_hierarchy)
1682                                         do_continue = css_is_ancestor(
1683                                                         &to->css,
1684                                                         &mem_over_limit->css);
1685                                 else
1686                                         do_continue = (to == mem_over_limit);
1687                                 css_put(&to->css);
1688                         }
1689                         if (do_continue) {
1690                                 DEFINE_WAIT(wait);
1691                                 prepare_to_wait(&mc.waitq, &wait,
1692                                                         TASK_INTERRUPTIBLE);
1693                                 /* moving charge context might have finished. */
1694                                 if (mc.moving_task)
1695                                         schedule();
1696                                 finish_wait(&mc.waitq, &wait);
1697                                 continue;
1698                         }
1699                 }
1700
1701                 if (!nr_retries--) {
1702                         if (!oom)
1703                                 goto nomem;
1704                         if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) {
1705                                 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1706                                 continue;
1707                         }
1708                         /* When we reach here, current task is dying .*/
1709                         css_put(&mem->css);
1710                         goto bypass;
1711                 }
1712         }
1713         if (csize > PAGE_SIZE)
1714                 refill_stock(mem, csize - PAGE_SIZE);
1715 done:
1716         return 0;
1717 nomem:
1718         css_put(&mem->css);
1719         return -ENOMEM;
1720 bypass:
1721         *memcg = NULL;
1722         return 0;
1723 }
1724
1725 /*
1726  * Somemtimes we have to undo a charge we got by try_charge().
1727  * This function is for that and do uncharge, put css's refcnt.
1728  * gotten by try_charge().
1729  */
1730 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1731                                                         unsigned long count)
1732 {
1733         if (!mem_cgroup_is_root(mem)) {
1734                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1735                 if (do_swap_account)
1736                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1737                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1738                 WARN_ON_ONCE(count > INT_MAX);
1739                 __css_put(&mem->css, (int)count);
1740         }
1741         /* we don't need css_put for root */
1742 }
1743
1744 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1745 {
1746         __mem_cgroup_cancel_charge(mem, 1);
1747 }
1748
1749 /*
1750  * A helper function to get mem_cgroup from ID. must be called under
1751  * rcu_read_lock(). The caller must check css_is_removed() or some if
1752  * it's concern. (dropping refcnt from swap can be called against removed
1753  * memcg.)
1754  */
1755 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1756 {
1757         struct cgroup_subsys_state *css;
1758
1759         /* ID 0 is unused ID */
1760         if (!id)
1761                 return NULL;
1762         css = css_lookup(&mem_cgroup_subsys, id);
1763         if (!css)
1764                 return NULL;
1765         return container_of(css, struct mem_cgroup, css);
1766 }
1767
1768 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1769 {
1770         struct mem_cgroup *mem = NULL;
1771         struct page_cgroup *pc;
1772         unsigned short id;
1773         swp_entry_t ent;
1774
1775         VM_BUG_ON(!PageLocked(page));
1776
1777         pc = lookup_page_cgroup(page);
1778         lock_page_cgroup(pc);
1779         if (PageCgroupUsed(pc)) {
1780                 mem = pc->mem_cgroup;
1781                 if (mem && !css_tryget(&mem->css))
1782                         mem = NULL;
1783         } else if (PageSwapCache(page)) {
1784                 ent.val = page_private(page);
1785                 id = lookup_swap_cgroup(ent);
1786                 rcu_read_lock();
1787                 mem = mem_cgroup_lookup(id);
1788                 if (mem && !css_tryget(&mem->css))
1789                         mem = NULL;
1790                 rcu_read_unlock();
1791         }
1792         unlock_page_cgroup(pc);
1793         return mem;
1794 }
1795
1796 /*
1797  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1798  * USED state. If already USED, uncharge and return.
1799  */
1800
1801 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1802                                      struct page_cgroup *pc,
1803                                      enum charge_type ctype)
1804 {
1805         /* try_charge() can return NULL to *memcg, taking care of it. */
1806         if (!mem)
1807                 return;
1808
1809         lock_page_cgroup(pc);
1810         if (unlikely(PageCgroupUsed(pc))) {
1811                 unlock_page_cgroup(pc);
1812                 mem_cgroup_cancel_charge(mem);
1813                 return;
1814         }
1815
1816         pc->mem_cgroup = mem;
1817         /*
1818          * We access a page_cgroup asynchronously without lock_page_cgroup().
1819          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1820          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1821          * before USED bit, we need memory barrier here.
1822          * See mem_cgroup_add_lru_list(), etc.
1823          */
1824         smp_wmb();
1825         switch (ctype) {
1826         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1827         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1828                 SetPageCgroupCache(pc);
1829                 SetPageCgroupUsed(pc);
1830                 break;
1831         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1832                 ClearPageCgroupCache(pc);
1833                 SetPageCgroupUsed(pc);
1834                 break;
1835         default:
1836                 break;
1837         }
1838
1839         mem_cgroup_charge_statistics(mem, pc, true);
1840
1841         unlock_page_cgroup(pc);
1842         /*
1843          * "charge_statistics" updated event counter. Then, check it.
1844          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1845          * if they exceeds softlimit.
1846          */
1847         memcg_check_events(mem, pc->page);
1848 }
1849
1850 /**
1851  * __mem_cgroup_move_account - move account of the page
1852  * @pc: page_cgroup of the page.
1853  * @from: mem_cgroup which the page is moved from.
1854  * @to: mem_cgroup which the page is moved to. @from != @to.
1855  * @uncharge: whether we should call uncharge and css_put against @from.
1856  *
1857  * The caller must confirm following.
1858  * - page is not on LRU (isolate_page() is useful.)
1859  * - the pc is locked, used, and ->mem_cgroup points to @from.
1860  *
1861  * This function doesn't do "charge" nor css_get to new cgroup. It should be
1862  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1863  * true, this function does "uncharge" from old cgroup, but it doesn't if
1864  * @uncharge is false, so a caller should do "uncharge".
1865  */
1866
1867 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1868         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1869 {
1870         VM_BUG_ON(from == to);
1871         VM_BUG_ON(PageLRU(pc->page));
1872         VM_BUG_ON(!PageCgroupLocked(pc));
1873         VM_BUG_ON(!PageCgroupUsed(pc));
1874         VM_BUG_ON(pc->mem_cgroup != from);
1875
1876         if (PageCgroupFileMapped(pc)) {
1877                 /* Update mapped_file data for mem_cgroup */
1878                 preempt_disable();
1879                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1880                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1881                 preempt_enable();
1882         }
1883         mem_cgroup_charge_statistics(from, pc, false);
1884         if (uncharge)
1885                 /* This is not "cancel", but cancel_charge does all we need. */
1886                 mem_cgroup_cancel_charge(from);
1887
1888         /* caller should have done css_get */
1889         pc->mem_cgroup = to;
1890         mem_cgroup_charge_statistics(to, pc, true);
1891         /*
1892          * We charges against "to" which may not have any tasks. Then, "to"
1893          * can be under rmdir(). But in current implementation, caller of
1894          * this function is just force_empty() and move charge, so it's
1895          * garanteed that "to" is never removed. So, we don't check rmdir
1896          * status here.
1897          */
1898 }
1899
1900 /*
1901  * check whether the @pc is valid for moving account and call
1902  * __mem_cgroup_move_account()
1903  */
1904 static int mem_cgroup_move_account(struct page_cgroup *pc,
1905                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1906 {
1907         int ret = -EINVAL;
1908         lock_page_cgroup(pc);
1909         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1910                 __mem_cgroup_move_account(pc, from, to, uncharge);
1911                 ret = 0;
1912         }
1913         unlock_page_cgroup(pc);
1914         /*
1915          * check events
1916          */
1917         memcg_check_events(to, pc->page);
1918         memcg_check_events(from, pc->page);
1919         return ret;
1920 }
1921
1922 /*
1923  * move charges to its parent.
1924  */
1925
1926 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1927                                   struct mem_cgroup *child,
1928                                   gfp_t gfp_mask)
1929 {
1930         struct page *page = pc->page;
1931         struct cgroup *cg = child->css.cgroup;
1932         struct cgroup *pcg = cg->parent;
1933         struct mem_cgroup *parent;
1934         int ret;
1935
1936         /* Is ROOT ? */
1937         if (!pcg)
1938                 return -EINVAL;
1939
1940         ret = -EBUSY;
1941         if (!get_page_unless_zero(page))
1942                 goto out;
1943         if (isolate_lru_page(page))
1944                 goto put;
1945
1946         parent = mem_cgroup_from_cont(pcg);
1947         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1948         if (ret || !parent)
1949                 goto put_back;
1950
1951         ret = mem_cgroup_move_account(pc, child, parent, true);
1952         if (ret)
1953                 mem_cgroup_cancel_charge(parent);
1954 put_back:
1955         putback_lru_page(page);
1956 put:
1957         put_page(page);
1958 out:
1959         return ret;
1960 }
1961
1962 /*
1963  * Charge the memory controller for page usage.
1964  * Return
1965  * 0 if the charge was successful
1966  * < 0 if the cgroup is over its limit
1967  */
1968 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1969                                 gfp_t gfp_mask, enum charge_type ctype,
1970                                 struct mem_cgroup *memcg)
1971 {
1972         struct mem_cgroup *mem;
1973         struct page_cgroup *pc;
1974         int ret;
1975
1976         pc = lookup_page_cgroup(page);
1977         /* can happen at boot */
1978         if (unlikely(!pc))
1979                 return 0;
1980         prefetchw(pc);
1981
1982         mem = memcg;
1983         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1984         if (ret || !mem)
1985                 return ret;
1986
1987         __mem_cgroup_commit_charge(mem, pc, ctype);
1988         return 0;
1989 }
1990
1991 int mem_cgroup_newpage_charge(struct page *page,
1992                               struct mm_struct *mm, gfp_t gfp_mask)
1993 {
1994         if (mem_cgroup_disabled())
1995                 return 0;
1996         if (PageCompound(page))
1997                 return 0;
1998         /*
1999          * If already mapped, we don't have to account.
2000          * If page cache, page->mapping has address_space.
2001          * But page->mapping may have out-of-use anon_vma pointer,
2002          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2003          * is NULL.
2004          */
2005         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2006                 return 0;
2007         if (unlikely(!mm))
2008                 mm = &init_mm;
2009         return mem_cgroup_charge_common(page, mm, gfp_mask,
2010                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
2011 }
2012
2013 static void
2014 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2015                                         enum charge_type ctype);
2016
2017 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2018                                 gfp_t gfp_mask)
2019 {
2020         struct mem_cgroup *mem = NULL;
2021         int ret;
2022
2023         if (mem_cgroup_disabled())
2024                 return 0;
2025         if (PageCompound(page))
2026                 return 0;
2027         /*
2028          * Corner case handling. This is called from add_to_page_cache()
2029          * in usual. But some FS (shmem) precharges this page before calling it
2030          * and call add_to_page_cache() with GFP_NOWAIT.
2031          *
2032          * For GFP_NOWAIT case, the page may be pre-charged before calling
2033          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2034          * charge twice. (It works but has to pay a bit larger cost.)
2035          * And when the page is SwapCache, it should take swap information
2036          * into account. This is under lock_page() now.
