Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[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 struct mem_cgroup_threshold_ary {
153         /* An array index points to threshold just below usage. */
154         atomic_t current_threshold;
155         /* Size of entries[] */
156         unsigned int size;
157         /* Array of thresholds */
158         struct mem_cgroup_threshold entries[0];
159 };
160
161 static void mem_cgroup_threshold(struct mem_cgroup *mem);
162
163 /*
164  * The memory controller data structure. The memory controller controls both
165  * page cache and RSS per cgroup. We would eventually like to provide
166  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
167  * to help the administrator determine what knobs to tune.
168  *
169  * TODO: Add a water mark for the memory controller. Reclaim will begin when
170  * we hit the water mark. May be even add a low water mark, such that
171  * no reclaim occurs from a cgroup at it's low water mark, this is
172  * a feature that will be implemented much later in the future.
173  */
174 struct mem_cgroup {
175         struct cgroup_subsys_state css;
176         /*
177          * the counter to account for memory usage
178          */
179         struct res_counter res;
180         /*
181          * the counter to account for mem+swap usage.
182          */
183         struct res_counter memsw;
184         /*
185          * Per cgroup active and inactive list, similar to the
186          * per zone LRU lists.
187          */
188         struct mem_cgroup_lru_info info;
189
190         /*
191           protect against reclaim related member.
192         */
193         spinlock_t reclaim_param_lock;
194
195         int     prev_priority;  /* for recording reclaim priority */
196
197         /*
198          * While reclaiming in a hierarchy, we cache the last child we
199          * reclaimed from.
200          */
201         int last_scanned_child;
202         /*
203          * Should the accounting and control be hierarchical, per subtree?
204          */
205         bool use_hierarchy;
206         atomic_t        oom_lock;
207         atomic_t        refcnt;
208
209         unsigned int    swappiness;
210
211         /* set when res.limit == memsw.limit */
212         bool            memsw_is_minimum;
213
214         /* protect arrays of thresholds */
215         struct mutex thresholds_lock;
216
217         /* thresholds for memory usage. RCU-protected */
218         struct mem_cgroup_threshold_ary *thresholds;
219
220         /* thresholds for mem+swap usage. RCU-protected */
221         struct mem_cgroup_threshold_ary *memsw_thresholds;
222
223         /*
224          * Should we move charges of a task when a task is moved into this
225          * mem_cgroup ? And what type of charges should we move ?
226          */
227         unsigned long   move_charge_at_immigrate;
228
229         /*
230          * percpu counter.
231          */
232         struct mem_cgroup_stat_cpu *stat;
233 };
234
235 /* Stuffs for move charges at task migration. */
236 /*
237  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
238  * left-shifted bitmap of these types.
239  */
240 enum move_type {
241         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
242         NR_MOVE_TYPE,
243 };
244
245 /* "mc" and its members are protected by cgroup_mutex */
246 static struct move_charge_struct {
247         struct mem_cgroup *from;
248         struct mem_cgroup *to;
249         unsigned long precharge;
250         unsigned long moved_charge;
251         unsigned long moved_swap;
252         struct task_struct *moving_task;        /* a task moving charges */
253         wait_queue_head_t waitq;                /* a waitq for other context */
254 } mc = {
255         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
256 };
257
258 /*
259  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
260  * limit reclaim to prevent infinite loops, if they ever occur.
261  */
262 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
263 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
264
265 enum charge_type {
266         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
267         MEM_CGROUP_CHARGE_TYPE_MAPPED,
268         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
269         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
270         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
271         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
272         NR_CHARGE_TYPE,
273 };
274
275 /* only for here (for easy reading.) */
276 #define PCGF_CACHE      (1UL << PCG_CACHE)
277 #define PCGF_USED       (1UL << PCG_USED)
278 #define PCGF_LOCK       (1UL << PCG_LOCK)
279 /* Not used, but added here for completeness */
280 #define PCGF_ACCT       (1UL << PCG_ACCT)
281
282 /* for encoding cft->private value on file */
283 #define _MEM                    (0)
284 #define _MEMSWAP                (1)
285 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
286 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
287 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
288
289 /*
290  * Reclaim flags for mem_cgroup_hierarchical_reclaim
291  */
292 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
293 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
294 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
295 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
296 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
297 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
298
299 static void mem_cgroup_get(struct mem_cgroup *mem);
300 static void mem_cgroup_put(struct mem_cgroup *mem);
301 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
302 static void drain_all_stock_async(void);
303
304 static struct mem_cgroup_per_zone *
305 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
306 {
307         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
308 }
309
310 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
311 {
312         return &mem->css;
313 }
314
315 static struct mem_cgroup_per_zone *
316 page_cgroup_zoneinfo(struct page_cgroup *pc)
317 {
318         struct mem_cgroup *mem = pc->mem_cgroup;
319         int nid = page_cgroup_nid(pc);
320         int zid = page_cgroup_zid(pc);
321
322         if (!mem)
323                 return NULL;
324
325         return mem_cgroup_zoneinfo(mem, nid, zid);
326 }
327
328 static struct mem_cgroup_tree_per_zone *
329 soft_limit_tree_node_zone(int nid, int zid)
330 {
331         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
332 }
333
334 static struct mem_cgroup_tree_per_zone *
335 soft_limit_tree_from_page(struct page *page)
336 {
337         int nid = page_to_nid(page);
338         int zid = page_zonenum(page);
339
340         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
341 }
342
343 static void
344 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
345                                 struct mem_cgroup_per_zone *mz,
346                                 struct mem_cgroup_tree_per_zone *mctz,
347                                 unsigned long long new_usage_in_excess)
348 {
349         struct rb_node **p = &mctz->rb_root.rb_node;
350         struct rb_node *parent = NULL;
351         struct mem_cgroup_per_zone *mz_node;
352
353         if (mz->on_tree)
354                 return;
355
356         mz->usage_in_excess = new_usage_in_excess;
357         if (!mz->usage_in_excess)
358                 return;
359         while (*p) {
360                 parent = *p;
361                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
362                                         tree_node);
363                 if (mz->usage_in_excess < mz_node->usage_in_excess)
364                         p = &(*p)->rb_left;
365                 /*
366                  * We can't avoid mem cgroups that are over their soft
367                  * limit by the same amount
368                  */
369                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
370                         p = &(*p)->rb_right;
371         }
372         rb_link_node(&mz->tree_node, parent, p);
373         rb_insert_color(&mz->tree_node, &mctz->rb_root);
374         mz->on_tree = true;
375 }
376
377 static void
378 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
379                                 struct mem_cgroup_per_zone *mz,
380                                 struct mem_cgroup_tree_per_zone *mctz)
381 {
382         if (!mz->on_tree)
383                 return;
384         rb_erase(&mz->tree_node, &mctz->rb_root);
385         mz->on_tree = false;
386 }
387
388 static void
389 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
390                                 struct mem_cgroup_per_zone *mz,
391                                 struct mem_cgroup_tree_per_zone *mctz)
392 {
393         spin_lock(&mctz->lock);
394         __mem_cgroup_remove_exceeded(mem, mz, mctz);
395         spin_unlock(&mctz->lock);
396 }
397
398
399 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
400 {
401         unsigned long long excess;
402         struct mem_cgroup_per_zone *mz;
403         struct mem_cgroup_tree_per_zone *mctz;
404         int nid = page_to_nid(page);
405         int zid = page_zonenum(page);
406         mctz = soft_limit_tree_from_page(page);
407
408         /*
409          * Necessary to update all ancestors when hierarchy is used.
410          * because their event counter is not touched.
411          */
412         for (; mem; mem = parent_mem_cgroup(mem)) {
413                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
414                 excess = res_counter_soft_limit_excess(&mem->res);
415                 /*
416                  * We have to update the tree if mz is on RB-tree or
417                  * mem is over its softlimit.
418                  */
419                 if (excess || mz->on_tree) {
420                         spin_lock(&mctz->lock);
421                         /* if on-tree, remove it */
422                         if (mz->on_tree)
423                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
424                         /*
425                          * Insert again. mz->usage_in_excess will be updated.
426                          * If excess is 0, no tree ops.
427                          */
428                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
429                         spin_unlock(&mctz->lock);
430                 }
431         }
432 }
433
434 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
435 {
436         int node, zone;
437         struct mem_cgroup_per_zone *mz;
438         struct mem_cgroup_tree_per_zone *mctz;
439
440         for_each_node_state(node, N_POSSIBLE) {
441                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
442                         mz = mem_cgroup_zoneinfo(mem, node, zone);
443                         mctz = soft_limit_tree_node_zone(node, zone);
444                         mem_cgroup_remove_exceeded(mem, mz, mctz);
445                 }
446         }
447 }
448
449 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
450 {
451         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
452 }
453
454 static struct mem_cgroup_per_zone *
455 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
456 {
457         struct rb_node *rightmost = NULL;
458         struct mem_cgroup_per_zone *mz;
459
460 retry:
461         mz = NULL;
462         rightmost = rb_last(&mctz->rb_root);
463         if (!rightmost)
464                 goto done;              /* Nothing to reclaim from */
465
466         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
467         /*
468          * Remove the node now but someone else can add it back,
469          * we will to add it back at the end of reclaim to its correct
470          * position in the tree.
471          */
472         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
473         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
474                 !css_tryget(&mz->mem->css))
475                 goto retry;
476 done:
477         return mz;
478 }
479
480 static struct mem_cgroup_per_zone *
481 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
482 {
483         struct mem_cgroup_per_zone *mz;
484
485         spin_lock(&mctz->lock);
486         mz = __mem_cgroup_largest_soft_limit_node(mctz);
487         spin_unlock(&mctz->lock);
488         return mz;
489 }
490
491 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
492                 enum mem_cgroup_stat_index idx)
493 {
494         int cpu;
495         s64 val = 0;
496
497         for_each_possible_cpu(cpu)
498                 val += per_cpu(mem->stat->count[idx], cpu);
499         return val;
500 }
501
502 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
503 {
504         s64 ret;
505
506         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
507         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
508         return ret;
509 }
510
511 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
512                                          bool charge)
513 {
514         int val = (charge) ? 1 : -1;
515         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
516 }
517
518 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
519                                          struct page_cgroup *pc,
520                                          bool charge)
521 {
522         int val = (charge) ? 1 : -1;
523
524         preempt_disable();
525
526         if (PageCgroupCache(pc))
527                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
528         else
529                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
530
531         if (charge)
532                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
533         else
534                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
535         __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
536
537         preempt_enable();
538 }
539
540 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
541                                         enum lru_list idx)
542 {
543         int nid, zid;
544         struct mem_cgroup_per_zone *mz;
545         u64 total = 0;
546
547         for_each_online_node(nid)
548                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
549                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
550                         total += MEM_CGROUP_ZSTAT(mz, idx);
551                 }
552         return total;
553 }
554
555 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
556 {
557         s64 val;
558
559         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
560
561         return !(val & ((1 << event_mask_shift) - 1));
562 }
563
564 /*
565  * Check events in order.
566  *
567  */
568 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
569 {
570         /* threshold event is triggered in finer grain than soft limit */
571         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
572                 mem_cgroup_threshold(mem);
573                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
574                         mem_cgroup_update_tree(mem, page);
575         }
576 }
577
578 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
579 {
580         return container_of(cgroup_subsys_state(cont,
581                                 mem_cgroup_subsys_id), struct mem_cgroup,
582                                 css);
583 }
584
585 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
586 {
587         /*
588          * mm_update_next_owner() may clear mm->owner to NULL
589          * if it races with swapoff, page migration, etc.
590          * So this can be called with p == NULL.
591          */
592         if (unlikely(!p))
593                 return NULL;
594
595         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
596                                 struct mem_cgroup, css);
597 }
598
599 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
600 {
601         struct mem_cgroup *mem = NULL;
602
603         if (!mm)
604                 return NULL;
605         /*
606          * Because we have no locks, mm->owner's may be being moved to other
607          * cgroup. We use css_tryget() here even if this looks
608          * pessimistic (rather than adding locks here).
609          */
610         rcu_read_lock();
611         do {
612                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
613                 if (unlikely(!mem))
614                         break;
615         } while (!css_tryget(&mem->css));
616         rcu_read_unlock();
617         return mem;
618 }
619
620 /*
621  * Call callback function against all cgroup under hierarchy tree.
622  */
623 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
624                           int (*func)(struct mem_cgroup *, void *))
625 {
626         int found, ret, nextid;
627         struct cgroup_subsys_state *css;
628         struct mem_cgroup *mem;
629
630         if (!root->use_hierarchy)
631                 return (*func)(root, data);
632
633         nextid = 1;
634         do {
635                 ret = 0;
636                 mem = NULL;
637
638                 rcu_read_lock();
639                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
640                                    &found);
641                 if (css && css_tryget(css))
642                         mem = container_of(css, struct mem_cgroup, css);
643                 rcu_read_unlock();
644
645                 if (mem) {
646                         ret = (*func)(mem, data);
647                         css_put(&mem->css);
648                 }
649                 nextid = found + 1;
650         } while (!ret && css);
651
652         return ret;
653 }
654
655 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
656 {
657         return (mem == root_mem_cgroup);
658 }
659
660 /*
661  * Following LRU functions are allowed to be used without PCG_LOCK.
662  * Operations are called by routine of global LRU independently from memcg.
663  * What we have to take care of here is validness of pc->mem_cgroup.
664  *
665  * Changes to pc->mem_cgroup happens when
666  * 1. charge
667  * 2. moving account
668  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
669  * It is added to LRU before charge.
670  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
671  * When moving account, the page is not on LRU. It's isolated.
672  */
673
674 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
675 {
676         struct page_cgroup *pc;
677         struct mem_cgroup_per_zone *mz;
678
679         if (mem_cgroup_disabled())
680                 return;
681         pc = lookup_page_cgroup(page);
682         /* can happen while we handle swapcache. */
683         if (!TestClearPageCgroupAcctLRU(pc))
684                 return;
685         VM_BUG_ON(!pc->mem_cgroup);
686         /*
687          * We don't check PCG_USED bit. It's cleared when the "page" is finally
688          * removed from global LRU.