2037          */
2038         if (!(gfp_mask & __GFP_WAIT)) {
2039                 struct page_cgroup *pc;
2040
2041
2042                 pc = lookup_page_cgroup(page);
2043                 if (!pc)
2044                         return 0;
2045                 lock_page_cgroup(pc);
2046                 if (PageCgroupUsed(pc)) {
2047                         unlock_page_cgroup(pc);
2048                         return 0;
2049                 }
2050                 unlock_page_cgroup(pc);
2051         }
2052
2053         if (unlikely(!mm && !mem))
2054                 mm = &init_mm;
2055
2056         if (page_is_file_cache(page))
2057                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2058                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
2059
2060         /* shmem */
2061         if (PageSwapCache(page)) {
2062                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2063                 if (!ret)
2064                         __mem_cgroup_commit_charge_swapin(page, mem,
2065                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2066         } else
2067                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2068                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
2069
2070         return ret;
2071 }
2072
2073 /*
2074  * While swap-in, try_charge -> commit or cancel, the page is locked.
2075  * And when try_charge() successfully returns, one refcnt to memcg without
2076  * struct page_cgroup is acquired. This refcnt will be consumed by
2077  * "commit()" or removed by "cancel()"
2078  */
2079 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2080                                  struct page *page,
2081                                  gfp_t mask, struct mem_cgroup **ptr)
2082 {
2083         struct mem_cgroup *mem;
2084         int ret;
2085
2086         if (mem_cgroup_disabled())
2087                 return 0;
2088
2089         if (!do_swap_account)
2090                 goto charge_cur_mm;
2091         /*
2092          * A racing thread's fault, or swapoff, may have already updated
2093          * the pte, and even removed page from swap cache: in those cases
2094          * do_swap_page()'s pte_same() test will fail; but there's also a
2095          * KSM case which does need to charge the page.
2096          */
2097         if (!PageSwapCache(page))
2098                 goto charge_cur_mm;
2099         mem = try_get_mem_cgroup_from_page(page);
2100         if (!mem)
2101                 goto charge_cur_mm;
2102         *ptr = mem;
2103         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2104         /* drop extra refcnt from tryget */
2105         css_put(&mem->css);
2106         return ret;
2107 charge_cur_mm:
2108         if (unlikely(!mm))
2109                 mm = &init_mm;
2110         return __mem_cgroup_try_charge(mm, mask, ptr, true);
2111 }
2112
2113 static void
2114 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2115                                         enum charge_type ctype)
2116 {
2117         struct page_cgroup *pc;
2118
2119         if (mem_cgroup_disabled())
2120                 return;
2121         if (!ptr)
2122                 return;
2123         cgroup_exclude_rmdir(&ptr->css);
2124         pc = lookup_page_cgroup(page);
2125         mem_cgroup_lru_del_before_commit_swapcache(page);
2126         __mem_cgroup_commit_charge(ptr, pc, ctype);
2127         mem_cgroup_lru_add_after_commit_swapcache(page);
2128         /*
2129          * Now swap is on-memory. This means this page may be
2130          * counted both as mem and swap....double count.
2131          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2132          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2133          * may call delete_from_swap_cache() before reach here.
2134          */
2135         if (do_swap_account && PageSwapCache(page)) {
2136                 swp_entry_t ent = {.val = page_private(page)};
2137                 unsigned short id;
2138                 struct mem_cgroup *memcg;
2139
2140                 id = swap_cgroup_record(ent, 0);
2141                 rcu_read_lock();
2142                 memcg = mem_cgroup_lookup(id);
2143                 if (memcg) {
2144                         /*
2145                          * This recorded memcg can be obsolete one. So, avoid
2146                          * calling css_tryget
2147                          */
2148                         if (!mem_cgroup_is_root(memcg))
2149                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2150                         mem_cgroup_swap_statistics(memcg, false);
2151                         mem_cgroup_put(memcg);
2152                 }
2153                 rcu_read_unlock();
2154         }
2155         /*
2156          * At swapin, we may charge account against cgroup which has no tasks.
2157          * So, rmdir()->pre_destroy() can be called while we do this charge.
2158          * In that case, we need to call pre_destroy() again. check it here.
2159          */
2160         cgroup_release_and_wakeup_rmdir(&ptr->css);
2161 }
2162
2163 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2164 {
2165         __mem_cgroup_commit_charge_swapin(page, ptr,
2166                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2167 }
2168
2169 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2170 {
2171         if (mem_cgroup_disabled())
2172                 return;
2173         if (!mem)
2174                 return;
2175         mem_cgroup_cancel_charge(mem);
2176 }
2177
2178 static void
2179 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2180 {
2181         struct memcg_batch_info *batch = NULL;
2182         bool uncharge_memsw = true;
2183         /* If swapout, usage of swap doesn't decrease */
2184         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2185                 uncharge_memsw = false;
2186
2187         batch = &current->memcg_batch;
2188         /*
2189          * In usual, we do css_get() when we remember memcg pointer.
2190          * But in this case, we keep res->usage until end of a series of
2191          * uncharges. Then, it's ok to ignore memcg's refcnt.
2192          */
2193         if (!batch->memcg)
2194                 batch->memcg = mem;
2195         /*
2196          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2197          * In those cases, all pages freed continously can be expected to be in
2198          * the same cgroup and we have chance to coalesce uncharges.
2199          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2200          * because we want to do uncharge as soon as possible.
2201          */
2202
2203         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2204                 goto direct_uncharge;
2205
2206         /*
2207          * In typical case, batch->memcg == mem. This means we can
2208          * merge a series of uncharges to an uncharge of res_counter.
2209          * If not, we uncharge res_counter ony by one.
2210          */
2211         if (batch->memcg != mem)
2212                 goto direct_uncharge;
2213         /* remember freed charge and uncharge it later */
2214         batch->bytes += PAGE_SIZE;
2215         if (uncharge_memsw)
2216                 batch->memsw_bytes += PAGE_SIZE;
2217         return;
2218 direct_uncharge:
2219         res_counter_uncharge(&mem->res, PAGE_SIZE);
2220         if (uncharge_memsw)
2221                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2222         if (unlikely(batch->memcg != mem))
2223                 memcg_oom_recover(mem);
2224         return;
2225 }
2226
2227 /*
2228  * uncharge if !page_mapped(page)
2229  */
2230 static struct mem_cgroup *
2231 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2232 {
2233         struct page_cgroup *pc;
2234         struct mem_cgroup *mem = NULL;
2235         struct mem_cgroup_per_zone *mz;
2236
2237         if (mem_cgroup_disabled())
2238                 return NULL;
2239
2240         if (PageSwapCache(page))
2241                 return NULL;
2242
2243         /*
2244          * Check if our page_cgroup is valid
2245          */
2246         pc = lookup_page_cgroup(page);
2247         if (unlikely(!pc || !PageCgroupUsed(pc)))
2248                 return NULL;
2249
2250         lock_page_cgroup(pc);
2251
2252         mem = pc->mem_cgroup;
2253
2254         if (!PageCgroupUsed(pc))
2255                 goto unlock_out;
2256
2257         switch (ctype) {
2258         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2259         case MEM_CGROUP_CHARGE_TYPE_DROP:
2260                 /* See mem_cgroup_prepare_migration() */
2261                 if (page_mapped(page) || PageCgroupMigration(pc))
2262                         goto unlock_out;
2263                 break;
2264         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2265                 if (!PageAnon(page)) {  /* Shared memory */
2266                         if (page->mapping && !page_is_file_cache(page))
2267                                 goto unlock_out;
2268                 } else if (page_mapped(page)) /* Anon */
2269                                 goto unlock_out;
2270                 break;
2271         default:
2272                 break;
2273         }
2274
2275         if (!mem_cgroup_is_root(mem))
2276                 __do_uncharge(mem, ctype);
2277         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2278                 mem_cgroup_swap_statistics(mem, true);
2279         mem_cgroup_charge_statistics(mem, pc, false);
2280
2281         ClearPageCgroupUsed(pc);
2282         /*
2283          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2284          * freed from LRU. This is safe because uncharged page is expected not
2285          * to be reused (freed soon). Exception is SwapCache, it's handled by
2286          * special functions.
2287          */
2288
2289         mz = page_cgroup_zoneinfo(pc);
2290         unlock_page_cgroup(pc);
2291
2292         memcg_check_events(mem, page);
2293         /* at swapout, this memcg will be accessed to record to swap */
2294         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2295                 css_put(&mem->css);
2296
2297         return mem;
2298
2299 unlock_out:
2300         unlock_page_cgroup(pc);
2301         return NULL;
2302 }
2303
2304 void mem_cgroup_uncharge_page(struct page *page)
2305 {
2306         /* early check. */
2307         if (page_mapped(page))
2308                 return;
2309         if (page->mapping && !PageAnon(page))
2310                 return;
2311         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2312 }
2313
2314 void mem_cgroup_uncharge_cache_page(struct page *page)
2315 {
2316         VM_BUG_ON(page_mapped(page));
2317         VM_BUG_ON(page->mapping);
2318         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2319 }
2320
2321 /*
2322  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2323  * In that cases, pages are freed continuously and we can expect pages
2324  * are in the same memcg. All these calls itself limits the number of
2325  * pages freed at once, then uncharge_start/end() is called properly.
2326  * This may be called prural(2) times in a context,
2327  */
2328
2329 void mem_cgroup_uncharge_start(void)
2330 {
2331         current->memcg_batch.do_batch++;
2332         /* We can do nest. */
2333         if (current->memcg_batch.do_batch == 1) {
2334                 current->memcg_batch.memcg = NULL;
2335                 current->memcg_batch.bytes = 0;
2336                 current->memcg_batch.memsw_bytes = 0;
2337         }
2338 }
2339
2340 void mem_cgroup_uncharge_end(void)
2341 {
2342         struct memcg_batch_info *batch = &current->memcg_batch;
2343
2344         if (!batch->do_batch)
2345                 return;
2346
2347         batch->do_batch--;
2348         if (batch->do_batch) /* If stacked, do nothing. */
2349                 return;
2350
2351         if (!batch->memcg)
2352                 return;
2353         /*
2354          * This "batch->memcg" is valid without any css_get/put etc...
2355          * bacause we hide charges behind us.
2356          */
2357         if (batch->bytes)
2358                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2359         if (batch->memsw_bytes)
2360                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2361         memcg_oom_recover(batch->memcg);
2362         /* forget this pointer (for sanity check) */
2363         batch->memcg = NULL;
2364 }
2365
2366 #ifdef CONFIG_SWAP
2367 /*
2368  * called after __delete_from_swap_cache() and drop "page" account.
2369  * memcg information is recorded to swap_cgroup of "ent"
2370  */
2371 void
2372 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2373 {
2374         struct mem_cgroup *memcg;
2375         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2376
2377         if (!swapout) /* this was a swap cache but the swap is unused ! */
2378                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2379
2380         memcg = __mem_cgroup_uncharge_common(page, ctype);
2381
2382         /* record memcg information */
2383         if (do_swap_account && swapout && memcg) {
2384                 swap_cgroup_record(ent, css_id(&memcg->css));
2385                 mem_cgroup_get(memcg);
2386         }
2387         if (swapout && memcg)
2388                 css_put(&memcg->css);
2389 }
2390 #endif
2391
2392 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2393 /*
2394  * called from swap_entry_free(). remove record in swap_cgroup and
2395  * uncharge "memsw" account.