689          */
690         mz = page_cgroup_zoneinfo(pc);
691         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
692         if (mem_cgroup_is_root(pc->mem_cgroup))
693                 return;
694         VM_BUG_ON(list_empty(&pc->lru));
695         list_del_init(&pc->lru);
696         return;
697 }
698
699 void mem_cgroup_del_lru(struct page *page)
700 {
701         mem_cgroup_del_lru_list(page, page_lru(page));
702 }
703
704 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
705 {
706         struct mem_cgroup_per_zone *mz;
707         struct page_cgroup *pc;
708
709         if (mem_cgroup_disabled())
710                 return;
711
712         pc = lookup_page_cgroup(page);
713         /*
714          * Used bit is set without atomic ops but after smp_wmb().
715          * For making pc->mem_cgroup visible, insert smp_rmb() here.
716          */
717         smp_rmb();
718         /* unused or root page is not rotated. */
719         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
720                 return;
721         mz = page_cgroup_zoneinfo(pc);
722         list_move(&pc->lru, &mz->lists[lru]);
723 }
724
725 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
726 {
727         struct page_cgroup *pc;
728         struct mem_cgroup_per_zone *mz;
729
730         if (mem_cgroup_disabled())
731                 return;
732         pc = lookup_page_cgroup(page);
733         VM_BUG_ON(PageCgroupAcctLRU(pc));
734         /*
735          * Used bit is set without atomic ops but after smp_wmb().
736          * For making pc->mem_cgroup visible, insert smp_rmb() here.
737          */
738         smp_rmb();
739         if (!PageCgroupUsed(pc))
740                 return;
741
742         mz = page_cgroup_zoneinfo(pc);
743         MEM_CGROUP_ZSTAT(mz, lru) += 1;
744         SetPageCgroupAcctLRU(pc);
745         if (mem_cgroup_is_root(pc->mem_cgroup))
746                 return;
747         list_add(&pc->lru, &mz->lists[lru]);
748 }
749
750 /*
751  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
752  * lru because the page may.be reused after it's fully uncharged (because of
753  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
754  * it again. This function is only used to charge SwapCache. It's done under
755  * lock_page and expected that zone->lru_lock is never held.
756  */
757 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
758 {
759         unsigned long flags;
760         struct zone *zone = page_zone(page);
761         struct page_cgroup *pc = lookup_page_cgroup(page);
762
763         spin_lock_irqsave(&zone->lru_lock, flags);
764         /*
765          * Forget old LRU when this page_cgroup is *not* used. This Used bit
766          * is guarded by lock_page() because the page is SwapCache.
767          */
768         if (!PageCgroupUsed(pc))
769                 mem_cgroup_del_lru_list(page, page_lru(page));
770         spin_unlock_irqrestore(&zone->lru_lock, flags);
771 }
772
773 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
774 {
775         unsigned long flags;
776         struct zone *zone = page_zone(page);
777         struct page_cgroup *pc = lookup_page_cgroup(page);
778
779         spin_lock_irqsave(&zone->lru_lock, flags);
780         /* link when the page is linked to LRU but page_cgroup isn't */
781         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
782                 mem_cgroup_add_lru_list(page, page_lru(page));
783         spin_unlock_irqrestore(&zone->lru_lock, flags);
784 }
785
786
787 void mem_cgroup_move_lists(struct page *page,
788                            enum lru_list from, enum lru_list to)
789 {
790         if (mem_cgroup_disabled())
791                 return;
792         mem_cgroup_del_lru_list(page, from);
793         mem_cgroup_add_lru_list(page, to);
794 }
795
796 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
797 {
798         int ret;
799         struct mem_cgroup *curr = NULL;
800
801         task_lock(task);
802         rcu_read_lock();
803         curr = try_get_mem_cgroup_from_mm(task->mm);
804         rcu_read_unlock();
805         task_unlock(task);
806         if (!curr)
807                 return 0;
808         /*
809          * We should check use_hierarchy of "mem" not "curr". Because checking
810          * use_hierarchy of "curr" here make this function true if hierarchy is
811          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
812          * hierarchy(even if use_hierarchy is disabled in "mem").
813          */
814         if (mem->use_hierarchy)
815                 ret = css_is_ancestor(&curr->css, &mem->css);
816         else
817                 ret = (curr == mem);
818         css_put(&curr->css);
819         return ret;
820 }
821
822 /*
823  * prev_priority control...this will be used in memory reclaim path.
824  */
825 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
826 {
827         int prev_priority;
828
829         spin_lock(&mem->reclaim_param_lock);
830         prev_priority = mem->prev_priority;
831         spin_unlock(&mem->reclaim_param_lock);
832
833         return prev_priority;
834 }
835
836 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
837 {
838         spin_lock(&mem->reclaim_param_lock);
839         if (priority < mem->prev_priority)
840                 mem->prev_priority = priority;
841         spin_unlock(&mem->reclaim_param_lock);
842 }
843
844 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
845 {
846         spin_lock(&mem->reclaim_param_lock);
847         mem->prev_priority = priority;
848         spin_unlock(&mem->reclaim_param_lock);
849 }
850
851 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
852 {
853         unsigned long active;
854         unsigned long inactive;
855         unsigned long gb;
856         unsigned long inactive_ratio;
857
858         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
859         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
860
861         gb = (inactive + active) >> (30 - PAGE_SHIFT);
862         if (gb)
863                 inactive_ratio = int_sqrt(10 * gb);
864         else
865                 inactive_ratio = 1;
866
867         if (present_pages) {
868                 present_pages[0] = inactive;
869                 present_pages[1] = active;
870         }
871
872         return inactive_ratio;
873 }
874
875 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
876 {
877         unsigned long active;
878         unsigned long inactive;
879         unsigned long present_pages[2];
880         unsigned long inactive_ratio;
881
882         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
883
884         inactive = present_pages[0];
885         active = present_pages[1];
886
887         if (inactive * inactive_ratio < active)
888                 return 1;
889
890         return 0;
891 }
892
893 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
894 {
895         unsigned long active;
896         unsigned long inactive;
897
898         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
899         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
900
901         return (active > inactive);
902 }
903
904 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
905                                        struct zone *zone,
906                                        enum lru_list lru)
907 {
908         int nid = zone->zone_pgdat->node_id;
909         int zid = zone_idx(zone);
910         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
911
912         return MEM_CGROUP_ZSTAT(mz, lru);
913 }
914
915 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
916                                                       struct zone *zone)
917 {
918         int nid = zone->zone_pgdat->node_id;
919         int zid = zone_idx(zone);
920         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
921
922         return &mz->reclaim_stat;
923 }
924
925 struct zone_reclaim_stat *
926 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
927 {
928         struct page_cgroup *pc;
929         struct mem_cgroup_per_zone *mz;
930
931         if (mem_cgroup_disabled())
932                 return NULL;
933
934         pc = lookup_page_cgroup(page);
935         /*
936          * Used bit is set without atomic ops but after smp_wmb().
937          * For making pc->mem_cgroup visible, insert smp_rmb() here.
938          */
939         smp_rmb();
940         if (!PageCgroupUsed(pc))
941                 return NULL;
942
943         mz = page_cgroup_zoneinfo(pc);
944         if (!mz)
945                 return NULL;
946
947         return &mz->reclaim_stat;
948 }
949
950 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
951                                         struct list_head *dst,
952                                         unsigned long *scanned, int order,
953                                         int mode, struct zone *z,
954                                         struct mem_cgroup *mem_cont,
955                                         int active, int file)
956 {
957         unsigned long nr_taken = 0;
958         struct page *page;
959         unsigned long scan;
960         LIST_HEAD(pc_list);
961         struct list_head *src;
962         struct page_cgroup *pc, *tmp;
963         int nid = z->zone_pgdat->node_id;
964         int zid = zone_idx(z);
965         struct mem_cgroup_per_zone *mz;
966         int lru = LRU_FILE * file + active;
967         int ret;
968
969         BUG_ON(!mem_cont);
970         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
971         src = &mz->lists[lru];
972
973         scan = 0;
974         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
975                 if (scan >= nr_to_scan)
976                         break;
977
978                 page = pc->page;
979                 if (unlikely(!PageCgroupUsed(pc)))
980                         continue;
981                 if (unlikely(!PageLRU(page)))
982                         continue;
983
984                 scan++;
985                 ret = __isolate_lru_page(page, mode, file);
986                 switch (ret) {
987                 case 0:
988                         list_move(&page->lru, dst);
989                         mem_cgroup_del_lru(page);
990                         nr_taken++;
991                         break;
992                 case -EBUSY:
993                         /* we don't affect global LRU but rotate in our LRU */
994                         mem_cgroup_rotate_lru_list(page, page_lru(page));
995                         break;
996                 default:
997                         break;
998                 }
999         }
1000
1001         *scanned = scan;
1002         return nr_taken;
1003 }
1004
1005 #define mem_cgroup_from_res_counter(counter, member)    \
1006         container_of(counter, struct mem_cgroup, member)
1007
1008 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1009 {
1010         if (do_swap_account) {
1011                 if (res_counter_check_under_limit(&mem->res) &&
1012                         res_counter_check_under_limit(&mem->memsw))
1013                         return true;
1014         } else
1015                 if (res_counter_check_under_limit(&mem->res))
1016                         return true;
1017         return false;
1018 }
1019
1020 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1021 {
1022         struct cgroup *cgrp = memcg->css.cgroup;
1023         unsigned int swappiness;
1024
1025         /* root ? */
1026         if (cgrp->parent == NULL)
1027                 return vm_swappiness;
1028
1029         spin_lock(&memcg->reclaim_param_lock);
1030         swappiness = memcg->swappiness;
1031         spin_unlock(&memcg->reclaim_param_lock);
1032
1033         return swappiness;
1034 }
1035
1036 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1037 {
1038         int *val = data;
1039         (*val)++;
1040         return 0;
1041 }
1042
1043 /**
1044  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1045  * @memcg: The memory cgroup that went over limit
1046  * @p: Task that is going to be killed
1047  *
1048  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1049  * enabled
1050  */
1051 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1052 {
1053         struct cgroup *task_cgrp;
1054         struct cgroup *mem_cgrp;
1055         /*
1056          * Need a buffer in BSS, can't rely on allocations. The code relies
1057          * on the assumption that OOM is serialized for memory controller.
1058          * If this assumption is broken, revisit this code.
1059          */
1060         static char memcg_name[PATH_MAX];
1061         int ret;
1062
1063         if (!memcg || !p)
1064                 return;
1065
1066
1067         rcu_read_lock();
1068
1069         mem_cgrp = memcg->css.cgroup;
1070         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1071
1072         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1073         if (ret < 0) {
1074                 /*
1075                  * Unfortunately, we are unable to convert to a useful name
1076                  * But we'll still print out the usage information
1077                  */
1078                 rcu_read_unlock();
1079                 goto done;
1080         }
1081         rcu_read_unlock();
1082
1083         printk(KERN_INFO "Task in %s killed", memcg_name);
1084
1085         rcu_read_lock();
1086         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1087         if (ret < 0) {
1088                 rcu_read_unlock();
1089                 goto done;
1090         }
1091         rcu_read_unlock();
1092
1093         /*
1094          * Continues from above, so we don't need an KERN_ level
1095          */
1096         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1097 done:
1098
1099         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1100                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1101                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1102                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1103         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1104                 "failcnt %llu\n",
1105                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1106                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1107                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1108 }
1109
1110 /*
1111  * This function returns the number of memcg under hierarchy tree. Returns
1112  * 1(self count) if no children.
1113  */
1114 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1115 {
1116         int num = 0;
1117         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1118         return num;
1119 }
1120
1121 /*
1122  * Visit the first child (need not be the first child as per the ordering
1123  * of the cgroup list, since we track last_scanned_child) of @mem and use
1124  * that to reclaim free pages from.
1125  */
1126 static struct mem_cgroup *
1127 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1128 {
1129         struct mem_cgroup *ret = NULL;
1130         struct cgroup_subsys_state *css;
1131         int nextid, found;
1132
1133         if (!root_mem->use_hierarchy) {
1134                 css_get(&root_mem->css);
1135                 ret = root_mem;
1136         }
1137
1138         while (!ret) {
1139                 rcu_read_lock();
1140                 nextid = root_mem->last_scanned_child + 1;
1141                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1142                                    &found);
1143                 if (css && css_tryget(css))
1144                         ret = container_of(css, struct mem_cgroup, css);
1145
1146                 rcu_read_unlock();
1147                 /* Updates scanning parameter */
1148                 spin_lock(&root_mem->reclaim_param_lock);
1149                 if (!css) {
1150                         /* this means start scan from ID:1 */
1151                         root_mem->last_scanned_child = 0;
1152                 } else
1153                         root_mem->last_scanned_child = found;
1154                 spin_unlock(&root_mem->reclaim_param_lock);
1155         }
1156
1157         return ret;
1158 }
1159
1160 /*
1161  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1162  * we reclaimed from, so that we don't end up penalizing one child extensively
1163  * based on its position in the children list.
1164  *
1165  * root_mem is the original ancestor that we've been reclaim from.
1166  *
1167  * We give up and return to the caller when we visit root_mem twice.
1168  * (other groups can be removed while we're walking....)
1169  *
1170  * If shrink==true, for avoiding to free too much, this returns immedieately.
1171  */
1172 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1173                                                 struct zone *zone,
1174                                                 gfp_t gfp_mask,
1175                                                 unsigned long reclaim_options)
1176 {
1177         struct mem_cgroup *victim;
1178         int ret, total = 0;
1179         int loop = 0;
1180         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1181         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1182         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1183         unsigned long excess = mem_cgroup_get_excess(root_mem);
1184
1185         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1186         if (root_mem->memsw_is_minimum)
1187                 noswap = true;
1188
1189         while (1) {
1190                 victim = mem_cgroup_select_victim(root_mem);
1191                 if (victim == root_mem) {
1192                         loop++;
1193                         if (loop >= 1)
1194                                 drain_all_stock_async();
1195                         if (loop >= 2) {
1196                                 /*
1197                                  * If we have not been able to reclaim
1198                                  * anything, it might because there are
1199                                  * no reclaimable pages under this hierarchy
1200                                  */
1201                                 if (!check_soft || !total) {
1202                                         css_put(&victim->css);
1203                                         break;
1204                                 }
1205                                 /*
1206                                  * We want to do more targetted reclaim.
1207                                  * excess >> 2 is not to excessive so as to
1208                                  * reclaim too much, nor too less that we keep
1209                                  * coming back to reclaim from this cgroup
1210                                  */
1211                                 if (total >= (excess >> 2) ||
1212                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1213                                         css_put(&victim->css);
1214                                         break;
1215                                 }
1216                         }
1217                 }
1218                 if (!mem_cgroup_local_usage(victim)) {
1219                         /* this cgroup's local usage == 0 */
1220                         css_put(&victim->css);
1221                         continue;
1222                 }
1223                 /* we use swappiness of local cgroup */
1224                 if (check_soft)
1225                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1226                                 noswap, get_swappiness(victim), zone,
1227                                 zone->zone_pgdat->node_id);
1228                 else
1229                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1230                                                 noswap, get_swappiness(victim));
1231                 css_put(&victim->css);
1232                 /*
1233                  * At shrinking usage, we can't check we should stop here or
1234                  * reclaim more. It's depends on callers. last_scanned_child
1235                  * will work enough for keeping fairness under tree.
1236                  */
1237                 if (shrink)
1238                         return ret;
1239                 total += ret;
1240                 if (check_soft) {
1241                         if (res_counter_check_under_soft_limit(&root_mem->res))
1242                                 return total;
1243                 } else if (mem_cgroup_check_under_limit(root_mem))
1244                         return 1 + total;
1245         }
1246         return total;
1247 }
1248
1249 static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1250 {
1251         int *val = (int *)data;
1252         int x;
1253         /*
1254          * Logically, we can stop scanning immediately when we find
1255          * a memcg is already locked. But condidering unlock ops and
1256          * creation/removal of memcg, scan-all is simple operation.
1257          */
1258         x = atomic_inc_return(&mem->oom_lock);
1259         *val = max(x, *val);
1260         return 0;
1261 }
1262 /*
1263  * Check OOM-Killer is already running under our hierarchy.
1264  * If someone is running, return false.
1265  */
1266 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1267 {
1268         int lock_count = 0;
1269
1270         mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1271
1272         if (lock_count == 1)
1273                 return true;
1274         return false;
1275 }
1276
1277 static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1278 {
1279         /*
1280          * When a new child is created while the hierarchy is under oom,
1281          * mem_cgroup_oom_lock() may not be called. We have to use
1282          * atomic_add_unless() here.
1283          */
1284         atomic_add_unless(&mem->oom_lock, -1, 0);
1285         return 0;
1286 }
1287
1288 static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1289 {
1290         mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1291 }
1292
1293 static DEFINE_MUTEX(memcg_oom_mutex);
1294 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1295
1296 /*
1297  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1298  */
1299 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1300 {
1301         DEFINE_WAIT(wait);
1302         bool locked;
1303
1304         /* At first, try to OOM lock hierarchy under mem.*/
1305         mutex_lock(&memcg_oom_mutex);
1306         locked = mem_cgroup_oom_lock(mem);
1307         /*
1308          * Even if signal_pending(), we can't quit charge() loop without
1309          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1310          * under OOM is always welcomed, use TASK_KILLABLE here.
1311          */
1312         if (!locked)
1313                 prepare_to_wait(&memcg_oom_waitq, &wait, TASK_KILLABLE);
1314         mutex_unlock(&memcg_oom_mutex);
1315
1316         if (locked)
1317                 mem_cgroup_out_of_memory(mem, mask);
1318         else {
1319                 schedule();
1320                 finish_wait(&memcg_oom_waitq, &wait);
1321         }
1322         mutex_lock(&memcg_oom_mutex);
1323         mem_cgroup_oom_unlock(mem);
1324         /*
1325          * Here, we use global waitq .....more fine grained waitq ?
1326          * Assume following hierarchy.
1327          * A/
1328          *   01
1329          *   02
1330          * assume OOM happens both in A and 01 at the same time. Tthey are
1331          * mutually exclusive by lock. (kill in 01 helps A.)
1332          * When we use per memcg waitq, we have to wake up waiters on A and 02
1333          * in addtion to waiters on 01. We use global waitq for avoiding mess.
1334          * It will not be a big problem.
1335          * (And a task may be moved to other groups while it's waiting for OOM.)
1336          */
1337         wake_up_all(&memcg_oom_waitq);
1338         mutex_unlock(&memcg_oom_mutex);
1339
1340         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1341                 return false;
1342         /* Give chance to dying process */
1343         schedule_timeout(1);
1344         return true;
1345 }
1346
1347 /*
1348  * Currently used to update mapped file statistics, but the routine can be
1349  * generalized to update other statistics as well.
1350  */
1351 void mem_cgroup_update_file_mapped(struct page *page, int val)
1352 {
1353         struct mem_cgroup *mem;
1354         struct page_cgroup *pc;
1355
1356         pc = lookup_page_cgroup(page);
1357         if (unlikely(!pc))
1358                 return;
1359
1360         lock_page_cgroup(pc);
1361         mem = pc->mem_cgroup;
1362         if (!mem)
1363                 goto done;
1364
1365         if (!PageCgroupUsed(pc))
1366                 goto done;
1367
1368         /*
1369          * Preemption is already disabled. We can use __this_cpu_xxx
1370          */
1371         __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], val);
1372
1373 done:
1374         unlock_page_cgroup(pc);
1375 }
1376
1377 /*
1378  * size of first charge trial. "32" comes from vmscan.c's magic value.
1379  * TODO: maybe necessary to use big numbers in big irons.
1380  */
1381 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1382 struct memcg_stock_pcp {
1383         struct mem_cgroup *cached; /* this never be root cgroup */
1384         int charge;
1385         struct work_struct work;
1386 };
1387 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1388 static atomic_t memcg_drain_count;
1389
1390 /*
1391  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1392  * from local stock and true is returned. If the stock is 0 or charges from a
1393  * cgroup which is not current target, returns false. This stock will be
1394  * refilled.
1395  */
1396 static bool consume_stock(struct mem_cgroup *mem)
1397 {
1398         struct memcg_stock_pcp *stock;
1399         bool ret = true;
1400
1401         stock = &get_cpu_var(memcg_stock);
1402         if (mem == stock->cached && stock->charge)
1403                 stock->charge -= PAGE_SIZE;
1404         else /* need to call res_counter_charge */
1405                 ret = false;
1406         put_cpu_var(memcg_stock);
1407         return ret;
1408 }
1409
1410 /*
1411  * Returns stocks cached in percpu to res_counter and reset cached information.
1412  */
1413 static void drain_stock(struct memcg_stock_pcp *stock)
1414 {
1415         struct mem_cgroup *old = stock->cached;
1416
1417         if (stock->charge) {
1418                 res_counter_uncharge(&old->res, stock->charge);
1419                 if (do_swap_account)
1420                         res_counter_uncharge(&old->memsw, stock->charge);
1421         }
1422         stock->cached = NULL;
1423         stock->charge = 0;
1424 }
1425
1426 /*
1427  * This must be called under preempt disabled or must be called by
1428  * a thread which is pinned to local cpu.
1429  */
1430 static void drain_local_stock(struct work_struct *dummy)
1431 {
1432         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1433         drain_stock(stock);
1434 }
1435
1436 /*
1437  * Cache charges(val) which is from res_counter, to local per_cpu area.
1438  * This will be consumed by consumt_stock() function, later.
1439  */
1440 static void refill_stock(struct mem_cgroup *mem, int val)
1441 {
1442         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1443
1444         if (stock->cached != mem) { /* reset if necessary */
1445                 drain_stock(stock);
1446                 stock->cached = mem;
1447         }
1448         stock->charge += val;
1449         put_cpu_var(memcg_stock);
1450 }
1451
1452 /*
1453  * Tries to drain stocked charges in other cpus. This function is asynchronous
1454  * and just put a work per cpu for draining localy on each cpu. Caller can
1455  * expects some charges will be back to res_counter later but cannot wait for
1456  * it.
1457  */
1458 static void drain_all_stock_async(void)
1459 {
1460         int cpu;
1461         /* This function is for scheduling "drain" in asynchronous way.
1462          * The result of "drain" is not directly handled by callers. Then,
1463          * if someone is calling drain, we don't have to call drain more.
1464          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1465          * there is a race. We just do loose check here.
1466          */
1467         if (atomic_read(&memcg_drain_count))
1468                 return;
1469         /* Notify other cpus that system-wide "drain" is running */
1470         atomic_inc(&memcg_drain_count);
1471         get_online_cpus();
1472         for_each_online_cpu(cpu) {
1473                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1474                 schedule_work_on(cpu, &stock->work);
1475         }
1476         put_online_cpus();
1477         atomic_dec(&memcg_drain_count);
1478         /* We don't wait for flush_work */
1479 }
1480
1481 /* This is a synchronous drain interface. */
1482 static void drain_all_stock_sync(void)
1483 {
1484         /* called when force_empty is called */
1485         atomic_inc(&memcg_drain_count);
1486         schedule_on_each_cpu(drain_local_stock);
1487         atomic_dec(&memcg_drain_count);
1488 }
1489
1490 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1491                                         unsigned long action,
1492                                         void *hcpu)
1493 {
1494         int cpu = (unsigned long)hcpu;
1495         struct memcg_stock_pcp *stock;
1496
1497         if (action != CPU_DEAD)
1498                 return NOTIFY_OK;
1499         stock = &per_cpu(memcg_stock, cpu);
1500         drain_stock(stock);
1501         return NOTIFY_OK;
1502 }
1503
1504 /*
1505  * Unlike exported interface, "oom" parameter is added. if oom==true,
1506  * oom-killer can be invoked.
1507  */
1508 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1509                         gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1510 {
1511         struct mem_cgroup *mem, *mem_over_limit;
1512         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1513         struct res_counter *fail_res;
1514         int csize = CHARGE_SIZE;
1515
1516         /*
1517          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1518          * in system level. So, allow to go ahead dying process in addition to
1519          * MEMDIE process.
1520          */
1521         if (unlikely(test_thread_flag(TIF_MEMDIE)
1522                      || fatal_signal_pending(current)))
1523                 goto bypass;
1524
1525         /*
1526          * We always charge the cgroup the mm_struct belongs to.
1527          * The mm_struct's mem_cgroup changes on task migration if the
1528          * thread group leader migrates. It's possible that mm is not
1529          * set, if so charge the init_mm (happens for pagecache usage).
1530          */
1531         mem = *memcg;
1532         if (likely(!mem)) {
1533                 mem = try_get_mem_cgroup_from_mm(mm);
1534                 *memcg = mem;
1535         } else {
1536                 css_get(&mem->css);
1537         }
1538         if (unlikely(!mem))
1539                 return 0;
1540
1541         VM_BUG_ON(css_is_removed(&mem->css));
1542         if (mem_cgroup_is_root(mem))
1543                 goto done;
1544
1545         while (1) {
1546                 int ret = 0;
1547                 unsigned long flags = 0;
1548
1549                 if (consume_stock(mem))
1550                         goto done;
1551
1552                 ret = res_counter_charge(&mem->res, csize, &fail_res);
1553                 if (likely(!ret)) {
1554                         if (!do_swap_account)
1555                                 break;
1556                         ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1557                         if (likely(!ret))
1558                                 break;
1559                         /* mem+swap counter fails */
1560                         res_counter_uncharge(&mem->res, csize);
1561                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1562                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1563                                                                         memsw);
1564                 } else
1565                         /* mem counter fails */
1566                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1567                                                                         res);
1568
1569                 /* reduce request size and retry */
1570                 if (csize > PAGE_SIZE) {
1571                         csize = PAGE_SIZE;
1572                         continue;
1573                 }
1574                 if (!(gfp_mask & __GFP_WAIT))
1575                         goto nomem;
1576
1577                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1578                                                 gfp_mask, flags);
1579                 if (ret)
1580                         continue;
1581
1582                 /*
1583                  * try_to_free_mem_cgroup_pages() might not give us a full
1584                  * picture of reclaim. Some pages are reclaimed and might be
1585                  * moved to swap cache or just unmapped from the cgroup.
1586                  * Check the limit again to see if the reclaim reduced the
1587                  * current usage of the cgroup before giving up
1588                  *
1589                  */
1590                 if (mem_cgroup_check_under_limit(mem_over_limit))
1591                         continue;
1592
1593                 /* try to avoid oom while someone is moving charge */
1594                 if (mc.moving_task && current != mc.moving_task) {
1595                         struct mem_cgroup *from, *to;
1596                         bool do_continue = false;
1597                         /*
1598                          * There is a small race that "from" or "to" can be
1599                          * freed by rmdir, so we use css_tryget().
1600                          */
1601                         rcu_read_lock();
1602                         from = mc.from;
1603                         to = mc.to;
1604                         if (from && css_tryget(&from->css)) {
1605                                 if (mem_over_limit->use_hierarchy)
1606                                         do_continue = css_is_ancestor(
1607                                                         &from->css,
1608                                                         &mem_over_limit->css);
1609                                 else
1610                                         do_continue = (from == mem_over_limit);
1611                                 css_put(&from->css);
1612                         }
1613                         if (!do_continue && to && css_tryget(&to->css)) {
1614                                 if (mem_over_limit->use_hierarchy)
1615                                         do_continue = css_is_ancestor(
1616                                                         &to->css,
1617                                                         &mem_over_limit->css);
1618                                 else
1619                                         do_continue = (to == mem_over_limit);
1620                                 css_put(&to->css);
1621                         }
1622                         rcu_read_unlock();
1623                         if (do_continue) {
1624                                 DEFINE_WAIT(wait);
1625                                 prepare_to_wait(&mc.waitq, &wait,
1626                                                         TASK_INTERRUPTIBLE);
1627                                 /* moving charge context might have finished. */
1628                                 if (mc.moving_task)
1629                                         schedule();
1630                                 finish_wait(&mc.waitq, &wait);
1631                                 continue;
1632                         }
1633                 }
1634
1635                 if (!nr_retries--) {
1636                         if (!oom)
1637                                 goto nomem;
1638                         if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) {
1639                                 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1640                                 continue;
1641                         }
1642                         /* When we reach here, current task is dying .*/
1643                         css_put(&mem->css);
1644                         goto bypass;
1645                 }
1646         }
1647         if (csize > PAGE_SIZE)
1648                 refill_stock(mem, csize - PAGE_SIZE);
1649 done:
1650         return 0;
1651 nomem:
1652         css_put(&mem->css);
1653         return -ENOMEM;
1654 bypass:
1655         *memcg = NULL;
1656         return 0;
1657 }
1658
1659 /*
1660  * Somemtimes we have to undo a charge we got by try_charge().
1661  * This function is for that and do uncharge, put css's refcnt.