2396  */
2397 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2398 {
2399         struct mem_cgroup *memcg;
2400         unsigned short id;
2401
2402         if (!do_swap_account)
2403                 return;
2404
2405         id = swap_cgroup_record(ent, 0);
2406         rcu_read_lock();
2407         memcg = mem_cgroup_lookup(id);
2408         if (memcg) {
2409                 /*
2410                  * We uncharge this because swap is freed.
2411                  * This memcg can be obsolete one. We avoid calling css_tryget
2412                  */
2413                 if (!mem_cgroup_is_root(memcg))
2414                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2415                 mem_cgroup_swap_statistics(memcg, false);
2416                 mem_cgroup_put(memcg);
2417         }
2418         rcu_read_unlock();
2419 }
2420
2421 /**
2422  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2423  * @entry: swap entry to be moved
2424  * @from:  mem_cgroup which the entry is moved from
2425  * @to:  mem_cgroup which the entry is moved to
2426  * @need_fixup: whether we should fixup res_counters and refcounts.
2427  *
2428  * It succeeds only when the swap_cgroup's record for this entry is the same
2429  * as the mem_cgroup's id of @from.
2430  *
2431  * Returns 0 on success, -EINVAL on failure.
2432  *
2433  * The caller must have charged to @to, IOW, called res_counter_charge() about
2434  * both res and memsw, and called css_get().
2435  */
2436 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2437                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2438 {
2439         unsigned short old_id, new_id;
2440
2441         old_id = css_id(&from->css);
2442         new_id = css_id(&to->css);
2443
2444         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2445                 mem_cgroup_swap_statistics(from, false);
2446                 mem_cgroup_swap_statistics(to, true);
2447                 /*
2448                  * This function is only called from task migration context now.
2449                  * It postpones res_counter and refcount handling till the end
2450                  * of task migration(mem_cgroup_clear_mc()) for performance
2451                  * improvement. But we cannot postpone mem_cgroup_get(to)
2452                  * because if the process that has been moved to @to does
2453                  * swap-in, the refcount of @to might be decreased to 0.
2454                  */
2455                 mem_cgroup_get(to);
2456                 if (need_fixup) {
2457                         if (!mem_cgroup_is_root(from))
2458                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2459                         mem_cgroup_put(from);
2460                         /*
2461                          * we charged both to->res and to->memsw, so we should
2462                          * uncharge to->res.
2463                          */
2464                         if (!mem_cgroup_is_root(to))
2465                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2466                         css_put(&to->css);
2467                 }
2468                 return 0;
2469         }
2470         return -EINVAL;
2471 }
2472 #else
2473 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2474                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2475 {
2476         return -EINVAL;
2477 }
2478 #endif
2479
2480 /*
2481  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2482  * page belongs to.
2483  */
2484 int mem_cgroup_prepare_migration(struct page *page,
2485         struct page *newpage, struct mem_cgroup **ptr)
2486 {
2487         struct page_cgroup *pc;
2488         struct mem_cgroup *mem = NULL;
2489         enum charge_type ctype;
2490         int ret = 0;
2491
2492         if (mem_cgroup_disabled())
2493                 return 0;
2494
2495         pc = lookup_page_cgroup(page);
2496         lock_page_cgroup(pc);
2497         if (PageCgroupUsed(pc)) {
2498                 mem = pc->mem_cgroup;
2499                 css_get(&mem->css);
2500                 /*
2501                  * At migrating an anonymous page, its mapcount goes down
2502                  * to 0 and uncharge() will be called. But, even if it's fully
2503                  * unmapped, migration may fail and this page has to be
2504                  * charged again. We set MIGRATION flag here and delay uncharge
2505                  * until end_migration() is called
2506                  *
2507                  * Corner Case Thinking
2508                  * A)
2509                  * When the old page was mapped as Anon and it's unmap-and-freed
2510                  * while migration was ongoing.
2511                  * If unmap finds the old page, uncharge() of it will be delayed
2512                  * until end_migration(). If unmap finds a new page, it's
2513                  * uncharged when it make mapcount to be 1->0. If unmap code
2514                  * finds swap_migration_entry, the new page will not be mapped
2515                  * and end_migration() will find it(mapcount==0).
2516                  *
2517                  * B)
2518                  * When the old page was mapped but migraion fails, the kernel
2519                  * remaps it. A charge for it is kept by MIGRATION flag even
2520                  * if mapcount goes down to 0. We can do remap successfully
2521                  * without charging it again.
2522                  *
2523                  * C)
2524                  * The "old" page is under lock_page() until the end of
2525                  * migration, so, the old page itself will not be swapped-out.
2526                  * If the new page is swapped out before end_migraton, our
2527                  * hook to usual swap-out path will catch the event.
2528                  */
2529                 if (PageAnon(page))
2530                         SetPageCgroupMigration(pc);
2531         }
2532         unlock_page_cgroup(pc);
2533         /*
2534          * If the page is not charged at this point,
2535          * we return here.
2536          */
2537         if (!mem)
2538                 return 0;
2539
2540         *ptr = mem;
2541         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2542         css_put(&mem->css);/* drop extra refcnt */
2543         if (ret || *ptr == NULL) {
2544                 if (PageAnon(page)) {
2545                         lock_page_cgroup(pc);
2546                         ClearPageCgroupMigration(pc);
2547                         unlock_page_cgroup(pc);
2548                         /*
2549                          * The old page may be fully unmapped while we kept it.
2550                          */
2551                         mem_cgroup_uncharge_page(page);
2552                 }
2553                 return -ENOMEM;
2554         }
2555         /*
2556          * We charge new page before it's used/mapped. So, even if unlock_page()
2557          * is called before end_migration, we can catch all events on this new
2558          * page. In the case new page is migrated but not remapped, new page's
2559          * mapcount will be finally 0 and we call uncharge in end_migration().
2560          */
2561         pc = lookup_page_cgroup(newpage);
2562         if (PageAnon(page))
2563                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2564         else if (page_is_file_cache(page))
2565                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2566         else
2567                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2568         __mem_cgroup_commit_charge(mem, pc, ctype);
2569         return ret;
2570 }
2571
2572 /* remove redundant charge if migration failed*/
2573 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2574         struct page *oldpage, struct page *newpage)
2575 {
2576         struct page *used, *unused;
2577         struct page_cgroup *pc;
2578
2579         if (!mem)
2580                 return;
2581         /* blocks rmdir() */
2582         cgroup_exclude_rmdir(&mem->css);
2583         /* at migration success, oldpage->mapping is NULL. */
2584         if (oldpage->mapping) {
2585                 used = oldpage;
2586                 unused = newpage;
2587         } else {
2588                 used = newpage;
2589                 unused = oldpage;
2590         }
2591         /*
2592          * We disallowed uncharge of pages under migration because mapcount
2593          * of the page goes down to zero, temporarly.
2594          * Clear the flag and check the page should be charged.
2595          */
2596         pc = lookup_page_cgroup(oldpage);
2597         lock_page_cgroup(pc);
2598         ClearPageCgroupMigration(pc);
2599         unlock_page_cgroup(pc);
2600
2601         if (unused != oldpage)
2602                 pc = lookup_page_cgroup(unused);
2603         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2604
2605         pc = lookup_page_cgroup(used);
2606         /*
2607          * If a page is a file cache, radix-tree replacement is very atomic
2608          * and we can skip this check. When it was an Anon page, its mapcount
2609          * goes down to 0. But because we added MIGRATION flage, it's not
2610          * uncharged yet. There are several case but page->mapcount check
2611          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2612          * check. (see prepare_charge() also)
2613          */
2614         if (PageAnon(used))
2615                 mem_cgroup_uncharge_page(used);
2616         /*
2617          * At migration, we may charge account against cgroup which has no
2618          * tasks.
2619          * So, rmdir()->pre_destroy() can be called while we do this charge.
2620          * In that case, we need to call pre_destroy() again. check it here.
2621          */
2622         cgroup_release_and_wakeup_rmdir(&mem->css);
2623 }
2624
2625 /*
2626  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2627  * Calling hierarchical_reclaim is not enough because we should update
2628  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2629  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2630  * not from the memcg which this page would be charged to.
2631  * try_charge_swapin does all of these works properly.
2632  */
2633 int mem_cgroup_shmem_charge_fallback(struct page *page,
2634                             struct mm_struct *mm,
2635                             gfp_t gfp_mask)
2636 {
2637         struct mem_cgroup *mem = NULL;
2638         int ret;
2639
2640         if (mem_cgroup_disabled())
2641                 return 0;
2642
2643         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2644         if (!ret)
2645                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2646
2647         return ret;
2648 }
2649
2650 static DEFINE_MUTEX(set_limit_mutex);
2651
2652 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2653                                 unsigned long long val)
2654 {
2655         int retry_count;
2656         u64 memswlimit, memlimit;
2657         int ret = 0;
2658         int children = mem_cgroup_count_children(memcg);
2659         u64 curusage, oldusage;
2660         int enlarge;
2661
2662         /*
2663          * For keeping hierarchical_reclaim simple, how long we should retry
2664          * is depends on callers. We set our retry-count to be function
2665          * of # of children which we should visit in this loop.
2666          */
2667         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2668
2669         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2670
2671         enlarge = 0;
2672         while (retry_count) {
2673                 if (signal_pending(current)) {
2674                         ret = -EINTR;
2675                         break;
2676                 }
2677                 /*
2678                  * Rather than hide all in some function, I do this in
2679                  * open coded manner. You see what this really does.
2680                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2681                  */
2682                 mutex_lock(&set_limit_mutex);
2683                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2684                 if (memswlimit < val) {
2685                         ret = -EINVAL;
2686                         mutex_unlock(&set_limit_mutex);
2687                         break;
2688                 }
2689
2690                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2691                 if (memlimit < val)
2692                         enlarge = 1;
2693
2694                 ret = res_counter_set_limit(&memcg->res, val);
2695                 if (!ret) {
2696                         if (memswlimit == val)
2697                                 memcg->memsw_is_minimum = true;
2698                         else
2699                                 memcg->memsw_is_minimum = false;
2700                 }
2701                 mutex_unlock(&set_limit_mutex);
2702
2703                 if (!ret)
2704                         break;
2705
2706                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2707                                                 MEM_CGROUP_RECLAIM_SHRINK);
2708                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2709                 /* Usage is reduced ? */
2710                 if (curusage >= oldusage)
2711                         retry_count--;
2712                 else
2713                         oldusage = curusage;
2714         }
2715         if (!ret && enlarge)
2716                 memcg_oom_recover(memcg);
2717
2718         return ret;
2719 }
2720
2721 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2722                                         unsigned long long val)
2723 {
2724         int retry_count;
2725         u64 memlimit, memswlimit, oldusage, curusage;
2726         int children = mem_cgroup_count_children(memcg);
2727         int ret = -EBUSY;
2728         int enlarge = 0;
2729
2730         /* see mem_cgroup_resize_res_limit */
2731         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2732         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2733         while (retry_count) {
2734                 if (signal_pending(current)) {
2735                         ret = -EINTR;
2736                         break;
2737                 }
2738                 /*
2739                  * Rather than hide all in some function, I do this in
2740                  * open coded manner. You see what this really does.