1662  * gotten by try_charge().
1663  */
1664 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1665                                                         unsigned long count)
1666 {
1667         if (!mem_cgroup_is_root(mem)) {
1668                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1669                 if (do_swap_account)
1670                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1671                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1672                 WARN_ON_ONCE(count > INT_MAX);
1673                 __css_put(&mem->css, (int)count);
1674         }
1675         /* we don't need css_put for root */
1676 }
1677
1678 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1679 {
1680         __mem_cgroup_cancel_charge(mem, 1);
1681 }
1682
1683 /*
1684  * A helper function to get mem_cgroup from ID. must be called under
1685  * rcu_read_lock(). The caller must check css_is_removed() or some if
1686  * it's concern. (dropping refcnt from swap can be called against removed
1687  * memcg.)
1688  */
1689 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1690 {
1691         struct cgroup_subsys_state *css;
1692
1693         /* ID 0 is unused ID */
1694         if (!id)
1695                 return NULL;
1696         css = css_lookup(&mem_cgroup_subsys, id);
1697         if (!css)
1698                 return NULL;
1699         return container_of(css, struct mem_cgroup, css);
1700 }
1701
1702 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1703 {
1704         struct mem_cgroup *mem = NULL;
1705         struct page_cgroup *pc;
1706         unsigned short id;
1707         swp_entry_t ent;
1708
1709         VM_BUG_ON(!PageLocked(page));
1710
1711         pc = lookup_page_cgroup(page);
1712         lock_page_cgroup(pc);
1713         if (PageCgroupUsed(pc)) {
1714                 mem = pc->mem_cgroup;
1715                 if (mem && !css_tryget(&mem->css))
1716                         mem = NULL;
1717         } else if (PageSwapCache(page)) {
1718                 ent.val = page_private(page);
1719                 id = lookup_swap_cgroup(ent);
1720                 rcu_read_lock();
1721                 mem = mem_cgroup_lookup(id);
1722                 if (mem && !css_tryget(&mem->css))
1723                         mem = NULL;
1724                 rcu_read_unlock();
1725         }
1726         unlock_page_cgroup(pc);
1727         return mem;
1728 }
1729
1730 /*
1731  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1732  * USED state. If already USED, uncharge and return.
1733  */
1734
1735 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1736                                      struct page_cgroup *pc,
1737                                      enum charge_type ctype)
1738 {
1739         /* try_charge() can return NULL to *memcg, taking care of it. */
1740         if (!mem)
1741                 return;
1742
1743         lock_page_cgroup(pc);
1744         if (unlikely(PageCgroupUsed(pc))) {
1745                 unlock_page_cgroup(pc);
1746                 mem_cgroup_cancel_charge(mem);
1747                 return;
1748         }
1749
1750         pc->mem_cgroup = mem;
1751         /*
1752          * We access a page_cgroup asynchronously without lock_page_cgroup().
1753          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1754          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1755          * before USED bit, we need memory barrier here.
1756          * See mem_cgroup_add_lru_list(), etc.
1757          */
1758         smp_wmb();
1759         switch (ctype) {
1760         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1761         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1762                 SetPageCgroupCache(pc);
1763                 SetPageCgroupUsed(pc);
1764                 break;
1765         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1766                 ClearPageCgroupCache(pc);
1767                 SetPageCgroupUsed(pc);
1768                 break;
1769         default:
1770                 break;
1771         }
1772
1773         mem_cgroup_charge_statistics(mem, pc, true);
1774
1775         unlock_page_cgroup(pc);
1776         /*
1777          * "charge_statistics" updated event counter. Then, check it.
1778          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1779          * if they exceeds softlimit.
1780          */
1781         memcg_check_events(mem, pc->page);
1782 }
1783
1784 /**
1785  * __mem_cgroup_move_account - move account of the page
1786  * @pc: page_cgroup of the page.
1787  * @from: mem_cgroup which the page is moved from.
1788  * @to: mem_cgroup which the page is moved to. @from != @to.
1789  * @uncharge: whether we should call uncharge and css_put against @from.
1790  *
1791  * The caller must confirm following.
1792  * - page is not on LRU (isolate_page() is useful.)
1793  * - the pc is locked, used, and ->mem_cgroup points to @from.
1794  *
1795  * This function doesn't do "charge" nor css_get to new cgroup. It should be
1796  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1797  * true, this function does "uncharge" from old cgroup, but it doesn't if
1798  * @uncharge is false, so a caller should do "uncharge".
1799  */
1800
1801 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1802         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1803 {
1804         struct page *page;
1805
1806         VM_BUG_ON(from == to);
1807         VM_BUG_ON(PageLRU(pc->page));
1808         VM_BUG_ON(!PageCgroupLocked(pc));
1809         VM_BUG_ON(!PageCgroupUsed(pc));
1810         VM_BUG_ON(pc->mem_cgroup != from);
1811
1812         page = pc->page;
1813         if (page_mapped(page) && !PageAnon(page)) {
1814                 /* Update mapped_file data for mem_cgroup */
1815                 preempt_disable();
1816                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1817                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1818                 preempt_enable();
1819         }
1820         mem_cgroup_charge_statistics(from, pc, false);
1821         if (uncharge)
1822                 /* This is not "cancel", but cancel_charge does all we need. */
1823                 mem_cgroup_cancel_charge(from);
1824
1825         /* caller should have done css_get */
1826         pc->mem_cgroup = to;
1827         mem_cgroup_charge_statistics(to, pc, true);
1828         /*
1829          * We charges against "to" which may not have any tasks. Then, "to"
1830          * can be under rmdir(). But in current implementation, caller of
1831          * this function is just force_empty() and move charge, so it's
1832          * garanteed that "to" is never removed. So, we don't check rmdir
1833          * status here.
1834          */
1835 }
1836
1837 /*
1838  * check whether the @pc is valid for moving account and call
1839  * __mem_cgroup_move_account()
1840  */
1841 static int mem_cgroup_move_account(struct page_cgroup *pc,
1842                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1843 {
1844         int ret = -EINVAL;
1845         lock_page_cgroup(pc);
1846         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1847                 __mem_cgroup_move_account(pc, from, to, uncharge);
1848                 ret = 0;
1849         }
1850         unlock_page_cgroup(pc);
1851         /*
1852          * check events
1853          */
1854         memcg_check_events(to, pc->page);
1855         memcg_check_events(from, pc->page);
1856         return ret;
1857 }
1858
1859 /*
1860  * move charges to its parent.
1861  */
1862
1863 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1864                                   struct mem_cgroup *child,
1865                                   gfp_t gfp_mask)
1866 {
1867         struct page *page = pc->page;
1868         struct cgroup *cg = child->css.cgroup;
1869         struct cgroup *pcg = cg->parent;
1870         struct mem_cgroup *parent;
1871         int ret;
1872
1873         /* Is ROOT ? */
1874         if (!pcg)
1875                 return -EINVAL;
1876
1877         ret = -EBUSY;
1878         if (!get_page_unless_zero(page))
1879                 goto out;
1880         if (isolate_lru_page(page))
1881                 goto put;
1882
1883         parent = mem_cgroup_from_cont(pcg);
1884         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1885         if (ret || !parent)
1886                 goto put_back;
1887
1888         ret = mem_cgroup_move_account(pc, child, parent, true);
1889         if (ret)
1890                 mem_cgroup_cancel_charge(parent);
1891 put_back:
1892         putback_lru_page(page);
1893 put:
1894         put_page(page);
1895 out:
1896         return ret;
1897 }
1898
1899 /*
1900  * Charge the memory controller for page usage.
1901  * Return
1902  * 0 if the charge was successful
1903  * < 0 if the cgroup is over its limit
1904  */
1905 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1906                                 gfp_t gfp_mask, enum charge_type ctype,
1907                                 struct mem_cgroup *memcg)
1908 {
1909         struct mem_cgroup *mem;
1910         struct page_cgroup *pc;
1911         int ret;
1912
1913         pc = lookup_page_cgroup(page);
1914         /* can happen at boot */
1915         if (unlikely(!pc))
1916                 return 0;
1917         prefetchw(pc);
1918
1919         mem = memcg;
1920         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1921         if (ret || !mem)
1922                 return ret;
1923
1924         __mem_cgroup_commit_charge(mem, pc, ctype);
1925         return 0;
1926 }
1927
1928 int mem_cgroup_newpage_charge(struct page *page,
1929                               struct mm_struct *mm, gfp_t gfp_mask)
1930 {
1931         if (mem_cgroup_disabled())
1932                 return 0;
1933         if (PageCompound(page))
1934                 return 0;
1935         /*
1936          * If already mapped, we don't have to account.
1937          * If page cache, page->mapping has address_space.
1938          * But page->mapping may have out-of-use anon_vma pointer,
1939          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1940          * is NULL.
1941          */
1942         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1943                 return 0;
1944         if (unlikely(!mm))
1945                 mm = &init_mm;
1946         return mem_cgroup_charge_common(page, mm, gfp_mask,
1947                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1948 }
1949
1950 static void
1951 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1952                                         enum charge_type ctype);
1953
1954 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1955                                 gfp_t gfp_mask)
1956 {
1957         struct mem_cgroup *mem = NULL;
1958         int ret;
1959
1960         if (mem_cgroup_disabled())
1961                 return 0;
1962         if (PageCompound(page))
1963                 return 0;
1964         /*
1965          * Corner case handling. This is called from add_to_page_cache()
1966          * in usual. But some FS (shmem) precharges this page before calling it
1967          * and call add_to_page_cache() with GFP_NOWAIT.
1968          *
1969          * For GFP_NOWAIT case, the page may be pre-charged before calling
1970          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1971          * charge twice. (It works but has to pay a bit larger cost.)
1972          * And when the page is SwapCache, it should take swap information
1973          * into account. This is under lock_page() now.
1974          */
1975         if (!(gfp_mask & __GFP_WAIT)) {
1976                 struct page_cgroup *pc;
1977
1978
1979                 pc = lookup_page_cgroup(page);
1980                 if (!pc)
1981                         return 0;
1982                 lock_page_cgroup(pc);
1983                 if (PageCgroupUsed(pc)) {
1984                         unlock_page_cgroup(pc);
1985                         return 0;
1986                 }
1987                 unlock_page_cgroup(pc);
1988         }
1989
1990         if (unlikely(!mm && !mem))
1991                 mm = &init_mm;
1992
1993         if (page_is_file_cache(page))
1994                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1995                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1996
1997         /* shmem */
1998         if (PageSwapCache(page)) {
1999                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2000                 if (!ret)
2001                         __mem_cgroup_commit_charge_swapin(page, mem,
2002                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2003         } else
2004                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2005                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
2006
2007         return ret;
2008 }
2009
2010 /*
2011  * While swap-in, try_charge -> commit or cancel, the page is locked.
2012  * And when try_charge() successfully returns, one refcnt to memcg without
2013  * struct page_cgroup is acquired. This refcnt will be consumed by
2014  * "commit()" or removed by "cancel()"
2015  */
2016 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2017                                  struct page *page,
2018                                  gfp_t mask, struct mem_cgroup **ptr)
2019 {
2020         struct mem_cgroup *mem;
2021         int ret;
2022
2023         if (mem_cgroup_disabled())
2024                 return 0;
2025
2026         if (!do_swap_account)
2027                 goto charge_cur_mm;
2028         /*
2029          * A racing thread's fault, or swapoff, may have already updated
2030          * the pte, and even removed page from swap cache: in those cases
2031          * do_swap_page()'s pte_same() test will fail; but there's also a
2032          * KSM case which does need to charge the page.
2033          */
2034         if (!PageSwapCache(page))
2035                 goto charge_cur_mm;
2036         mem = try_get_mem_cgroup_from_page(page);
2037         if (!mem)
2038                 goto charge_cur_mm;
2039         *ptr = mem;
2040         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2041         /* drop extra refcnt from tryget */
2042         css_put(&mem->css);
2043         return ret;
2044 charge_cur_mm:
2045         if (unlikely(!mm))
2046                 mm = &init_mm;
2047         return __mem_cgroup_try_charge(mm, mask, ptr, true);
2048 }
2049
2050 static void
2051 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2052                                         enum charge_type ctype)
2053 {
2054         struct page_cgroup *pc;
2055
2056         if (mem_cgroup_disabled())
2057                 return;
2058         if (!ptr)
2059                 return;
2060         cgroup_exclude_rmdir(&ptr->css);
2061         pc = lookup_page_cgroup(page);
2062         mem_cgroup_lru_del_before_commit_swapcache(page);
2063         __mem_cgroup_commit_charge(ptr, pc, ctype);
2064         mem_cgroup_lru_add_after_commit_swapcache(page);
2065         /*
2066          * Now swap is on-memory. This means this page may be
2067          * counted both as mem and swap....double count.
2068          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2069          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2070          * may call delete_from_swap_cache() before reach here.
2071          */
2072         if (do_swap_account && PageSwapCache(page)) {
2073                 swp_entry_t ent = {.val = page_private(page)};
2074                 unsigned short id;
2075                 struct mem_cgroup *memcg;
2076
2077                 id = swap_cgroup_record(ent, 0);
2078                 rcu_read_lock();
2079                 memcg = mem_cgroup_lookup(id);
2080                 if (memcg) {
2081                         /*
2082                          * This recorded memcg can be obsolete one. So, avoid
2083                          * calling css_tryget
2084                          */
2085                         if (!mem_cgroup_is_root(memcg))
2086                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2087                         mem_cgroup_swap_statistics(memcg, false);
2088                         mem_cgroup_put(memcg);
2089                 }
2090                 rcu_read_unlock();
2091         }
2092         /*
2093          * At swapin, we may charge account against cgroup which has no tasks.
2094          * So, rmdir()->pre_destroy() can be called while we do this charge.
2095          * In that case, we need to call pre_destroy() again. check it here.
2096          */
2097         cgroup_release_and_wakeup_rmdir(&ptr->css);
2098 }
2099
2100 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2101 {
2102         __mem_cgroup_commit_charge_swapin(page, ptr,
2103                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2104 }
2105
2106 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2107 {
2108         if (mem_cgroup_disabled())
2109                 return;
2110         if (!mem)
2111                 return;
2112         mem_cgroup_cancel_charge(mem);
2113 }
2114
2115 static void
2116 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2117 {
2118         struct memcg_batch_info *batch = NULL;
2119         bool uncharge_memsw = true;
2120         /* If swapout, usage of swap doesn't decrease */
2121         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2122                 uncharge_memsw = false;
2123         /*
2124          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2125          * In those cases, all pages freed continously can be expected to be in
2126          * the same cgroup and we have chance to coalesce uncharges.
2127          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2128          * because we want to do uncharge as soon as possible.
2129          */
2130         if (!current->memcg_batch.do_batch || test_thread_flag(TIF_MEMDIE))
2131                 goto direct_uncharge;
2132
2133         batch = &current->memcg_batch;
2134         /*
2135          * In usual, we do css_get() when we remember memcg pointer.
2136          * But in this case, we keep res->usage until end of a series of
2137          * uncharges. Then, it's ok to ignore memcg's refcnt.
2138          */
2139         if (!batch->memcg)
2140                 batch->memcg = mem;
2141         /*
2142          * In typical case, batch->memcg == mem. This means we can
2143          * merge a series of uncharges to an uncharge of res_counter.
2144          * If not, we uncharge res_counter ony by one.
2145          */
2146         if (batch->memcg != mem)
2147                 goto direct_uncharge;
2148         /* remember freed charge and uncharge it later */
2149         batch->bytes += PAGE_SIZE;
2150         if (uncharge_memsw)
2151                 batch->memsw_bytes += PAGE_SIZE;
2152         return;
2153 direct_uncharge:
2154         res_counter_uncharge(&mem->res, PAGE_SIZE);
2155         if (uncharge_memsw)
2156                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2157         return;
2158 }
2159
2160 /*
2161  * uncharge if !page_mapped(page)
2162  */
2163 static struct mem_cgroup *
2164 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2165 {
2166         struct page_cgroup *pc;
2167         struct mem_cgroup *mem = NULL;
2168         struct mem_cgroup_per_zone *mz;
2169
2170         if (mem_cgroup_disabled())
2171                 return NULL;
2172
2173         if (PageSwapCache(page))
2174                 return NULL;
2175
2176         /*
2177          * Check if our page_cgroup is valid
2178          */
2179         pc = lookup_page_cgroup(page);
2180         if (unlikely(!pc || !PageCgroupUsed(pc)))
2181                 return NULL;
2182
2183         lock_page_cgroup(pc);
2184
2185         mem = pc->mem_cgroup;
2186
2187         if (!PageCgroupUsed(pc))
2188                 goto unlock_out;
2189
2190         switch (ctype) {
2191         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2192         case MEM_CGROUP_CHARGE_TYPE_DROP:
2193                 if (page_mapped(page))
2194                         goto unlock_out;
2195                 break;
2196         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2197                 if (!PageAnon(page)) {  /* Shared memory */
2198                         if (page->mapping && !page_is_file_cache(page))
2199                                 goto unlock_out;
2200                 } else if (page_mapped(page)) /* Anon */
2201                                 goto unlock_out;
2202                 break;
2203         default:
2204                 break;
2205         }
2206
2207         if (!mem_cgroup_is_root(mem))
2208                 __do_uncharge(mem, ctype);
2209         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2210                 mem_cgroup_swap_statistics(mem, true);
2211         mem_cgroup_charge_statistics(mem, pc, false);
2212
2213         ClearPageCgroupUsed(pc);
2214         /*
2215          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2216          * freed from LRU. This is safe because uncharged page is expected not
2217          * to be reused (freed soon). Exception is SwapCache, it's handled by
2218          * special functions.
2219          */
2220
2221         mz = page_cgroup_zoneinfo(pc);
2222         unlock_page_cgroup(pc);
2223
2224         memcg_check_events(mem, page);
2225         /* at swapout, this memcg will be accessed to record to swap */
2226         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2227                 css_put(&mem->css);
2228
2229         return mem;
2230
2231 unlock_out:
2232         unlock_page_cgroup(pc);
2233         return NULL;
2234 }
2235
2236 void mem_cgroup_uncharge_page(struct page *page)
2237 {
2238         /* early check. */
2239         if (page_mapped(page))
2240                 return;
2241         if (page->mapping && !PageAnon(page))
2242                 return;
2243         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2244 }
2245
2246 void mem_cgroup_uncharge_cache_page(struct page *page)
2247 {
2248         VM_BUG_ON(page_mapped(page));
2249         VM_BUG_ON(page->mapping);
2250         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2251 }
2252
2253 /*
2254  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2255  * In that cases, pages are freed continuously and we can expect pages
2256  * are in the same memcg. All these calls itself limits the number of
2257  * pages freed at once, then uncharge_start/end() is called properly.
2258  * This may be called prural(2) times in a context,
2259  */
2260
2261 void mem_cgroup_uncharge_start(void)
2262 {
2263         current->memcg_batch.do_batch++;
2264         /* We can do nest. */
2265         if (current->memcg_batch.do_batch == 1) {
2266                 current->memcg_batch.memcg = NULL;
2267                 current->memcg_batch.bytes = 0;
2268                 current->memcg_batch.memsw_bytes = 0;
2269         }
2270 }
2271
2272 void mem_cgroup_uncharge_end(void)
2273 {
2274         struct memcg_batch_info *batch = &current->memcg_batch;
2275
2276         if (!batch->do_batch)
2277                 return;
2278
2279         batch->do_batch--;
2280         if (batch->do_batch) /* If stacked, do nothing. */
2281                 return;
2282
2283         if (!batch->memcg)
2284                 return;
2285         /*
2286          * This "batch->memcg" is valid without any css_get/put etc...
2287          * bacause we hide charges behind us.
2288          */
2289         if (batch->bytes)
2290                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2291         if (batch->memsw_bytes)
2292                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2293         /* forget this pointer (for sanity check) */
2294         batch->memcg = NULL;
2295 }
2296
2297 #ifdef CONFIG_SWAP
2298 /*
2299  * called after __delete_from_swap_cache() and drop "page" account.
2300  * memcg information is recorded to swap_cgroup of "ent"
2301  */
2302 void
2303 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2304 {
2305         struct mem_cgroup *memcg;
2306         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2307
2308         if (!swapout) /* this was a swap cache but the swap is unused ! */
2309                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2310
2311         memcg = __mem_cgroup_uncharge_common(page, ctype);
2312
2313         /* record memcg information */
2314         if (do_swap_account && swapout && memcg) {
2315                 swap_cgroup_record(ent, css_id(&memcg->css));
2316                 mem_cgroup_get(memcg);
2317         }
2318         if (swapout && memcg)
2319                 css_put(&memcg->css);
2320 }
2321 #endif
2322
2323 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2324 /*
2325  * called from swap_entry_free(). remove record in swap_cgroup and
2326  * uncharge "memsw" account.
2327  */
2328 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2329 {
2330         struct mem_cgroup *memcg;
2331         unsigned short id;
2332
2333         if (!do_swap_account)
2334                 return;
2335
2336         id = swap_cgroup_record(ent, 0);
2337         rcu_read_lock();
2338         memcg = mem_cgroup_lookup(id);
2339         if (memcg) {
2340                 /*
2341                  * We uncharge this because swap is freed.
2342                  * This memcg can be obsolete one. We avoid calling css_tryget
2343                  */
2344                 if (!mem_cgroup_is_root(memcg))
2345                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2346                 mem_cgroup_swap_statistics(memcg, false);
2347                 mem_cgroup_put(memcg);
2348         }
2349         rcu_read_unlock();
2350 }
2351
2352 /**
2353  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2354  * @entry: swap entry to be moved
2355  * @from:  mem_cgroup which the entry is moved from
2356  * @to:  mem_cgroup which the entry is moved to
2357  * @need_fixup: whether we should fixup res_counters and refcounts.
2358  *
2359  * It succeeds only when the swap_cgroup's record for this entry is the same
2360  * as the mem_cgroup's id of @from.
2361  *
2362  * Returns 0 on success, -EINVAL on failure.
2363  *
2364  * The caller must have charged to @to, IOW, called res_counter_charge() about
2365  * both res and memsw, and called css_get().
2366  */
2367 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2368                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2369 {
2370         unsigned short old_id, new_id;
2371
2372         old_id = css_id(&from->css);
2373         new_id = css_id(&to->css);
2374
2375         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2376                 mem_cgroup_swap_statistics(from, false);
2377                 mem_cgroup_swap_statistics(to, true);
2378                 /*
2379                  * This function is only called from task migration context now.
2380                  * It postpones res_counter and refcount handling till the end
2381                  * of task migration(mem_cgroup_clear_mc()) for performance
2382                  * improvement. But we cannot postpone mem_cgroup_get(to)
2383                  * because if the process that has been moved to @to does
2384                  * swap-in, the refcount of @to might be decreased to 0.
2385                  */
2386                 mem_cgroup_get(to);
2387                 if (need_fixup) {
2388                         if (!mem_cgroup_is_root(from))
2389                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2390                         mem_cgroup_put(from);
2391                         /*
2392                          * we charged both to->res and to->memsw, so we should
2393                          * uncharge to->res.
2394                          */
2395                         if (!mem_cgroup_is_root(to))
2396                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2397                         css_put(&to->css);
2398                 }
2399                 return 0;
2400         }
2401         return -EINVAL;
2402 }
2403 #else
2404 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2405                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2406 {
2407         return -EINVAL;
2408 }
2409 #endif
2410
2411 /*
2412  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2413  * page belongs to.
2414  */
2415 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2416 {
2417         struct page_cgroup *pc;
2418         struct mem_cgroup *mem = NULL;
2419         int ret = 0;
2420
2421         if (mem_cgroup_disabled())
2422                 return 0;
2423
2424         pc = lookup_page_cgroup(page);
2425         lock_page_cgroup(pc);
2426         if (PageCgroupUsed(pc)) {
2427                 mem = pc->mem_cgroup;
2428                 css_get(&mem->css);
2429         }
2430         unlock_page_cgroup(pc);
2431
2432         if (mem) {
2433                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
2434                 css_put(&mem->css);
2435         }
2436         *ptr = mem;
2437         return ret;
2438 }
2439
2440 /* remove redundant charge if migration failed*/
2441 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2442                 struct page *oldpage, struct page *newpage)
2443 {
2444         struct page *target, *unused;
2445         struct page_cgroup *pc;
2446         enum charge_type ctype;
2447
2448         if (!mem)
2449                 return;
2450         cgroup_exclude_rmdir(&mem->css);
2451         /* at migration success, oldpage->mapping is NULL. */
2452         if (oldpage->mapping) {
2453                 target = oldpage;
2454                 unused = NULL;
2455         } else {
2456                 target = newpage;
2457                 unused = oldpage;
2458         }
2459
2460         if (PageAnon(target))
2461                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2462         else if (page_is_file_cache(target))
2463                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2464         else
2465                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2466
2467         /* unused page is not on radix-tree now. */
2468         if (unused)
2469                 __mem_cgroup_uncharge_common(unused, ctype);
2470
2471         pc = lookup_page_cgroup(target);
2472         /*
2473          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2474          * So, double-counting is effectively avoided.
2475          */
2476         __mem_cgroup_commit_charge(mem, pc, ctype);
2477
2478         /*
2479          * Both of oldpage and newpage are still under lock_page().
2480          * Then, we don't have to care about race in radix-tree.
2481          * But we have to be careful that this page is unmapped or not.
2482          *
2483          * There is a case for !page_mapped(). At the start of
2484          * migration, oldpage was mapped. But now, it's zapped.
2485          * But we know *target* page is not freed/reused under us.
2486          * mem_cgroup_uncharge_page() does all necessary checks.
2487          */
2488         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2489                 mem_cgroup_uncharge_page(target);
2490         /*
2491          * At migration, we may charge account against cgroup which has no tasks
2492          * So, rmdir()->pre_destroy() can be called while we do this charge.
2493          * In that case, we need to call pre_destroy() again. check it here.
2494          */
2495         cgroup_release_and_wakeup_rmdir(&mem->css);
2496 }
2497
2498 /*
2499  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2500  * Calling hierarchical_reclaim is not enough because we should update
2501  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2502  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2503  * not from the memcg which this page would be charged to.
2504  * try_charge_swapin does all of these works properly.
2505  */
2506 int mem_cgroup_shmem_charge_fallback(struct page *page,
2507                             struct mm_struct *mm,
2508                             gfp_t gfp_mask)
2509 {
2510         struct mem_cgroup *mem = NULL;
2511         int ret;
2512
2513         if (mem_cgroup_disabled())
2514                 return 0;
2515
2516         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2517         if (!ret)
2518                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2519
2520         return ret;
2521 }
2522
2523 static DEFINE_MUTEX(set_limit_mutex);
2524
2525 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2526                                 unsigned long long val)
2527 {
2528         int retry_count;
2529         u64 memswlimit;
2530         int ret = 0;
2531         int children = mem_cgroup_count_children(memcg);
2532         u64 curusage, oldusage;
2533
2534         /*
2535          * For keeping hierarchical_reclaim simple, how long we should retry
2536          * is depends on callers. We set our retry-count to be function
2537          * of # of children which we should visit in this loop.
2538          */
2539         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2540
2541         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2542
2543         while (retry_count) {
2544                 if (signal_pending(current)) {
2545                         ret = -EINTR;
2546                         break;
2547                 }
2548                 /*
2549                  * Rather than hide all in some function, I do this in
2550                  * open coded manner. You see what this really does.
2551                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2552                  */
2553                 mutex_lock(&set_limit_mutex);
2554                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2555                 if (memswlimit < val) {
2556                         ret = -EINVAL;
2557                         mutex_unlock(&set_limit_mutex);
2558                         break;
2559                 }
2560                 ret = res_counter_set_limit(&memcg->res, val);
2561                 if (!ret) {
2562                         if (memswlimit == val)
2563                                 memcg->memsw_is_minimum = true;
2564                         else
2565                                 memcg->memsw_is_minimum = false;
2566                 }
2567                 mutex_unlock(&set_limit_mutex);
2568
2569                 if (!ret)
2570                         break;
2571
2572                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2573                                                 MEM_CGROUP_RECLAIM_SHRINK);
2574                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2575                 /* Usage is reduced ? */
2576                 if (curusage >= oldusage)
2577                         retry_count--;
2578                 else
2579                         oldusage = curusage;
2580         }
2581
2582         return ret;
2583 }
2584
2585 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2586                                         unsigned long long val)
2587 {
2588         int retry_count;
2589         u64 memlimit, oldusage, curusage;
2590         int children = mem_cgroup_count_children(memcg);
2591         int ret = -EBUSY;
2592
2593         /* see mem_cgroup_resize_res_limit */
2594         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2595         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2596         while (retry_count) {
2597                 if (signal_pending(current)) {
2598                         ret = -EINTR;
2599                         break;
2600                 }
2601                 /*
2602                  * Rather than hide all in some function, I do this in
2603                  * open coded manner. You see what this really does.