2741                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2742                  */
2743                 mutex_lock(&set_limit_mutex);
2744                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2745                 if (memlimit > val) {
2746                         ret = -EINVAL;
2747                         mutex_unlock(&set_limit_mutex);
2748                         break;
2749                 }
2750                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2751                 if (memswlimit < val)
2752                         enlarge = 1;
2753                 ret = res_counter_set_limit(&memcg->memsw, val);
2754                 if (!ret) {
2755                         if (memlimit == val)
2756                                 memcg->memsw_is_minimum = true;
2757                         else
2758                                 memcg->memsw_is_minimum = false;
2759                 }
2760                 mutex_unlock(&set_limit_mutex);
2761
2762                 if (!ret)
2763                         break;
2764
2765                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2766                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2767                                                 MEM_CGROUP_RECLAIM_SHRINK);
2768                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2769                 /* Usage is reduced ? */
2770                 if (curusage >= oldusage)
2771                         retry_count--;
2772                 else
2773                         oldusage = curusage;
2774         }
2775         if (!ret && enlarge)
2776                 memcg_oom_recover(memcg);
2777         return ret;
2778 }
2779
2780 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2781                                                 gfp_t gfp_mask, int nid,
2782                                                 int zid)
2783 {
2784         unsigned long nr_reclaimed = 0;
2785         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2786         unsigned long reclaimed;
2787         int loop = 0;
2788         struct mem_cgroup_tree_per_zone *mctz;
2789         unsigned long long excess;
2790
2791         if (order > 0)
2792                 return 0;
2793
2794         mctz = soft_limit_tree_node_zone(nid, zid);
2795         /*
2796          * This loop can run a while, specially if mem_cgroup's continuously
2797          * keep exceeding their soft limit and putting the system under
2798          * pressure
2799          */
2800         do {
2801                 if (next_mz)
2802                         mz = next_mz;
2803                 else
2804                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2805                 if (!mz)
2806                         break;
2807
2808                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2809                                                 gfp_mask,
2810                                                 MEM_CGROUP_RECLAIM_SOFT);
2811                 nr_reclaimed += reclaimed;
2812                 spin_lock(&mctz->lock);
2813
2814                 /*
2815                  * If we failed to reclaim anything from this memory cgroup
2816                  * it is time to move on to the next cgroup
2817                  */
2818                 next_mz = NULL;
2819                 if (!reclaimed) {
2820                         do {
2821                                 /*
2822                                  * Loop until we find yet another one.
2823                                  *
2824                                  * By the time we get the soft_limit lock
2825                                  * again, someone might have aded the
2826                                  * group back on the RB tree. Iterate to
2827                                  * make sure we get a different mem.
2828                                  * mem_cgroup_largest_soft_limit_node returns
2829                                  * NULL if no other cgroup is present on
2830                                  * the tree
2831                                  */
2832                                 next_mz =
2833                                 __mem_cgroup_largest_soft_limit_node(mctz);
2834                                 if (next_mz == mz) {
2835                                         css_put(&next_mz->mem->css);
2836                                         next_mz = NULL;
2837                                 } else /* next_mz == NULL or other memcg */
2838                                         break;
2839                         } while (1);
2840                 }
2841                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2842                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2843                 /*
2844                  * One school of thought says that we should not add
2845                  * back the node to the tree if reclaim returns 0.
2846                  * But our reclaim could return 0, simply because due
2847                  * to priority we are exposing a smaller subset of
2848                  * memory to reclaim from. Consider this as a longer
2849                  * term TODO.
2850                  */
2851                 /* If excess == 0, no tree ops */
2852                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2853                 spin_unlock(&mctz->lock);
2854                 css_put(&mz->mem->css);
2855                 loop++;
2856                 /*
2857                  * Could not reclaim anything and there are no more
2858                  * mem cgroups to try or we seem to be looping without
2859                  * reclaiming anything.
2860                  */
2861                 if (!nr_reclaimed &&
2862                         (next_mz == NULL ||
2863                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2864                         break;
2865         } while (!nr_reclaimed);
2866         if (next_mz)
2867                 css_put(&next_mz->mem->css);
2868         return nr_reclaimed;
2869 }
2870
2871 /*
2872  * This routine traverse page_cgroup in given list and drop them all.
2873  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2874  */
2875 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2876                                 int node, int zid, enum lru_list lru)
2877 {
2878         struct zone *zone;
2879         struct mem_cgroup_per_zone *mz;
2880         struct page_cgroup *pc, *busy;
2881         unsigned long flags, loop;
2882         struct list_head *list;
2883         int ret = 0;
2884
2885         zone = &NODE_DATA(node)->node_zones[zid];
2886         mz = mem_cgroup_zoneinfo(mem, node, zid);
2887         list = &mz->lists[lru];
2888
2889         loop = MEM_CGROUP_ZSTAT(mz, lru);
2890         /* give some margin against EBUSY etc...*/
2891         loop += 256;
2892         busy = NULL;
2893         while (loop--) {
2894                 ret = 0;
2895                 spin_lock_irqsave(&zone->lru_lock, flags);
2896                 if (list_empty(list)) {
2897                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2898                         break;
2899                 }
2900                 pc = list_entry(list->prev, struct page_cgroup, lru);
2901                 if (busy == pc) {
2902                         list_move(&pc->lru, list);
2903                         busy = NULL;
2904                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2905                         continue;
2906                 }
2907                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2908
2909                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2910                 if (ret == -ENOMEM)
2911                         break;
2912
2913                 if (ret == -EBUSY || ret == -EINVAL) {
2914                         /* found lock contention or "pc" is obsolete. */
2915                         busy = pc;
2916                         cond_resched();
2917                 } else
2918                         busy = NULL;
2919         }
2920
2921         if (!ret && !list_empty(list))
2922                 return -EBUSY;
2923         return ret;
2924 }
2925
2926 /*
2927  * make mem_cgroup's charge to be 0 if there is no task.
2928  * This enables deleting this mem_cgroup.
2929  */
2930 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2931 {
2932         int ret;
2933         int node, zid, shrink;
2934         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2935         struct cgroup *cgrp = mem->css.cgroup;
2936
2937         css_get(&mem->css);
2938
2939         shrink = 0;
2940         /* should free all ? */
2941         if (free_all)
2942                 goto try_to_free;
2943 move_account:
2944         do {
2945                 ret = -EBUSY;
2946                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2947                         goto out;
2948                 ret = -EINTR;
2949                 if (signal_pending(current))
2950                         goto out;
2951                 /* This is for making all *used* pages to be on LRU. */
2952                 lru_add_drain_all();
2953                 drain_all_stock_sync();
2954                 ret = 0;
2955                 for_each_node_state(node, N_HIGH_MEMORY) {
2956                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2957                                 enum lru_list l;
2958                                 for_each_lru(l) {
2959                                         ret = mem_cgroup_force_empty_list(mem,
2960                                                         node, zid, l);
2961                                         if (ret)
2962                                                 break;
2963                                 }
2964                         }
2965                         if (ret)
2966                                 break;
2967                 }
2968                 memcg_oom_recover(mem);
2969                 /* it seems parent cgroup doesn't have enough mem */
2970                 if (ret == -ENOMEM)
2971                         goto try_to_free;
2972                 cond_resched();
2973         /* "ret" should also be checked to ensure all lists are empty. */
2974         } while (mem->res.usage > 0 || ret);
2975 out:
2976         css_put(&mem->css);
2977         return ret;
2978
2979 try_to_free:
2980         /* returns EBUSY if there is a task or if we come here twice. */
2981         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2982                 ret = -EBUSY;
2983                 goto out;
2984         }
2985         /* we call try-to-free pages for make this cgroup empty */
2986         lru_add_drain_all();
2987         /* try to free all pages in this cgroup */
2988         shrink = 1;
2989         while (nr_retries && mem->res.usage > 0) {
2990                 int progress;
2991
2992                 if (signal_pending(current)) {
2993                         ret = -EINTR;
2994                         goto out;
2995                 }
2996                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2997                                                 false, get_swappiness(mem));
2998                 if (!progress) {
2999                         nr_retries--;
3000                         /* maybe some writeback is necessary */
3001                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3002                 }
3003
3004         }
3005         lru_add_drain();
3006         /* try move_account...there may be some *locked* pages. */
3007         goto move_account;
3008 }
3009
3010 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3011 {
3012         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3013 }
3014
3015
3016 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3017 {
3018         return mem_cgroup_from_cont(cont)->use_hierarchy;
3019 }
3020
3021 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3022                                         u64 val)
3023 {
3024         int retval = 0;
3025         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3026         struct cgroup *parent = cont->parent;
3027         struct mem_cgroup *parent_mem = NULL;
3028
3029         if (parent)
3030                 parent_mem = mem_cgroup_from_cont(parent);
3031
3032         cgroup_lock();
3033         /*
3034          * If parent's use_hierarchy is set, we can't make any modifications
3035          * in the child subtrees. If it is unset, then the change can
3036          * occur, provided the current cgroup has no children.
3037          *
3038          * For the root cgroup, parent_mem is NULL, we allow value to be
3039          * set if there are no children.
3040          */
3041         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3042                                 (val == 1 || val == 0)) {
3043                 if (list_empty(&cont->children))
3044                         mem->use_hierarchy = val;
3045                 else
3046                         retval = -EBUSY;
3047         } else
3048                 retval = -EINVAL;
3049         cgroup_unlock();
3050
3051         return retval;
3052 }
3053
3054 struct mem_cgroup_idx_data {
3055         s64 val;
3056         enum mem_cgroup_stat_index idx;
3057 };
3058
3059 static int
3060 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
3061 {
3062         struct mem_cgroup_idx_data *d = data;
3063         d->val += mem_cgroup_read_stat(mem, d->idx);
3064         return 0;
3065 }
3066
3067 static void
3068 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3069                                 enum mem_cgroup_stat_index idx, s64 *val)
3070 {
3071         struct mem_cgroup_idx_data d;
3072         d.idx = idx;
3073         d.val = 0;
3074         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
3075         *val = d.val;
3076 }
3077
3078 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3079 {
3080         u64 idx_val, val;
3081
3082         if (!mem_cgroup_is_root(mem)) {
3083                 if (!swap)
3084                         return res_counter_read_u64(&mem->res, RES_USAGE);
3085                 else
3086                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3087         }
3088
3089         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
3090         val = idx_val;
3091         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
3092         val += idx_val;
3093
3094         if (swap) {
3095                 mem_cgroup_get_recursive_idx_stat(mem,
3096                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
3097                 val += idx_val;
3098         }
3099
3100         return val << PAGE_SHIFT;
3101 }
3102
3103 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3104 {
3105         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3106         u64 val;
3107         int type, name;
3108
3109         type = MEMFILE_TYPE(cft->private);
3110         name = MEMFILE_ATTR(cft->private);
3111         switch (type) {
3112         case _MEM:
3113                 if (name == RES_USAGE)
3114                         val = mem_cgroup_usage(mem, false);
3115                 else
3116                         val = res_counter_read_u64(&mem->res, name);
3117                 break;
3118         case _MEMSWAP:
3119                 if (name == RES_USAGE)
3120                         val = mem_cgroup_usage(mem, true);
3121                 else
3122                         val = res_counter_read_u64(&mem->memsw, name);
3123                 break;
3124         default:
3125                 BUG();
3126                 break;
3127         }
3128         return val;
3129 }
3130 /*
3131  * The user of this function is...
3132  * RES_LIMIT.
3133  */
3134 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3135                             const char *buffer)
3136 {
3137         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3138         int type, name;
3139         unsigned long long val;
3140         int ret;
3141
3142         type = MEMFILE_TYPE(cft->private);
3143         name = MEMFILE_ATTR(cft->private);
3144         switch (name) {
3145         case RES_LIMIT:
3146                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3147                         ret = -EINVAL;
3148                         break;
3149                 }
3150                 /* This function does all necessary parse...reuse it */
3151                 ret = res_counter_memparse_write_strategy(buffer, &val);
3152                 if (ret)
3153                         break;
3154                 if (type == _MEM)
3155                         ret = mem_cgroup_resize_limit(memcg, val);
3156                 else
3157                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3158                 break;
3159         case RES_SOFT_LIMIT:
3160                 ret = res_counter_memparse_write_strategy(buffer, &val);
3161                 if (ret)
3162                         break;
3163                 /*
3164                  * For memsw, soft limits are hard to implement in terms
3165                  * of semantics, for now, we support soft limits for
3166                  * control without swap
3167                  */
3168                 if (type == _MEM)
3169                         ret = res_counter_set_soft_limit(&memcg->res, val);
3170                 else
3171                         ret = -EINVAL;
3172                 break;
3173         default:
3174                 ret = -EINVAL; /* should be BUG() ? */
3175                 break;
3176         }
3177         return ret;
3178 }
3179
3180 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3181                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3182 {
3183         struct cgroup *cgroup;
3184         unsigned long long min_limit, min_memsw_limit, tmp;
3185
3186         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3187         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3188         cgroup = memcg->css.cgroup;
3189         if (!memcg->use_hierarchy)
3190                 goto out;
3191
3192         while (cgroup->parent) {
3193                 cgroup = cgroup->parent;
3194                 memcg = mem_cgroup_from_cont(cgroup);
3195                 if (!memcg->use_hierarchy)
3196                         break;
3197                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3198                 min_limit = min(min_limit, tmp);
3199                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3200                 min_memsw_limit = min(min_memsw_limit, tmp);
3201         }
3202 out:
3203         *mem_limit = min_limit;
3204         *memsw_limit = min_memsw_limit;
3205         return;
3206 }
3207
3208 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3209 {
3210         struct mem_cgroup *mem;
3211         int type, name;
3212
3213         mem = mem_cgroup_from_cont(cont);
3214         type = MEMFILE_TYPE(event);
3215         name = MEMFILE_ATTR(event);
3216         switch (name) {
3217         case RES_MAX_USAGE:
3218                 if (type == _MEM)
3219                         res_counter_reset_max(&mem->res);
3220                 else
3221                         res_counter_reset_max(&mem->memsw);
3222                 break;
3223         case RES_FAILCNT:
3224                 if (type == _MEM)
3225                         res_counter_reset_failcnt(&mem->res);
3226                 else
3227                         res_counter_reset_failcnt(&mem->memsw);
3228                 break;
3229         }
3230
3231         return 0;
3232 }
3233
3234 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3235                                         struct cftype *cft)
3236 {
3237         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3238 }
3239
3240 #ifdef CONFIG_MMU
3241 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3242                                         struct cftype *cft, u64 val)
3243 {
3244         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3245
3246         if (val >= (1 << NR_MOVE_TYPE))
3247                 return -EINVAL;
3248         /*
3249          * We check this value several times in both in can_attach() and
3250          * attach(), so we need cgroup lock to prevent this value from being
3251          * inconsistent.
3252          */
3253         cgroup_lock();
3254         mem->move_charge_at_immigrate = val;
3255         cgroup_unlock();
3256
3257         return 0;
3258 }
3259 #else
3260 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3261                                         struct cftype *cft, u64 val)
3262 {
3263         return -ENOSYS;
3264 }
3265 #endif
3266
3267
3268 /* For read statistics */
3269 enum {
3270         MCS_CACHE,
3271         MCS_RSS,
3272         MCS_FILE_MAPPED,
3273         MCS_PGPGIN,
3274         MCS_PGPGOUT,
3275         MCS_SWAP,
3276         MCS_INACTIVE_ANON,
3277         MCS_ACTIVE_ANON,
3278         MCS_INACTIVE_FILE,
3279         MCS_ACTIVE_FILE,
3280         MCS_UNEVICTABLE,
3281         NR_MCS_STAT,
3282 };
3283
3284 struct mcs_total_stat {
3285         s64 stat[NR_MCS_STAT];
3286 };
3287
3288 struct {
3289         char *local_name;
3290         char *total_name;
3291 } memcg_stat_strings[NR_MCS_STAT] = {
3292         {"cache", "total_cache"},
3293         {"rss", "total_rss"},
3294         {"mapped_file", "total_mapped_file"},
3295         {"pgpgin", "total_pgpgin"},
3296         {"pgpgout", "total_pgpgout"},
3297         {"swap", "total_swap"},
3298         {"inactive_anon", "total_inactive_anon"},
3299         {"active_anon", "total_active_anon"},
3300         {"inactive_file", "total_inactive_file"},
3301         {"active_file", "total_active_file"},
3302         {"unevictable", "total_unevictable"}
3303 };
3304
3305
3306 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3307 {
3308         struct mcs_total_stat *s = data;
3309         s64 val;
3310
3311         /* per cpu stat */
3312         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3313         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3314         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3315         s->stat[MCS_RSS] += val * PAGE_SIZE;
3316         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3317         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3318         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3319         s->stat[MCS_PGPGIN] += val;
3320         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3321         s->stat[MCS_PGPGOUT] += val;
3322         if (do_swap_account) {
3323                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3324                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3325         }
3326
3327         /* per zone stat */
3328         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3329         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3330         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3331         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3332         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3333         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3334         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3335         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3336         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3337         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3338         return 0;
3339 }
3340
3341 static void
3342 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3343 {
3344         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3345 }
3346
3347 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3348                                  struct cgroup_map_cb *cb)
3349 {
3350         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3351         struct mcs_total_stat mystat;
3352         int i;
3353
3354         memset(&mystat, 0, sizeof(mystat));
3355         mem_cgroup_get_local_stat(mem_cont, &mystat);
3356
3357         for (i = 0; i < NR_MCS_STAT; i++) {
3358                 if (i == MCS_SWAP && !do_swap_account)
3359                         continue;
3360                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3361         }
3362
3363         /* Hierarchical information */
3364         {
3365                 unsigned long long limit, memsw_limit;
3366                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3367                 cb->fill(cb, "hierarchical_memory_limit", limit);
3368                 if (do_swap_account)
3369                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3370         }
3371
3372         memset(&mystat, 0, sizeof(mystat));
3373         mem_cgroup_get_total_stat(mem_cont, &mystat);
3374         for (i = 0; i < NR_MCS_STAT; i++) {
3375                 if (i == MCS_SWAP && !do_swap_account)
3376                         continue;
3377                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3378         }
3379
3380 #ifdef CONFIG_DEBUG_VM
3381         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3382
3383         {
3384                 int nid, zid;
3385                 struct mem_cgroup_per_zone *mz;
3386                 unsigned long recent_rotated[2] = {0, 0};
3387                 unsigned long recent_scanned[2] = {0, 0};
3388
3389                 for_each_online_node(nid)
3390                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3391                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3392
3393                                 recent_rotated[0] +=
3394                                         mz->reclaim_stat.recent_rotated[0];
3395                                 recent_rotated[1] +=
3396                                         mz->reclaim_stat.recent_rotated[1];
3397                                 recent_scanned[0] +=
3398                                         mz->reclaim_stat.recent_scanned[0];
3399                                 recent_scanned[1] +=
3400                                         mz->reclaim_stat.recent_scanned[1];
3401                         }
3402                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3403                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3404                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3405                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3406         }
3407 #endif
3408
3409         return 0;
3410 }
3411
3412 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3413 {
3414         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3415
3416         return get_swappiness(memcg);
3417 }
3418
3419 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3420                                        u64 val)
3421 {
3422         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3423         struct mem_cgroup *parent;
3424
3425         if (val > 100)
3426                 return -EINVAL;
3427
3428         if (cgrp->parent == NULL)
3429                 return -EINVAL;
3430
3431         parent = mem_cgroup_from_cont(cgrp->parent);
3432
3433         cgroup_lock();
3434
3435         /* If under hierarchy, only empty-root can set this value */
3436         if ((parent->use_hierarchy) ||
3437             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3438                 cgroup_unlock();
3439                 return -EINVAL;
3440         }
3441
3442         spin_lock(&memcg->reclaim_param_lock);
3443         memcg->swappiness = val;
3444         spin_unlock(&memcg->reclaim_param_lock);
3445
3446         cgroup_unlock();
3447
3448         return 0;
3449 }
3450
3451 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3452 {
3453         struct mem_cgroup_threshold_ary *t;
3454         u64 usage;
3455         int i;
3456
3457         rcu_read_lock();
3458         if (!swap)
3459                 t = rcu_dereference(memcg->thresholds.primary);
3460         else
3461                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3462
3463         if (!t)
3464                 goto unlock;
3465
3466         usage = mem_cgroup_usage(memcg, swap);
3467
3468         /*
3469          * current_threshold points to threshold just below usage.
3470          * If it's not true, a threshold was crossed after last
3471          * call of __mem_cgroup_threshold().
3472          */
3473         i = t->current_threshold;
3474
3475         /*
3476          * Iterate backward over array of thresholds starting from
3477          * current_threshold and check if a threshold is crossed.
3478          * If none of thresholds below usage is crossed, we read
3479          * only one element of the array here.
3480          */
3481         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3482                 eventfd_signal(t->entries[i].eventfd, 1);
3483
3484         /* i = current_threshold + 1 */
3485         i++;
3486
3487         /*
3488          * Iterate forward over array of thresholds starting from
3489          * current_threshold+1 and check if a threshold is crossed.
3490          * If none of thresholds above usage is crossed, we read
3491          * only one element of the array here.