2604                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2605                  */
2606                 mutex_lock(&set_limit_mutex);
2607                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2608                 if (memlimit > val) {
2609                         ret = -EINVAL;
2610                         mutex_unlock(&set_limit_mutex);
2611                         break;
2612                 }
2613                 ret = res_counter_set_limit(&memcg->memsw, val);
2614                 if (!ret) {
2615                         if (memlimit == val)
2616                                 memcg->memsw_is_minimum = true;
2617                         else
2618                                 memcg->memsw_is_minimum = false;
2619                 }
2620                 mutex_unlock(&set_limit_mutex);
2621
2622                 if (!ret)
2623                         break;
2624
2625                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2626                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2627                                                 MEM_CGROUP_RECLAIM_SHRINK);
2628                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2629                 /* Usage is reduced ? */
2630                 if (curusage >= oldusage)
2631                         retry_count--;
2632                 else
2633                         oldusage = curusage;
2634         }
2635         return ret;
2636 }
2637
2638 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2639                                                 gfp_t gfp_mask, int nid,
2640                                                 int zid)
2641 {
2642         unsigned long nr_reclaimed = 0;
2643         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2644         unsigned long reclaimed;
2645         int loop = 0;
2646         struct mem_cgroup_tree_per_zone *mctz;
2647         unsigned long long excess;
2648
2649         if (order > 0)
2650                 return 0;
2651
2652         mctz = soft_limit_tree_node_zone(nid, zid);
2653         /*
2654          * This loop can run a while, specially if mem_cgroup's continuously
2655          * keep exceeding their soft limit and putting the system under
2656          * pressure
2657          */
2658         do {
2659                 if (next_mz)
2660                         mz = next_mz;
2661                 else
2662                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2663                 if (!mz)
2664                         break;
2665
2666                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2667                                                 gfp_mask,
2668                                                 MEM_CGROUP_RECLAIM_SOFT);
2669                 nr_reclaimed += reclaimed;
2670                 spin_lock(&mctz->lock);
2671
2672                 /*
2673                  * If we failed to reclaim anything from this memory cgroup
2674                  * it is time to move on to the next cgroup
2675                  */
2676                 next_mz = NULL;
2677                 if (!reclaimed) {
2678                         do {
2679                                 /*
2680                                  * Loop until we find yet another one.
2681                                  *
2682                                  * By the time we get the soft_limit lock
2683                                  * again, someone might have aded the
2684                                  * group back on the RB tree. Iterate to
2685                                  * make sure we get a different mem.
2686                                  * mem_cgroup_largest_soft_limit_node returns
2687                                  * NULL if no other cgroup is present on
2688                                  * the tree
2689                                  */
2690                                 next_mz =
2691                                 __mem_cgroup_largest_soft_limit_node(mctz);
2692                                 if (next_mz == mz) {
2693                                         css_put(&next_mz->mem->css);
2694                                         next_mz = NULL;
2695                                 } else /* next_mz == NULL or other memcg */
2696                                         break;
2697                         } while (1);
2698                 }
2699                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2700                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2701                 /*
2702                  * One school of thought says that we should not add
2703                  * back the node to the tree if reclaim returns 0.
2704                  * But our reclaim could return 0, simply because due
2705                  * to priority we are exposing a smaller subset of
2706                  * memory to reclaim from. Consider this as a longer
2707                  * term TODO.
2708                  */
2709                 /* If excess == 0, no tree ops */
2710                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2711                 spin_unlock(&mctz->lock);
2712                 css_put(&mz->mem->css);
2713                 loop++;
2714                 /*
2715                  * Could not reclaim anything and there are no more
2716                  * mem cgroups to try or we seem to be looping without
2717                  * reclaiming anything.
2718                  */
2719                 if (!nr_reclaimed &&
2720                         (next_mz == NULL ||
2721                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2722                         break;
2723         } while (!nr_reclaimed);
2724         if (next_mz)
2725                 css_put(&next_mz->mem->css);
2726         return nr_reclaimed;
2727 }
2728
2729 /*
2730  * This routine traverse page_cgroup in given list and drop them all.
2731  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2732  */
2733 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2734                                 int node, int zid, enum lru_list lru)
2735 {
2736         struct zone *zone;
2737         struct mem_cgroup_per_zone *mz;
2738         struct page_cgroup *pc, *busy;
2739         unsigned long flags, loop;
2740         struct list_head *list;
2741         int ret = 0;
2742
2743         zone = &NODE_DATA(node)->node_zones[zid];
2744         mz = mem_cgroup_zoneinfo(mem, node, zid);
2745         list = &mz->lists[lru];
2746
2747         loop = MEM_CGROUP_ZSTAT(mz, lru);
2748         /* give some margin against EBUSY etc...*/
2749         loop += 256;
2750         busy = NULL;
2751         while (loop--) {
2752                 ret = 0;
2753                 spin_lock_irqsave(&zone->lru_lock, flags);
2754                 if (list_empty(list)) {
2755                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2756                         break;
2757                 }
2758                 pc = list_entry(list->prev, struct page_cgroup, lru);
2759                 if (busy == pc) {
2760                         list_move(&pc->lru, list);
2761                         busy = NULL;
2762                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2763                         continue;
2764                 }
2765                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2766
2767                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2768                 if (ret == -ENOMEM)
2769                         break;
2770
2771                 if (ret == -EBUSY || ret == -EINVAL) {
2772                         /* found lock contention or "pc" is obsolete. */
2773                         busy = pc;
2774                         cond_resched();
2775                 } else
2776                         busy = NULL;
2777         }
2778
2779         if (!ret && !list_empty(list))
2780                 return -EBUSY;
2781         return ret;
2782 }
2783
2784 /*
2785  * make mem_cgroup's charge to be 0 if there is no task.
2786  * This enables deleting this mem_cgroup.
2787  */
2788 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2789 {
2790         int ret;
2791         int node, zid, shrink;
2792         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2793         struct cgroup *cgrp = mem->css.cgroup;
2794
2795         css_get(&mem->css);
2796
2797         shrink = 0;
2798         /* should free all ? */
2799         if (free_all)
2800                 goto try_to_free;
2801 move_account:
2802         do {
2803                 ret = -EBUSY;
2804                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2805                         goto out;
2806                 ret = -EINTR;
2807                 if (signal_pending(current))
2808                         goto out;
2809                 /* This is for making all *used* pages to be on LRU. */
2810                 lru_add_drain_all();
2811                 drain_all_stock_sync();
2812                 ret = 0;
2813                 for_each_node_state(node, N_HIGH_MEMORY) {
2814                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2815                                 enum lru_list l;
2816                                 for_each_lru(l) {
2817                                         ret = mem_cgroup_force_empty_list(mem,
2818                                                         node, zid, l);
2819                                         if (ret)
2820                                                 break;
2821                                 }
2822                         }
2823                         if (ret)
2824                                 break;
2825                 }
2826                 /* it seems parent cgroup doesn't have enough mem */
2827                 if (ret == -ENOMEM)
2828                         goto try_to_free;
2829                 cond_resched();
2830         /* "ret" should also be checked to ensure all lists are empty. */
2831         } while (mem->res.usage > 0 || ret);
2832 out:
2833         css_put(&mem->css);
2834         return ret;
2835
2836 try_to_free:
2837         /* returns EBUSY if there is a task or if we come here twice. */
2838         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2839                 ret = -EBUSY;
2840                 goto out;
2841         }
2842         /* we call try-to-free pages for make this cgroup empty */
2843         lru_add_drain_all();
2844         /* try to free all pages in this cgroup */
2845         shrink = 1;
2846         while (nr_retries && mem->res.usage > 0) {
2847                 int progress;
2848
2849                 if (signal_pending(current)) {
2850                         ret = -EINTR;
2851                         goto out;
2852                 }
2853                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2854                                                 false, get_swappiness(mem));
2855                 if (!progress) {
2856                         nr_retries--;
2857                         /* maybe some writeback is necessary */
2858                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2859                 }
2860
2861         }
2862         lru_add_drain();
2863         /* try move_account...there may be some *locked* pages. */
2864         goto move_account;
2865 }
2866
2867 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2868 {
2869         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2870 }
2871
2872
2873 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2874 {
2875         return mem_cgroup_from_cont(cont)->use_hierarchy;
2876 }
2877
2878 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2879                                         u64 val)
2880 {
2881         int retval = 0;
2882         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2883         struct cgroup *parent = cont->parent;
2884         struct mem_cgroup *parent_mem = NULL;
2885
2886         if (parent)
2887                 parent_mem = mem_cgroup_from_cont(parent);
2888
2889         cgroup_lock();
2890         /*
2891          * If parent's use_hierarchy is set, we can't make any modifications
2892          * in the child subtrees. If it is unset, then the change can
2893          * occur, provided the current cgroup has no children.
2894          *
2895          * For the root cgroup, parent_mem is NULL, we allow value to be
2896          * set if there are no children.
2897          */
2898         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2899                                 (val == 1 || val == 0)) {
2900                 if (list_empty(&cont->children))
2901                         mem->use_hierarchy = val;
2902                 else
2903                         retval = -EBUSY;
2904         } else
2905                 retval = -EINVAL;
2906         cgroup_unlock();
2907
2908         return retval;
2909 }
2910
2911 struct mem_cgroup_idx_data {
2912         s64 val;
2913         enum mem_cgroup_stat_index idx;
2914 };
2915
2916 static int
2917 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2918 {
2919         struct mem_cgroup_idx_data *d = data;
2920         d->val += mem_cgroup_read_stat(mem, d->idx);
2921         return 0;
2922 }
2923
2924 static void
2925 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2926                                 enum mem_cgroup_stat_index idx, s64 *val)
2927 {
2928         struct mem_cgroup_idx_data d;
2929         d.idx = idx;
2930         d.val = 0;
2931         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2932         *val = d.val;
2933 }
2934
2935 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
2936 {
2937         u64 idx_val, val;
2938
2939         if (!mem_cgroup_is_root(mem)) {
2940                 if (!swap)
2941                         return res_counter_read_u64(&mem->res, RES_USAGE);
2942                 else
2943                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
2944         }
2945
2946         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
2947         val = idx_val;
2948         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
2949         val += idx_val;
2950
2951         if (swap) {
2952                 mem_cgroup_get_recursive_idx_stat(mem,
2953                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2954                 val += idx_val;
2955         }
2956
2957         return val << PAGE_SHIFT;
2958 }
2959
2960 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2961 {
2962         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2963         u64 val;
2964         int type, name;
2965
2966         type = MEMFILE_TYPE(cft->private);
2967         name = MEMFILE_ATTR(cft->private);
2968         switch (type) {
2969         case _MEM:
2970                 if (name == RES_USAGE)
2971                         val = mem_cgroup_usage(mem, false);
2972                 else
2973                         val = res_counter_read_u64(&mem->res, name);
2974                 break;
2975         case _MEMSWAP:
2976                 if (name == RES_USAGE)
2977                         val = mem_cgroup_usage(mem, true);
2978                 else
2979                         val = res_counter_read_u64(&mem->memsw, name);
2980                 break;
2981         default:
2982                 BUG();
2983                 break;
2984         }
2985         return val;
2986 }
2987 /*
2988  * The user of this function is...
2989  * RES_LIMIT.
2990  */
2991 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2992                             const char *buffer)
2993 {
2994         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2995         int type, name;
2996         unsigned long long val;
2997         int ret;
2998
2999         type = MEMFILE_TYPE(cft->private);
3000         name = MEMFILE_ATTR(cft->private);
3001         switch (name) {
3002         case RES_LIMIT:
3003                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3004                         ret = -EINVAL;
3005                         break;
3006                 }
3007                 /* This function does all necessary parse...reuse it */
3008                 ret = res_counter_memparse_write_strategy(buffer, &val);
3009                 if (ret)
3010                         break;
3011                 if (type == _MEM)
3012                         ret = mem_cgroup_resize_limit(memcg, val);
3013                 else
3014                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3015                 break;
3016         case RES_SOFT_LIMIT:
3017                 ret = res_counter_memparse_write_strategy(buffer, &val);
3018                 if (ret)
3019                         break;
3020                 /*
3021                  * For memsw, soft limits are hard to implement in terms
3022                  * of semantics, for now, we support soft limits for
3023                  * control without swap
3024                  */
3025                 if (type == _MEM)
3026                         ret = res_counter_set_soft_limit(&memcg->res, val);
3027                 else
3028                         ret = -EINVAL;
3029                 break;
3030         default:
3031                 ret = -EINVAL; /* should be BUG() ? */
3032                 break;
3033         }
3034         return ret;
3035 }
3036
3037 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3038                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3039 {
3040         struct cgroup *cgroup;
3041         unsigned long long min_limit, min_memsw_limit, tmp;
3042
3043         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3044         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3045         cgroup = memcg->css.cgroup;
3046         if (!memcg->use_hierarchy)
3047                 goto out;
3048
3049         while (cgroup->parent) {
3050                 cgroup = cgroup->parent;
3051                 memcg = mem_cgroup_from_cont(cgroup);
3052                 if (!memcg->use_hierarchy)
3053                         break;
3054                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3055                 min_limit = min(min_limit, tmp);
3056                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3057                 min_memsw_limit = min(min_memsw_limit, tmp);
3058         }
3059 out:
3060         *mem_limit = min_limit;
3061         *memsw_limit = min_memsw_limit;
3062         return;
3063 }
3064
3065 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3066 {
3067         struct mem_cgroup *mem;
3068         int type, name;
3069
3070         mem = mem_cgroup_from_cont(cont);
3071         type = MEMFILE_TYPE(event);
3072         name = MEMFILE_ATTR(event);
3073         switch (name) {
3074         case RES_MAX_USAGE:
3075                 if (type == _MEM)
3076                         res_counter_reset_max(&mem->res);
3077                 else
3078                         res_counter_reset_max(&mem->memsw);
3079                 break;
3080         case RES_FAILCNT:
3081                 if (type == _MEM)
3082                         res_counter_reset_failcnt(&mem->res);
3083                 else
3084                         res_counter_reset_failcnt(&mem->memsw);
3085                 break;
3086         }
3087
3088         return 0;
3089 }
3090
3091 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3092                                         struct cftype *cft)
3093 {
3094         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3095 }
3096
3097 #ifdef CONFIG_MMU
3098 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3099                                         struct cftype *cft, u64 val)
3100 {
3101         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3102
3103         if (val >= (1 << NR_MOVE_TYPE))
3104                 return -EINVAL;
3105         /*
3106          * We check this value several times in both in can_attach() and
3107          * attach(), so we need cgroup lock to prevent this value from being
3108          * inconsistent.
3109          */
3110         cgroup_lock();
3111         mem->move_charge_at_immigrate = val;
3112         cgroup_unlock();
3113
3114         return 0;
3115 }
3116 #else
3117 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3118                                         struct cftype *cft, u64 val)
3119 {
3120         return -ENOSYS;
3121 }
3122 #endif
3123
3124
3125 /* For read statistics */
3126 enum {
3127         MCS_CACHE,
3128         MCS_RSS,
3129         MCS_FILE_MAPPED,
3130         MCS_PGPGIN,
3131         MCS_PGPGOUT,
3132         MCS_SWAP,
3133         MCS_INACTIVE_ANON,
3134         MCS_ACTIVE_ANON,
3135         MCS_INACTIVE_FILE,
3136         MCS_ACTIVE_FILE,
3137         MCS_UNEVICTABLE,
3138         NR_MCS_STAT,
3139 };
3140
3141 struct mcs_total_stat {
3142         s64 stat[NR_MCS_STAT];
3143 };
3144
3145 struct {
3146         char *local_name;
3147         char *total_name;
3148 } memcg_stat_strings[NR_MCS_STAT] = {
3149         {"cache", "total_cache"},
3150         {"rss", "total_rss"},
3151         {"mapped_file", "total_mapped_file"},
3152         {"pgpgin", "total_pgpgin"},
3153         {"pgpgout", "total_pgpgout"},
3154         {"swap", "total_swap"},
3155         {"inactive_anon", "total_inactive_anon"},
3156         {"active_anon", "total_active_anon"},
3157         {"inactive_file", "total_inactive_file"},
3158         {"active_file", "total_active_file"},
3159         {"unevictable", "total_unevictable"}
3160 };
3161
3162
3163 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3164 {
3165         struct mcs_total_stat *s = data;
3166         s64 val;
3167
3168         /* per cpu stat */
3169         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3170         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3171         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3172         s->stat[MCS_RSS] += val * PAGE_SIZE;
3173         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3174         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3175         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3176         s->stat[MCS_PGPGIN] += val;
3177         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3178         s->stat[MCS_PGPGOUT] += val;
3179         if (do_swap_account) {
3180                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3181                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3182         }
3183
3184         /* per zone stat */
3185         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3186         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3187         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3188         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3189         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3190         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3191         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3192         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3193         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3194         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3195         return 0;
3196 }
3197
3198 static void
3199 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3200 {
3201         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3202 }
3203
3204 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3205                                  struct cgroup_map_cb *cb)
3206 {
3207         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3208         struct mcs_total_stat mystat;
3209         int i;
3210
3211         memset(&mystat, 0, sizeof(mystat));
3212         mem_cgroup_get_local_stat(mem_cont, &mystat);
3213
3214         for (i = 0; i < NR_MCS_STAT; i++) {
3215                 if (i == MCS_SWAP && !do_swap_account)
3216                         continue;
3217                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3218         }
3219
3220         /* Hierarchical information */
3221         {
3222                 unsigned long long limit, memsw_limit;
3223                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3224                 cb->fill(cb, "hierarchical_memory_limit", limit);
3225                 if (do_swap_account)
3226                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3227         }
3228
3229         memset(&mystat, 0, sizeof(mystat));
3230         mem_cgroup_get_total_stat(mem_cont, &mystat);
3231         for (i = 0; i < NR_MCS_STAT; i++) {
3232                 if (i == MCS_SWAP && !do_swap_account)
3233                         continue;
3234                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3235         }
3236
3237 #ifdef CONFIG_DEBUG_VM
3238         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3239
3240         {
3241                 int nid, zid;
3242                 struct mem_cgroup_per_zone *mz;
3243                 unsigned long recent_rotated[2] = {0, 0};
3244                 unsigned long recent_scanned[2] = {0, 0};
3245
3246                 for_each_online_node(nid)
3247                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3248                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3249
3250                                 recent_rotated[0] +=
3251                                         mz->reclaim_stat.recent_rotated[0];
3252                                 recent_rotated[1] +=
3253                                         mz->reclaim_stat.recent_rotated[1];
3254                                 recent_scanned[0] +=
3255                                         mz->reclaim_stat.recent_scanned[0];
3256                                 recent_scanned[1] +=
3257                                         mz->reclaim_stat.recent_scanned[1];
3258                         }
3259                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3260                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3261                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3262                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3263         }
3264 #endif
3265
3266         return 0;
3267 }
3268
3269 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3270 {
3271         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3272
3273         return get_swappiness(memcg);
3274 }
3275
3276 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3277                                        u64 val)
3278 {
3279         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3280         struct mem_cgroup *parent;
3281
3282         if (val > 100)
3283                 return -EINVAL;
3284
3285         if (cgrp->parent == NULL)
3286                 return -EINVAL;
3287
3288         parent = mem_cgroup_from_cont(cgrp->parent);
3289
3290         cgroup_lock();
3291
3292         /* If under hierarchy, only empty-root can set this value */
3293         if ((parent->use_hierarchy) ||
3294             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3295                 cgroup_unlock();
3296                 return -EINVAL;
3297         }
3298
3299         spin_lock(&memcg->reclaim_param_lock);
3300         memcg->swappiness = val;
3301         spin_unlock(&memcg->reclaim_param_lock);
3302
3303         cgroup_unlock();
3304
3305         return 0;
3306 }
3307
3308 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3309 {
3310         struct mem_cgroup_threshold_ary *t;
3311         u64 usage;
3312         int i;
3313
3314         rcu_read_lock();
3315         if (!swap)
3316                 t = rcu_dereference(memcg->thresholds);
3317         else
3318                 t = rcu_dereference(memcg->memsw_thresholds);
3319
3320         if (!t)
3321                 goto unlock;
3322
3323         usage = mem_cgroup_usage(memcg, swap);
3324
3325         /*
3326          * current_threshold points to threshold just below usage.
3327          * If it's not true, a threshold was crossed after last
3328          * call of __mem_cgroup_threshold().
3329          */
3330         i = atomic_read(&t->current_threshold);
3331
3332         /*
3333          * Iterate backward over array of thresholds starting from
3334          * current_threshold and check if a threshold is crossed.
3335          * If none of thresholds below usage is crossed, we read
3336          * only one element of the array here.
3337          */
3338         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3339                 eventfd_signal(t->entries[i].eventfd, 1);
3340
3341         /* i = current_threshold + 1 */
3342         i++;
3343
3344         /*
3345          * Iterate forward over array of thresholds starting from
3346          * current_threshold+1 and check if a threshold is crossed.
3347          * If none of thresholds above usage is crossed, we read
3348          * only one element of the array here.
3349          */
3350         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3351                 eventfd_signal(t->entries[i].eventfd, 1);
3352
3353         /* Update current_threshold */
3354         atomic_set(&t->current_threshold, i - 1);
3355 unlock:
3356         rcu_read_unlock();
3357 }
3358
3359 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3360 {
3361         __mem_cgroup_threshold(memcg, false);
3362         if (do_swap_account)
3363                 __mem_cgroup_threshold(memcg, true);
3364 }
3365
3366 static int compare_thresholds(const void *a, const void *b)
3367 {
3368         const struct mem_cgroup_threshold *_a = a;
3369         const struct mem_cgroup_threshold *_b = b;
3370
3371         return _a->threshold - _b->threshold;
3372 }
3373
3374 static int mem_cgroup_register_event(struct cgroup *cgrp, struct cftype *cft,
3375                 struct eventfd_ctx *eventfd, const char *args)
3376 {
3377         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3378         struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3379         int type = MEMFILE_TYPE(cft->private);
3380         u64 threshold, usage;
3381         int size;
3382         int i, ret;
3383
3384         ret = res_counter_memparse_write_strategy(args, &threshold);
3385         if (ret)
3386                 return ret;
3387
3388         mutex_lock(&memcg->thresholds_lock);
3389         if (type == _MEM)
3390                 thresholds = memcg->thresholds;
3391         else if (type == _MEMSWAP)
3392                 thresholds = memcg->memsw_thresholds;
3393         else
3394                 BUG();
3395
3396         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3397
3398         /* Check if a threshold crossed before adding a new one */
3399         if (thresholds)
3400                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3401
3402         if (thresholds)
3403                 size = thresholds->size + 1;
3404         else
3405                 size = 1;
3406
3407         /* Allocate memory for new array of thresholds */
3408         thresholds_new = kmalloc(sizeof(*thresholds_new) +
3409                         size * sizeof(struct mem_cgroup_threshold),
3410                         GFP_KERNEL);
3411         if (!thresholds_new) {
3412                 ret = -ENOMEM;
3413                 goto unlock;
3414         }
3415         thresholds_new->size = size;
3416
3417         /* Copy thresholds (if any) to new array */
3418         if (thresholds)
3419                 memcpy(thresholds_new->entries, thresholds->entries,
3420                                 thresholds->size *
3421                                 sizeof(struct mem_cgroup_threshold));
3422         /* Add new threshold */
3423         thresholds_new->entries[size - 1].eventfd = eventfd;
3424         thresholds_new->entries[size - 1].threshold = threshold;
3425
3426         /* Sort thresholds. Registering of new threshold isn't time-critical */
3427         sort(thresholds_new->entries, size,
3428                         sizeof(struct mem_cgroup_threshold),
3429                         compare_thresholds, NULL);
3430
3431         /* Find current threshold */
3432         atomic_set(&thresholds_new->current_threshold, -1);
3433         for (i = 0; i < size; i++) {
3434                 if (thresholds_new->entries[i].threshold < usage) {
3435                         /*
3436                          * thresholds_new->current_threshold will not be used
3437                          * until rcu_assign_pointer(), so it's safe to increment
3438                          * it here.
3439                          */
3440                         atomic_inc(&thresholds_new->current_threshold);
3441                 }
3442         }
3443
3444         if (type == _MEM)
3445                 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3446         else
3447                 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3448
3449         /* To be sure that nobody uses thresholds before freeing it */
3450         synchronize_rcu();
3451
3452         kfree(thresholds);
3453 unlock:
3454         mutex_unlock(&memcg->thresholds_lock);
3455
3456         return ret;
3457 }
3458
3459 static int mem_cgroup_unregister_event(struct cgroup *cgrp, struct cftype *cft,
3460                 struct eventfd_ctx *eventfd)
3461 {
3462         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3463         struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3464         int type = MEMFILE_TYPE(cft->private);
3465         u64 usage;
3466         int size = 0;
3467         int i, j, ret;
3468
3469         mutex_lock(&memcg->thresholds_lock);
3470         if (type == _MEM)
3471                 thresholds = memcg->thresholds;
3472         else if (type == _MEMSWAP)
3473                 thresholds = memcg->memsw_thresholds;
3474         else
3475                 BUG();
3476
3477         /*
3478          * Something went wrong if we trying to unregister a threshold
3479          * if we don't have thresholds
3480          */
3481         BUG_ON(!thresholds);
3482
3483         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3484
3485         /* Check if a threshold crossed before removing */
3486         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3487
3488         /* Calculate new number of threshold */
3489         for (i = 0; i < thresholds->size; i++) {
3490                 if (thresholds->entries[i].eventfd != eventfd)
3491                         size++;
3492         }
3493
3494         /* Set thresholds array to NULL if we don't have thresholds */
3495         if (!size) {
3496                 thresholds_new = NULL;
3497                 goto assign;
3498         }
3499
3500         /* Allocate memory for new array of thresholds */
3501         thresholds_new = kmalloc(sizeof(*thresholds_new) +
3502                         size * sizeof(struct mem_cgroup_threshold),
3503                         GFP_KERNEL);
3504         if (!thresholds_new) {
3505                 ret = -ENOMEM;
3506                 goto unlock;
3507         }
3508         thresholds_new->size = size;
3509
3510         /* Copy thresholds and find current threshold */
3511         atomic_set(&thresholds_new->current_threshold, -1);
3512         for (i = 0, j = 0; i < thresholds->size; i++) {
3513                 if (thresholds->entries[i].eventfd == eventfd)
3514                         continue;
3515
3516                 thresholds_new->entries[j] = thresholds->entries[i];
3517                 if (thresholds_new->entries[j].threshold < usage) {
3518                         /*
3519                          * thresholds_new->current_threshold will not be used
3520                          * until rcu_assign_pointer(), so it's safe to increment
3521                          * it here.