3492          */
3493         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3494                 eventfd_signal(t->entries[i].eventfd, 1);
3495
3496         /* Update current_threshold */
3497         t->current_threshold = i - 1;
3498 unlock:
3499         rcu_read_unlock();
3500 }
3501
3502 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3503 {
3504         __mem_cgroup_threshold(memcg, false);
3505         if (do_swap_account)
3506                 __mem_cgroup_threshold(memcg, true);
3507 }
3508
3509 static int compare_thresholds(const void *a, const void *b)
3510 {
3511         const struct mem_cgroup_threshold *_a = a;
3512         const struct mem_cgroup_threshold *_b = b;
3513
3514         return _a->threshold - _b->threshold;
3515 }
3516
3517 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem, void *data)
3518 {
3519         struct mem_cgroup_eventfd_list *ev;
3520
3521         list_for_each_entry(ev, &mem->oom_notify, list)
3522                 eventfd_signal(ev->eventfd, 1);
3523         return 0;
3524 }
3525
3526 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3527 {
3528         mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_notify_cb);
3529 }
3530
3531 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3532         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3533 {
3534         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3535         struct mem_cgroup_thresholds *thresholds;
3536         struct mem_cgroup_threshold_ary *new;
3537         int type = MEMFILE_TYPE(cft->private);
3538         u64 threshold, usage;
3539         int i, size, ret;
3540
3541         ret = res_counter_memparse_write_strategy(args, &threshold);
3542         if (ret)
3543                 return ret;
3544
3545         mutex_lock(&memcg->thresholds_lock);
3546
3547         if (type == _MEM)
3548                 thresholds = &memcg->thresholds;
3549         else if (type == _MEMSWAP)
3550                 thresholds = &memcg->memsw_thresholds;
3551         else
3552                 BUG();
3553
3554         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3555
3556         /* Check if a threshold crossed before adding a new one */
3557         if (thresholds->primary)
3558                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3559
3560         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3561
3562         /* Allocate memory for new array of thresholds */
3563         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3564                         GFP_KERNEL);
3565         if (!new) {
3566                 ret = -ENOMEM;
3567                 goto unlock;
3568         }
3569         new->size = size;
3570
3571         /* Copy thresholds (if any) to new array */
3572         if (thresholds->primary) {
3573                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3574                                 sizeof(struct mem_cgroup_threshold));
3575         }
3576
3577         /* Add new threshold */
3578         new->entries[size - 1].eventfd = eventfd;
3579         new->entries[size - 1].threshold = threshold;
3580
3581         /* Sort thresholds. Registering of new threshold isn't time-critical */
3582         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3583                         compare_thresholds, NULL);
3584
3585         /* Find current threshold */
3586         new->current_threshold = -1;
3587         for (i = 0; i < size; i++) {
3588                 if (new->entries[i].threshold < usage) {
3589                         /*
3590                          * new->current_threshold will not be used until
3591                          * rcu_assign_pointer(), so it's safe to increment
3592                          * it here.
3593                          */
3594                         ++new->current_threshold;
3595                 }
3596         }
3597
3598         /* Free old spare buffer and save old primary buffer as spare */
3599         kfree(thresholds->spare);
3600         thresholds->spare = thresholds->primary;
3601
3602         rcu_assign_pointer(thresholds->primary, new);
3603
3604         /* To be sure that nobody uses thresholds */
3605         synchronize_rcu();
3606
3607 unlock:
3608         mutex_unlock(&memcg->thresholds_lock);
3609
3610         return ret;
3611 }
3612
3613 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3614         struct cftype *cft, struct eventfd_ctx *eventfd)
3615 {
3616         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3617         struct mem_cgroup_thresholds *thresholds;
3618         struct mem_cgroup_threshold_ary *new;
3619         int type = MEMFILE_TYPE(cft->private);
3620         u64 usage;
3621         int i, j, size;
3622
3623         mutex_lock(&memcg->thresholds_lock);
3624         if (type == _MEM)
3625                 thresholds = &memcg->thresholds;
3626         else if (type == _MEMSWAP)
3627                 thresholds = &memcg->memsw_thresholds;
3628         else
3629                 BUG();
3630
3631         /*
3632          * Something went wrong if we trying to unregister a threshold
3633          * if we don't have thresholds
3634          */
3635         BUG_ON(!thresholds);
3636
3637         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3638
3639         /* Check if a threshold crossed before removing */
3640         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3641
3642         /* Calculate new number of threshold */
3643         size = 0;
3644         for (i = 0; i < thresholds->primary->size; i++) {
3645                 if (thresholds->primary->entries[i].eventfd != eventfd)
3646                         size++;
3647         }
3648
3649         new = thresholds->spare;
3650
3651         /* Set thresholds array to NULL if we don't have thresholds */
3652         if (!size) {
3653                 kfree(new);
3654                 new = NULL;
3655                 goto swap_buffers;
3656         }
3657
3658         new->size = size;
3659
3660         /* Copy thresholds and find current threshold */
3661         new->current_threshold = -1;
3662         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3663                 if (thresholds->primary->entries[i].eventfd == eventfd)
3664                         continue;
3665
3666                 new->entries[j] = thresholds->primary->entries[i];
3667                 if (new->entries[j].threshold < usage) {
3668                         /*
3669                          * new->current_threshold will not be used
3670                          * until rcu_assign_pointer(), so it's safe to increment
3671                          * it here.
3672                          */
3673                         ++new->current_threshold;
3674                 }
3675                 j++;
3676         }
3677
3678 swap_buffers:
3679         /* Swap primary and spare array */
3680         thresholds->spare = thresholds->primary;
3681         rcu_assign_pointer(thresholds->primary, new);
3682
3683         /* To be sure that nobody uses thresholds */
3684         synchronize_rcu();
3685
3686         mutex_unlock(&memcg->thresholds_lock);
3687 }
3688
3689 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3690         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3691 {
3692         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3693         struct mem_cgroup_eventfd_list *event;
3694         int type = MEMFILE_TYPE(cft->private);
3695
3696         BUG_ON(type != _OOM_TYPE);
3697         event = kmalloc(sizeof(*event), GFP_KERNEL);
3698         if (!event)
3699                 return -ENOMEM;
3700
3701         mutex_lock(&memcg_oom_mutex);
3702
3703         event->eventfd = eventfd;
3704         list_add(&event->list, &memcg->oom_notify);
3705
3706         /* already in OOM ? */
3707         if (atomic_read(&memcg->oom_lock))
3708                 eventfd_signal(eventfd, 1);
3709         mutex_unlock(&memcg_oom_mutex);
3710
3711         return 0;
3712 }
3713
3714 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3715         struct cftype *cft, struct eventfd_ctx *eventfd)
3716 {
3717         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3718         struct mem_cgroup_eventfd_list *ev, *tmp;
3719         int type = MEMFILE_TYPE(cft->private);
3720
3721         BUG_ON(type != _OOM_TYPE);
3722
3723         mutex_lock(&memcg_oom_mutex);
3724
3725         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3726                 if (ev->eventfd == eventfd) {
3727                         list_del(&ev->list);
3728                         kfree(ev);
3729                 }
3730         }
3731
3732         mutex_unlock(&memcg_oom_mutex);
3733 }
3734
3735 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
3736         struct cftype *cft,  struct cgroup_map_cb *cb)
3737 {
3738         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3739
3740         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
3741
3742         if (atomic_read(&mem->oom_lock))
3743                 cb->fill(cb, "under_oom", 1);
3744         else
3745                 cb->fill(cb, "under_oom", 0);
3746         return 0;
3747 }
3748
3749 /*
3750  */
3751 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
3752         struct cftype *cft, u64 val)
3753 {
3754         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3755         struct mem_cgroup *parent;
3756
3757         /* cannot set to root cgroup and only 0 and 1 are allowed */
3758         if (!cgrp->parent || !((val == 0) || (val == 1)))
3759                 return -EINVAL;
3760
3761         parent = mem_cgroup_from_cont(cgrp->parent);
3762
3763         cgroup_lock();
3764         /* oom-kill-disable is a flag for subhierarchy. */
3765         if ((parent->use_hierarchy) ||
3766             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
3767                 cgroup_unlock();
3768                 return -EINVAL;
3769         }
3770         mem->oom_kill_disable = val;
3771         if (!val)
3772                 memcg_oom_recover(mem);
3773         cgroup_unlock();
3774         return 0;
3775 }
3776
3777 static struct cftype mem_cgroup_files[] = {
3778         {
3779                 .name = "usage_in_bytes",
3780                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3781                 .read_u64 = mem_cgroup_read,
3782                 .register_event = mem_cgroup_usage_register_event,
3783                 .unregister_event = mem_cgroup_usage_unregister_event,
3784         },
3785         {
3786                 .name = "max_usage_in_bytes",
3787                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3788                 .trigger = mem_cgroup_reset,
3789                 .read_u64 = mem_cgroup_read,
3790         },
3791         {
3792                 .name = "limit_in_bytes",
3793                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3794                 .write_string = mem_cgroup_write,
3795                 .read_u64 = mem_cgroup_read,
3796         },
3797         {
3798                 .name = "soft_limit_in_bytes",
3799                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3800                 .write_string = mem_cgroup_write,
3801                 .read_u64 = mem_cgroup_read,
3802         },
3803         {
3804                 .name = "failcnt",
3805                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3806                 .trigger = mem_cgroup_reset,
3807                 .read_u64 = mem_cgroup_read,
3808         },
3809         {
3810                 .name = "stat",
3811                 .read_map = mem_control_stat_show,
3812         },
3813         {
3814                 .name = "force_empty",
3815                 .trigger = mem_cgroup_force_empty_write,
3816         },
3817         {
3818                 .name = "use_hierarchy",
3819                 .write_u64 = mem_cgroup_hierarchy_write,
3820                 .read_u64 = mem_cgroup_hierarchy_read,
3821         },
3822         {
3823                 .name = "swappiness",
3824                 .read_u64 = mem_cgroup_swappiness_read,
3825                 .write_u64 = mem_cgroup_swappiness_write,
3826         },
3827         {
3828                 .name = "move_charge_at_immigrate",
3829                 .read_u64 = mem_cgroup_move_charge_read,
3830                 .write_u64 = mem_cgroup_move_charge_write,
3831         },
3832         {
3833                 .name = "oom_control",
3834                 .read_map = mem_cgroup_oom_control_read,
3835                 .write_u64 = mem_cgroup_oom_control_write,
3836                 .register_event = mem_cgroup_oom_register_event,
3837                 .unregister_event = mem_cgroup_oom_unregister_event,
3838                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3839         },
3840 };
3841
3842 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3843 static struct cftype memsw_cgroup_files[] = {
3844         {
3845                 .name = "memsw.usage_in_bytes",
3846                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3847                 .read_u64 = mem_cgroup_read,
3848                 .register_event = mem_cgroup_usage_register_event,
3849                 .unregister_event = mem_cgroup_usage_unregister_event,
3850         },
3851         {
3852                 .name = "memsw.max_usage_in_bytes",
3853                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3854                 .trigger = mem_cgroup_reset,
3855                 .read_u64 = mem_cgroup_read,
3856         },
3857         {
3858                 .name = "memsw.limit_in_bytes",
3859                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3860                 .write_string = mem_cgroup_write,
3861                 .read_u64 = mem_cgroup_read,
3862         },
3863         {
3864                 .name = "memsw.failcnt",
3865                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3866                 .trigger = mem_cgroup_reset,
3867                 .read_u64 = mem_cgroup_read,
3868         },
3869 };
3870
3871 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3872 {
3873         if (!do_swap_account)
3874                 return 0;
3875         return cgroup_add_files(cont, ss, memsw_cgroup_files,
3876                                 ARRAY_SIZE(memsw_cgroup_files));
3877 };
3878 #else
3879 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3880 {
3881         return 0;
3882 }
3883 #endif
3884
3885 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3886 {
3887         struct mem_cgroup_per_node *pn;
3888         struct mem_cgroup_per_zone *mz;
3889         enum lru_list l;
3890         int zone, tmp = node;
3891         /*
3892          * This routine is called against possible nodes.
3893          * But it's BUG to call kmalloc() against offline node.
3894          *
3895          * TODO: this routine can waste much memory for nodes which will
3896          *       never be onlined. It's better to use memory hotplug callback
3897          *       function.
3898          */
3899         if (!node_state(node, N_NORMAL_MEMORY))
3900                 tmp = -1;
3901         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3902         if (!pn)
3903                 return 1;
3904
3905         mem->info.nodeinfo[node] = pn;
3906         memset(pn, 0, sizeof(*pn));
3907
3908         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3909                 mz = &pn->zoneinfo[zone];
3910                 for_each_lru(l)
3911                         INIT_LIST_HEAD(&mz->lists[l]);
3912                 mz->usage_in_excess = 0;
3913                 mz->on_tree = false;
3914                 mz->mem = mem;
3915         }
3916         return 0;
3917 }
3918
3919 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3920 {
3921         kfree(mem->info.nodeinfo[node]);
3922 }
3923
3924 static struct mem_cgroup *mem_cgroup_alloc(void)
3925 {
3926         struct mem_cgroup *mem;
3927         int size = sizeof(struct mem_cgroup);
3928
3929         /* Can be very big if MAX_NUMNODES is very big */
3930         if (size < PAGE_SIZE)
3931                 mem = kmalloc(size, GFP_KERNEL);
3932         else
3933                 mem = vmalloc(size);
3934
3935         if (!mem)
3936                 return NULL;
3937
3938         memset(mem, 0, size);
3939         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3940         if (!mem->stat) {
3941                 if (size < PAGE_SIZE)
3942                         kfree(mem);
3943                 else
3944                         vfree(mem);
3945                 mem = NULL;
3946         }
3947         return mem;
3948 }
3949
3950 /*
3951  * At destroying mem_cgroup, references from swap_cgroup can remain.
3952  * (scanning all at force_empty is too costly...)
3953  *
3954  * Instead of clearing all references at force_empty, we remember
3955  * the number of reference from swap_cgroup and free mem_cgroup when
3956  * it goes down to 0.
3957  *
3958  * Removal of cgroup itself succeeds regardless of refs from swap.
3959  */
3960
3961 static void __mem_cgroup_free(struct mem_cgroup *mem)
3962 {
3963         int node;
3964
3965         mem_cgroup_remove_from_trees(mem);
3966         free_css_id(&mem_cgroup_subsys, &mem->css);
3967
3968         for_each_node_state(node, N_POSSIBLE)
3969                 free_mem_cgroup_per_zone_info(mem, node);
3970
3971         free_percpu(mem->stat);
3972         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
3973                 kfree(mem);
3974         else
3975                 vfree(mem);
3976 }
3977
3978 static void mem_cgroup_get(struct mem_cgroup *mem)
3979 {
3980         atomic_inc(&mem->refcnt);
3981 }
3982
3983 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
3984 {
3985         if (atomic_sub_and_test(count, &mem->refcnt)) {
3986                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3987                 __mem_cgroup_free(mem);
3988                 if (parent)
3989                         mem_cgroup_put(parent);
3990         }
3991 }
3992
3993 static void mem_cgroup_put(struct mem_cgroup *mem)
3994 {
3995         __mem_cgroup_put(mem, 1);
3996 }
3997
3998 /*
3999  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4000  */
4001 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4002 {
4003         if (!mem->res.parent)
4004                 return NULL;
4005         return mem_cgroup_from_res_counter(mem->res.parent, res);
4006 }
4007
4008 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4009 static void __init enable_swap_cgroup(void)
4010 {
4011         if (!mem_cgroup_disabled() && really_do_swap_account)
4012                 do_swap_account = 1;
4013 }
4014 #else
4015 static void __init enable_swap_cgroup(void)
4016 {
4017 }
4018 #endif
4019
4020 static int mem_cgroup_soft_limit_tree_init(void)
4021 {
4022         struct mem_cgroup_tree_per_node *rtpn;
4023         struct mem_cgroup_tree_per_zone *rtpz;
4024         int tmp, node, zone;
4025
4026         for_each_node_state(node, N_POSSIBLE) {
4027                 tmp = node;
4028                 if (!node_state(node, N_NORMAL_MEMORY))
4029                         tmp = -1;
4030                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4031                 if (!rtpn)
4032                         return 1;
4033
4034                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4035
4036                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4037                         rtpz = &rtpn->rb_tree_per_zone[zone];
4038                         rtpz->rb_root = RB_ROOT;
4039                         spin_lock_init(&rtpz->lock);
4040                 }
4041         }
4042         return 0;
4043 }
4044
4045 static struct cgroup_subsys_state * __ref
4046 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4047 {
4048         struct mem_cgroup *mem, *parent;
4049         long error = -ENOMEM;
4050         int node;
4051
4052         mem = mem_cgroup_alloc();
4053         if (!mem)
4054                 return ERR_PTR(error);
4055
4056         for_each_node_state(node, N_POSSIBLE)
4057                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4058                         goto free_out;
4059
4060         /* root ? */
4061         if (cont->parent == NULL) {
4062                 int cpu;
4063                 enable_swap_cgroup();
4064                 parent = NULL;
4065                 root_mem_cgroup = mem;
4066                 if (mem_cgroup_soft_limit_tree_init())
4067                         goto free_out;
4068                 for_each_possible_cpu(cpu) {
4069                         struct memcg_stock_pcp *stock =
4070                                                 &per_cpu(memcg_stock, cpu);
4071                         INIT_WORK(&stock->work, drain_local_stock);
4072                 }
4073                 hotcpu_notifier(memcg_stock_cpu_callback, 0);
4074         } else {
4075                 parent = mem_cgroup_from_cont(cont->parent);
4076                 mem->use_hierarchy = parent->use_hierarchy;
4077                 mem->oom_kill_disable = parent->oom_kill_disable;
4078         }
4079
4080         if (parent && parent->use_hierarchy) {
4081                 res_counter_init(&mem->res, &parent->res);
4082                 res_counter_init(&mem->memsw, &parent->memsw);
4083                 /*
4084                  * We increment refcnt of the parent to ensure that we can
4085                  * safely access it on res_counter_charge/uncharge.
4086                  * This refcnt will be decremented when freeing this
4087                  * mem_cgroup(see mem_cgroup_put).
4088                  */
4089                 mem_cgroup_get(parent);
4090         } else {
4091                 res_counter_init(&mem->res, NULL);
4092                 res_counter_init(&mem->memsw, NULL);
4093         }
4094         mem->last_scanned_child = 0;
4095         spin_lock_init(&mem->reclaim_param_lock);
4096         INIT_LIST_HEAD(&mem->oom_notify);
4097
4098         if (parent)
4099                 mem->swappiness = get_swappiness(parent);
4100         atomic_set(&mem->refcnt, 1);
4101         mem->move_charge_at_immigrate = 0;
4102         mutex_init(&mem->thresholds_lock);
4103         return &mem->css;
4104 free_out:
4105         __mem_cgroup_free(mem);
4106         root_mem_cgroup = NULL;
4107         return ERR_PTR(error);
4108 }
4109
4110 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4111                                         struct cgroup *cont)
4112 {
4113         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4114
4115         return mem_cgroup_force_empty(mem, false);
4116 }
4117
4118 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4119                                 struct cgroup *cont)
4120 {
4121         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4122
4123         mem_cgroup_put(mem);
4124 }
4125
4126 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4127                                 struct cgroup *cont)
4128 {
4129         int ret;
4130
4131         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4132                                 ARRAY_SIZE(mem_cgroup_files));
4133
4134         if (!ret)
4135                 ret = register_memsw_files(cont, ss);
4136         return ret;
4137 }
4138
4139 #ifdef CONFIG_MMU
4140 /* Handlers for move charge at task migration. */
4141 #define PRECHARGE_COUNT_AT_ONCE 256
4142 static int mem_cgroup_do_precharge(unsigned long count)
4143 {
4144         int ret = 0;
4145         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4146         struct mem_cgroup *mem = mc.to;
4147
4148         if (mem_cgroup_is_root(mem)) {
4149                 mc.precharge += count;
4150                 /* we don't need css_get for root */
4151                 return ret;
4152         }
4153         /* try to charge at once */
4154         if (count > 1) {
4155                 struct res_counter *dummy;
4156                 /*
4157                  * "mem" cannot be under rmdir() because we've already checked
4158                  * by cgroup_lock_live_cgroup() that it is not removed and we
4159                  * are still under the same cgroup_mutex. So we can postpone
4160                  * css_get().
4161                  */
4162                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4163                         goto one_by_one;
4164                 if (do_swap_account && res_counter_charge(&mem->memsw,
4165                                                 PAGE_SIZE * count, &dummy)) {
4166                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4167                         goto one_by_one;
4168                 }
4169                 mc.precharge += count;
4170                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
4171                 WARN_ON_ONCE(count > INT_MAX);
4172                 __css_get(&mem->css, (int)count);
4173                 return ret;
4174         }
4175 one_by_one:
4176         /* fall back to one by one charge */
4177         while (count--) {
4178                 if (signal_pending(current)) {
4179                         ret = -EINTR;
4180                         break;
4181                 }
4182                 if (!batch_count--) {
4183                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4184                         cond_resched();
4185                 }
4186                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4187                 if (ret || !mem)
4188                         /* mem_cgroup_clear_mc() will do uncharge later */
4189                         return -ENOMEM;
4190                 mc.precharge++;
4191         }
4192         return ret;
4193 }
4194
4195 /**
4196  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4197  * @vma: the vma the pte to be checked belongs
4198  * @addr: the address corresponding to the pte to be checked
4199  * @ptent: the pte to be checked
4200  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4201  *
4202  * Returns
4203  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4204  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4205  *     move charge. if @target is not NULL, the page is stored in target->page
4206  *     with extra refcnt got(Callers should handle it).
4207  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4208  *     target for charge migration. if @target is not NULL, the entry is stored
4209  *     in target->ent.
4210  *
4211  * Called with pte lock held.
4212  */
4213 union mc_target {
4214         struct page     *page;
4215         swp_entry_t     ent;
4216 };
4217
4218 enum mc_target_type {
4219         MC_TARGET_NONE, /* not used */
4220         MC_TARGET_PAGE,
4221         MC_TARGET_SWAP,
4222 };
4223
4224 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4225                                                 unsigned long addr, pte_t ptent)
4226 {
4227         struct page *page = vm_normal_page(vma, addr, ptent);
4228
4229         if (!page || !page_mapped(page))
4230                 return NULL;
4231         if (PageAnon(page)) {
4232                 /* we don't move shared anon */
4233                 if (!move_anon() || page_mapcount(page) > 2)
4234                         return NULL;
4235         } else if (!move_file())
4236                 /* we ignore mapcount for file pages */
4237                 return NULL;
4238         if (!get_page_unless_zero(page))
4239                 return NULL;
4240
4241         return page;
4242 }
4243
4244 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4245                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4246 {
4247         int usage_count;
4248         struct page *page = NULL;
4249         swp_entry_t ent = pte_to_swp_entry(ptent);
4250
4251         if (!move_anon() || non_swap_entry(ent))
4252                 return NULL;
4253         usage_count = mem_cgroup_count_swap_user(ent, &page);
4254         if (usage_count > 1) { /* we don't move shared anon */
4255                 if (page)
4256                         put_page(page);
4257                 return NULL;
4258         }
4259         if (do_swap_account)
4260                 entry->val = ent.val;
4261
4262         return page;
4263 }
4264
4265 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4266                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4267 {
4268         struct page *page = NULL;
4269         struct inode *inode;
4270         struct address_space *mapping;
4271         pgoff_t pgoff;
4272
4273         if (!vma->vm_file) /* anonymous vma */
4274                 return NULL;
4275         if (!move_file())
4276                 return NULL;
4277
4278         inode = vma->vm_file->f_path.dentry->d_inode;
4279         mapping = vma->vm_file->f_mapping;
4280         if (pte_none(ptent))
4281                 pgoff = linear_page_index(vma, addr);
4282         else /* pte_file(ptent) is true */
4283                 pgoff = pte_to_pgoff(ptent);
4284
4285         /* page is moved even if it's not RSS of this task(page-faulted). */
4286         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4287                 page = find_get_page(mapping, pgoff);
4288         } else { /* shmem/tmpfs file. we should take account of swap too. */
4289                 swp_entry_t ent;
4290                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4291                 if (do_swap_account)
4292                         entry->val = ent.val;
4293         }
4294
4295         return page;
4296 }
4297
4298 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4299                 unsigned long addr, pte_t ptent, union mc_target *target)
4300 {
4301         struct page *page = NULL;
4302         struct page_cgroup *pc;
4303         int ret = 0;
4304         swp_entry_t ent = { .val = 0 };
4305
4306         if (pte_present(ptent))
4307                 page = mc_handle_present_pte(vma, addr, ptent);
4308         else if (is_swap_pte(ptent))
4309                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4310         else if (pte_none(ptent) || pte_file(ptent))
4311                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4312
4313         if (!page && !ent.val)
4314                 return 0;
4315         if (page) {
4316                 pc = lookup_page_cgroup(page);
4317                 /*
4318                  * Do only loose check w/o page_cgroup lock.
4319                  * mem_cgroup_move_account() checks the pc is valid or not under
4320                  * the lock.
4321                  */
4322                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4323                         ret = MC_TARGET_PAGE;
4324                         if (target)
4325                                 target->page = page;
4326                 }
4327                 if (!ret || !target)
4328                         put_page(page);
4329         }
4330         /* There is a swap entry and a page doesn't exist or isn't charged */
4331         if (ent.val && !ret &&
4332                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4333                 ret = MC_TARGET_SWAP;
4334                 if (target)
4335                         target->ent = ent;
4336         }
4337         return ret;
4338 }
4339
4340 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4341                                         unsigned long addr, unsigned long end,
4342                                         struct mm_walk *walk)
4343 {
4344         struct vm_area_struct *vma = walk->private;
4345         pte_t *pte;
4346         spinlock_t *ptl;
4347
4348         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4349         for (; addr != end; pte++, addr += PAGE_SIZE)
4350                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4351                         mc.precharge++; /* increment precharge temporarily */
4352         pte_unmap_unlock(pte - 1, ptl);
4353         cond_resched();
4354
4355         return 0;
4356 }
4357
4358 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4359 {
4360         unsigned long precharge;
4361         struct vm_area_struct *vma;
4362
4363         down_read(&mm->mmap_sem);
4364         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4365                 struct mm_walk mem_cgroup_count_precharge_walk = {
4366                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4367                         .mm = mm,
4368                         .private = vma,
4369                 };
4370                 if (is_vm_hugetlb_page(vma))
4371                         continue;
4372                 walk_page_range(vma->vm_start, vma->vm_end,
4373                                         &mem_cgroup_count_precharge_walk);
4374         }
4375         up_read(&mm->mmap_sem);
4376
4377         precharge = mc.precharge;
4378         mc.precharge = 0;
4379
4380         return precharge;
4381 }
4382
4383 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4384 {
4385         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4386 }
4387
4388 static void mem_cgroup_clear_mc(void)
4389 {
4390         /* we must uncharge all the leftover precharges from mc.to */
4391         if (mc.precharge) {
4392                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4393                 mc.precharge = 0;
4394                 memcg_oom_recover(mc.to);
4395         }
4396         /*
4397          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4398          * we must uncharge here.
4399          */
4400         if (mc.moved_charge) {
4401                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4402                 mc.moved_charge = 0;
4403                 memcg_oom_recover(mc.from);
4404         }
4405         /* we must fixup refcnts and charges */
4406         if (mc.moved_swap) {
4407                 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4408                 /* uncharge swap account from the old cgroup */
4409                 if (!mem_cgroup_is_root(mc.from))
4410                         res_counter_uncharge(&mc.from->memsw,
4411                                                 PAGE_SIZE * mc.moved_swap);
4412                 __mem_cgroup_put(mc.from, mc.moved_swap);
4413
4414                 if (!mem_cgroup_is_root(mc.to)) {
4415                         /*
4416                          * we charged both to->res and to->memsw, so we should
4417                          * uncharge to->res.
4418                          */
4419                         res_counter_uncharge(&mc.to->res,
4420                                                 PAGE_SIZE * mc.moved_swap);
4421                         VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4422                         __css_put(&mc.to->css, mc.moved_swap);
4423                 }
4424                 /* we've already done mem_cgroup_get(mc.to) */
4425
4426                 mc.moved_swap = 0;
4427         }
4428         mc.from = NULL;
4429         mc.to = NULL;
4430         mc.moving_task = NULL;
4431         wake_up_all(&mc.waitq);
4432 }
4433
4434 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4435                                 struct cgroup *cgroup,
4436                                 struct task_struct *p,
4437                                 bool threadgroup)
4438 {
4439         int ret = 0;
4440         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4441
4442         if (mem->move_charge_at_immigrate) {
4443                 struct mm_struct *mm;
4444                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4445
4446                 VM_BUG_ON(from == mem);
4447
4448                 mm = get_task_mm(p);
4449                 if (!mm)
4450                         return 0;
4451                 /* We move charges only when we move a owner of the mm */
4452                 if (mm->owner == p) {
4453                         VM_BUG_ON(mc.from);
4454                         VM_BUG_ON(mc.to);
4455                         VM_BUG_ON(mc.precharge);
4456                         VM_BUG_ON(mc.moved_charge);
4457                         VM_BUG_ON(mc.moved_swap);
4458                         VM_BUG_ON(mc.moving_task);
4459                         mc.from = from;
4460                         mc.to = mem;
4461                         mc.precharge = 0;
4462                         mc.moved_charge = 0;
4463                         mc.moved_swap = 0;
4464                         mc.moving_task = current;
4465
4466                         ret = mem_cgroup_precharge_mc(mm);
4467                         if (ret)
4468                                 mem_cgroup_clear_mc();
4469                 }
4470                 mmput(mm);
4471         }
4472         return ret;
4473 }
4474
4475 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4476                                 struct cgroup *cgroup,
4477                                 struct task_struct *p,
4478                                 bool threadgroup)
4479 {
4480         mem_cgroup_clear_mc();
4481 }
4482
4483 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4484                                 unsigned long addr, unsigned long end,
4485                                 struct mm_walk *walk)
4486 {
4487         int ret = 0;
4488         struct vm_area_struct *vma = walk->private;
4489         pte_t *pte;
4490         spinlock_t *ptl;
4491
4492 retry:
4493         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4494         for (; addr != end; addr += PAGE_SIZE) {
4495                 pte_t ptent = *(pte++);
4496                 union mc_target target;
4497                 int type;
4498                 struct page *page;
4499                 struct page_cgroup *pc;
4500                 swp_entry_t ent;
4501
4502                 if (!mc.precharge)
4503                         break;
4504
4505                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4506                 switch (type) {
4507                 case MC_TARGET_PAGE:
4508                         page = target.page;
4509                         if (isolate_lru_page(page))
4510                                 goto put;
4511                         pc = lookup_page_cgroup(page);
4512                         if (!mem_cgroup_move_account(pc,
4513                                                 mc.from, mc.to, false)) {
4514                                 mc.precharge--;
4515                                 /* we uncharge from mc.from later. */
4516                                 mc.moved_charge++;
4517                         }
4518                         putback_lru_page(page);
4519 put:                    /* is_target_pte_for_mc() gets the page */
4520                         put_page(page);
4521                         break;
4522                 case MC_TARGET_SWAP:
4523                         ent = target.ent;
4524                         if (!mem_cgroup_move_swap_account(ent,
4525                                                 mc.from, mc.to, false)) {
4526                                 mc.precharge--;
4527                                 /* we fixup refcnts and charges later. */
4528                                 mc.moved_swap++;
4529                         }
4530                         break;
4531                 default:
4532                         break;
4533                 }
4534         }
4535         pte_unmap_unlock(pte - 1, ptl);
4536         cond_resched();
4537
4538         if (addr != end) {
4539                 /*
4540                  * We have consumed all precharges we got in can_attach().
4541                  * We try charge one by one, but don't do any additional
4542                  * charges to mc.to if we have failed in charge once in attach()
4543                  * phase.
4544                  */
4545                 ret = mem_cgroup_do_precharge(1);
4546                 if (!ret)
4547                         goto retry;
4548         }
4549
4550         return ret;
4551 }
4552
4553 static void mem_cgroup_move_charge(struct mm_struct *mm)
4554 {
4555         struct vm_area_struct *vma;
4556
4557         lru_add_drain_all();
4558         down_read(&mm->mmap_sem);
4559         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4560                 int ret;
4561                 struct mm_walk mem_cgroup_move_charge_walk = {
4562                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4563                         .mm = mm,
4564                         .private = vma,
4565                 };
4566                 if (is_vm_hugetlb_page(vma))
4567                         continue;
4568                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4569                                                 &mem_cgroup_move_charge_walk);
4570                 if (ret)
4571                         /*
4572                          * means we have consumed all precharges and failed in
4573                          * doing additional charge. Just abandon here.
4574                          */
4575                         break;
4576         }
4577         up_read(&mm->mmap_sem);
4578 }
4579
4580 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4581                                 struct cgroup *cont,
4582                                 struct cgroup *old_cont,
4583                                 struct task_struct *p,
4584                                 bool threadgroup)
4585 {
4586         struct mm_struct *mm;
4587
4588         if (!mc.to)
4589                 /* no need to move charge */
4590                 return;
4591
4592         mm = get_task_mm(p);
4593         if (mm) {
4594                 mem_cgroup_move_charge(mm);
4595                 mmput(mm);
4596         }
4597         mem_cgroup_clear_mc();
4598 }
4599 #else   /* !CONFIG_MMU */
4600 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4601                                 struct cgroup *cgroup,
4602                                 struct task_struct *p,
4603                                 bool threadgroup)
4604 {
4605         return 0;
4606 }
4607 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4608                                 struct cgroup *cgroup,
4609                                 struct task_struct *p,
4610                                 bool threadgroup)
4611 {
4612 }
4613 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4614                                 struct cgroup *cont,
4615                                 struct cgroup *old_cont,
4616                                 struct task_struct *p,
4617                                 bool threadgroup)
4618 {
4619 }
4620 #endif
4621
4622 struct cgroup_subsys mem_cgroup_subsys = {
4623         .name = "memory",
4624         .subsys_id = mem_cgroup_subsys_id,
4625         .create = mem_cgroup_create,
4626         .pre_destroy = mem_cgroup_pre_destroy,
4627         .destroy = mem_cgroup_destroy,
4628         .populate = mem_cgroup_populate,
4629         .can_attach = mem_cgroup_can_attach,
4630         .cancel_attach = mem_cgroup_cancel_attach,
4631         .attach = mem_cgroup_move_task,
4632         .early_init = 0,
4633         .use_id = 1,
4634 };
4635
4636 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4637
4638 static int __init disable_swap_account(char *s)
4639 {
4640         really_do_swap_account = 0;
4641         return 1;
4642 }
4643 __setup("noswapaccount", disable_swap_account);
4644 #endif