3522                          */
3523                         atomic_inc(&thresholds_new->current_threshold);
3524                 }
3525                 j++;
3526         }
3527
3528 assign:
3529         if (type == _MEM)
3530                 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3531         else
3532                 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3533
3534         /* To be sure that nobody uses thresholds before freeing it */
3535         synchronize_rcu();
3536
3537         kfree(thresholds);
3538 unlock:
3539         mutex_unlock(&memcg->thresholds_lock);
3540
3541         return ret;
3542 }
3543
3544 static struct cftype mem_cgroup_files[] = {
3545         {
3546                 .name = "usage_in_bytes",
3547                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3548                 .read_u64 = mem_cgroup_read,
3549                 .register_event = mem_cgroup_register_event,
3550                 .unregister_event = mem_cgroup_unregister_event,
3551         },
3552         {
3553                 .name = "max_usage_in_bytes",
3554                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3555                 .trigger = mem_cgroup_reset,
3556                 .read_u64 = mem_cgroup_read,
3557         },
3558         {
3559                 .name = "limit_in_bytes",
3560                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3561                 .write_string = mem_cgroup_write,
3562                 .read_u64 = mem_cgroup_read,
3563         },
3564         {
3565                 .name = "soft_limit_in_bytes",
3566                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3567                 .write_string = mem_cgroup_write,
3568                 .read_u64 = mem_cgroup_read,
3569         },
3570         {
3571                 .name = "failcnt",
3572                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3573                 .trigger = mem_cgroup_reset,
3574                 .read_u64 = mem_cgroup_read,
3575         },
3576         {
3577                 .name = "stat",
3578                 .read_map = mem_control_stat_show,
3579         },
3580         {
3581                 .name = "force_empty",
3582                 .trigger = mem_cgroup_force_empty_write,
3583         },
3584         {
3585                 .name = "use_hierarchy",
3586                 .write_u64 = mem_cgroup_hierarchy_write,
3587                 .read_u64 = mem_cgroup_hierarchy_read,
3588         },
3589         {
3590                 .name = "swappiness",
3591                 .read_u64 = mem_cgroup_swappiness_read,
3592                 .write_u64 = mem_cgroup_swappiness_write,
3593         },
3594         {
3595                 .name = "move_charge_at_immigrate",
3596                 .read_u64 = mem_cgroup_move_charge_read,
3597                 .write_u64 = mem_cgroup_move_charge_write,
3598         },
3599 };
3600
3601 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3602 static struct cftype memsw_cgroup_files[] = {
3603         {
3604                 .name = "memsw.usage_in_bytes",
3605                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3606                 .read_u64 = mem_cgroup_read,
3607                 .register_event = mem_cgroup_register_event,
3608                 .unregister_event = mem_cgroup_unregister_event,
3609         },
3610         {
3611                 .name = "memsw.max_usage_in_bytes",
3612                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3613                 .trigger = mem_cgroup_reset,
3614                 .read_u64 = mem_cgroup_read,
3615         },
3616         {
3617                 .name = "memsw.limit_in_bytes",
3618                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3619                 .write_string = mem_cgroup_write,
3620                 .read_u64 = mem_cgroup_read,
3621         },
3622         {
3623                 .name = "memsw.failcnt",
3624                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3625                 .trigger = mem_cgroup_reset,
3626                 .read_u64 = mem_cgroup_read,
3627         },
3628 };
3629
3630 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3631 {
3632         if (!do_swap_account)
3633                 return 0;
3634         return cgroup_add_files(cont, ss, memsw_cgroup_files,
3635                                 ARRAY_SIZE(memsw_cgroup_files));
3636 };
3637 #else
3638 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3639 {
3640         return 0;
3641 }
3642 #endif
3643
3644 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3645 {
3646         struct mem_cgroup_per_node *pn;
3647         struct mem_cgroup_per_zone *mz;
3648         enum lru_list l;
3649         int zone, tmp = node;
3650         /*
3651          * This routine is called against possible nodes.
3652          * But it's BUG to call kmalloc() against offline node.
3653          *
3654          * TODO: this routine can waste much memory for nodes which will
3655          *       never be onlined. It's better to use memory hotplug callback
3656          *       function.
3657          */
3658         if (!node_state(node, N_NORMAL_MEMORY))
3659                 tmp = -1;
3660         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3661         if (!pn)
3662                 return 1;
3663
3664         mem->info.nodeinfo[node] = pn;
3665         memset(pn, 0, sizeof(*pn));
3666
3667         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3668                 mz = &pn->zoneinfo[zone];
3669                 for_each_lru(l)
3670                         INIT_LIST_HEAD(&mz->lists[l]);
3671                 mz->usage_in_excess = 0;
3672                 mz->on_tree = false;
3673                 mz->mem = mem;
3674         }
3675         return 0;
3676 }
3677
3678 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3679 {
3680         kfree(mem->info.nodeinfo[node]);
3681 }
3682
3683 static struct mem_cgroup *mem_cgroup_alloc(void)
3684 {
3685         struct mem_cgroup *mem;
3686         int size = sizeof(struct mem_cgroup);
3687
3688         /* Can be very big if MAX_NUMNODES is very big */
3689         if (size < PAGE_SIZE)
3690                 mem = kmalloc(size, GFP_KERNEL);
3691         else
3692                 mem = vmalloc(size);
3693
3694         if (!mem)
3695                 return NULL;
3696
3697         memset(mem, 0, size);
3698         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3699         if (!mem->stat) {
3700                 if (size < PAGE_SIZE)
3701                         kfree(mem);
3702                 else
3703                         vfree(mem);
3704                 mem = NULL;
3705         }
3706         return mem;
3707 }
3708
3709 /*
3710  * At destroying mem_cgroup, references from swap_cgroup can remain.
3711  * (scanning all at force_empty is too costly...)
3712  *
3713  * Instead of clearing all references at force_empty, we remember
3714  * the number of reference from swap_cgroup and free mem_cgroup when
3715  * it goes down to 0.
3716  *
3717  * Removal of cgroup itself succeeds regardless of refs from swap.
3718  */
3719
3720 static void __mem_cgroup_free(struct mem_cgroup *mem)
3721 {
3722         int node;
3723
3724         mem_cgroup_remove_from_trees(mem);
3725         free_css_id(&mem_cgroup_subsys, &mem->css);
3726
3727         for_each_node_state(node, N_POSSIBLE)
3728                 free_mem_cgroup_per_zone_info(mem, node);
3729
3730         free_percpu(mem->stat);
3731         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
3732                 kfree(mem);
3733         else
3734                 vfree(mem);
3735 }
3736
3737 static void mem_cgroup_get(struct mem_cgroup *mem)
3738 {
3739         atomic_inc(&mem->refcnt);
3740 }
3741
3742 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
3743 {
3744         if (atomic_sub_and_test(count, &mem->refcnt)) {
3745                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3746                 __mem_cgroup_free(mem);
3747                 if (parent)
3748                         mem_cgroup_put(parent);
3749         }
3750 }
3751
3752 static void mem_cgroup_put(struct mem_cgroup *mem)
3753 {
3754         __mem_cgroup_put(mem, 1);
3755 }
3756
3757 /*
3758  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3759  */
3760 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3761 {
3762         if (!mem->res.parent)
3763                 return NULL;
3764         return mem_cgroup_from_res_counter(mem->res.parent, res);
3765 }
3766
3767 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3768 static void __init enable_swap_cgroup(void)
3769 {
3770         if (!mem_cgroup_disabled() && really_do_swap_account)
3771                 do_swap_account = 1;
3772 }
3773 #else
3774 static void __init enable_swap_cgroup(void)
3775 {
3776 }
3777 #endif
3778
3779 static int mem_cgroup_soft_limit_tree_init(void)
3780 {
3781         struct mem_cgroup_tree_per_node *rtpn;
3782         struct mem_cgroup_tree_per_zone *rtpz;
3783         int tmp, node, zone;
3784
3785         for_each_node_state(node, N_POSSIBLE) {
3786                 tmp = node;
3787                 if (!node_state(node, N_NORMAL_MEMORY))
3788                         tmp = -1;
3789                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3790                 if (!rtpn)
3791                         return 1;
3792
3793                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3794
3795                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3796                         rtpz = &rtpn->rb_tree_per_zone[zone];
3797                         rtpz->rb_root = RB_ROOT;
3798                         spin_lock_init(&rtpz->lock);
3799                 }
3800         }
3801         return 0;
3802 }
3803
3804 static struct cgroup_subsys_state * __ref
3805 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3806 {
3807         struct mem_cgroup *mem, *parent;
3808         long error = -ENOMEM;
3809         int node;
3810
3811         mem = mem_cgroup_alloc();
3812         if (!mem)
3813                 return ERR_PTR(error);
3814
3815         for_each_node_state(node, N_POSSIBLE)
3816                 if (alloc_mem_cgroup_per_zone_info(mem, node))
3817                         goto free_out;
3818
3819         /* root ? */
3820         if (cont->parent == NULL) {
3821                 int cpu;
3822                 enable_swap_cgroup();
3823                 parent = NULL;
3824                 root_mem_cgroup = mem;
3825                 if (mem_cgroup_soft_limit_tree_init())
3826                         goto free_out;
3827                 for_each_possible_cpu(cpu) {
3828                         struct memcg_stock_pcp *stock =
3829                                                 &per_cpu(memcg_stock, cpu);
3830                         INIT_WORK(&stock->work, drain_local_stock);
3831                 }
3832                 hotcpu_notifier(memcg_stock_cpu_callback, 0);
3833         } else {
3834                 parent = mem_cgroup_from_cont(cont->parent);
3835                 mem->use_hierarchy = parent->use_hierarchy;
3836         }
3837
3838         if (parent && parent->use_hierarchy) {
3839                 res_counter_init(&mem->res, &parent->res);
3840                 res_counter_init(&mem->memsw, &parent->memsw);
3841                 /*
3842                  * We increment refcnt of the parent to ensure that we can
3843                  * safely access it on res_counter_charge/uncharge.
3844                  * This refcnt will be decremented when freeing this
3845                  * mem_cgroup(see mem_cgroup_put).
3846                  */
3847                 mem_cgroup_get(parent);
3848         } else {
3849                 res_counter_init(&mem->res, NULL);
3850                 res_counter_init(&mem->memsw, NULL);
3851         }
3852         mem->last_scanned_child = 0;
3853         spin_lock_init(&mem->reclaim_param_lock);
3854
3855         if (parent)
3856                 mem->swappiness = get_swappiness(parent);
3857         atomic_set(&mem->refcnt, 1);
3858         mem->move_charge_at_immigrate = 0;
3859         mutex_init(&mem->thresholds_lock);
3860         return &mem->css;
3861 free_out:
3862         __mem_cgroup_free(mem);
3863         root_mem_cgroup = NULL;
3864         return ERR_PTR(error);
3865 }
3866
3867 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
3868                                         struct cgroup *cont)
3869 {
3870         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3871
3872         return mem_cgroup_force_empty(mem, false);
3873 }
3874
3875 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
3876                                 struct cgroup *cont)
3877 {
3878         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3879
3880         mem_cgroup_put(mem);
3881 }
3882
3883 static int mem_cgroup_populate(struct cgroup_subsys *ss,
3884                                 struct cgroup *cont)
3885 {
3886         int ret;
3887
3888         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
3889                                 ARRAY_SIZE(mem_cgroup_files));
3890
3891         if (!ret)
3892                 ret = register_memsw_files(cont, ss);
3893         return ret;
3894 }
3895
3896 #ifdef CONFIG_MMU
3897 /* Handlers for move charge at task migration. */
3898 #define PRECHARGE_COUNT_AT_ONCE 256
3899 static int mem_cgroup_do_precharge(unsigned long count)
3900 {
3901         int ret = 0;
3902         int batch_count = PRECHARGE_COUNT_AT_ONCE;
3903         struct mem_cgroup *mem = mc.to;
3904
3905         if (mem_cgroup_is_root(mem)) {
3906                 mc.precharge += count;
3907                 /* we don't need css_get for root */
3908                 return ret;
3909         }
3910         /* try to charge at once */
3911         if (count > 1) {
3912                 struct res_counter *dummy;
3913                 /*
3914                  * "mem" cannot be under rmdir() because we've already checked
3915                  * by cgroup_lock_live_cgroup() that it is not removed and we
3916                  * are still under the same cgroup_mutex. So we can postpone
3917                  * css_get().
3918                  */
3919                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
3920                         goto one_by_one;
3921                 if (do_swap_account && res_counter_charge(&mem->memsw,
3922                                                 PAGE_SIZE * count, &dummy)) {
3923                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
3924                         goto one_by_one;
3925                 }
3926                 mc.precharge += count;
3927                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
3928                 WARN_ON_ONCE(count > INT_MAX);
3929                 __css_get(&mem->css, (int)count);
3930                 return ret;
3931         }
3932 one_by_one:
3933         /* fall back to one by one charge */
3934         while (count--) {
3935                 if (signal_pending(current)) {
3936                         ret = -EINTR;
3937                         break;
3938                 }
3939                 if (!batch_count--) {
3940                         batch_count = PRECHARGE_COUNT_AT_ONCE;
3941                         cond_resched();
3942                 }
3943                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
3944                 if (ret || !mem)
3945                         /* mem_cgroup_clear_mc() will do uncharge later */
3946                         return -ENOMEM;
3947                 mc.precharge++;
3948         }
3949         return ret;
3950 }
3951
3952 /**
3953  * is_target_pte_for_mc - check a pte whether it is valid for move charge
3954  * @vma: the vma the pte to be checked belongs
3955  * @addr: the address corresponding to the pte to be checked
3956  * @ptent: the pte to be checked
3957  * @target: the pointer the target page or swap ent will be stored(can be NULL)
3958  *
3959  * Returns
3960  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
3961  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
3962  *     move charge. if @target is not NULL, the page is stored in target->page
3963  *     with extra refcnt got(Callers should handle it).
3964  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
3965  *     target for charge migration. if @target is not NULL, the entry is stored
3966  *     in target->ent.
3967  *
3968  * Called with pte lock held.
3969  */
3970 union mc_target {
3971         struct page     *page;
3972         swp_entry_t     ent;
3973 };
3974
3975 enum mc_target_type {
3976         MC_TARGET_NONE, /* not used */
3977         MC_TARGET_PAGE,
3978         MC_TARGET_SWAP,
3979 };
3980
3981 static int is_target_pte_for_mc(struct vm_area_struct *vma,
3982                 unsigned long addr, pte_t ptent, union mc_target *target)
3983 {
3984         struct page *page = NULL;
3985         struct page_cgroup *pc;
3986         int ret = 0;
3987         swp_entry_t ent = { .val = 0 };
3988         int usage_count = 0;
3989         bool move_anon = test_bit(MOVE_CHARGE_TYPE_ANON,
3990                                         &mc.to->move_charge_at_immigrate);
3991
3992         if (!pte_present(ptent)) {
3993                 /* TODO: handle swap of shmes/tmpfs */
3994                 if (pte_none(ptent) || pte_file(ptent))
3995                         return 0;
3996                 else if (is_swap_pte(ptent)) {
3997                         ent = pte_to_swp_entry(ptent);
3998                         if (!move_anon || non_swap_entry(ent))
3999                                 return 0;
4000                         usage_count = mem_cgroup_count_swap_user(ent, &page);
4001                 }
4002         } else {
4003                 page = vm_normal_page(vma, addr, ptent);
4004                 if (!page || !page_mapped(page))
4005                         return 0;
4006                 /*
4007                  * TODO: We don't move charges of file(including shmem/tmpfs)
4008                  * pages for now.
4009                  */
4010                 if (!move_anon || !PageAnon(page))
4011                         return 0;
4012                 if (!get_page_unless_zero(page))
4013                         return 0;
4014                 usage_count = page_mapcount(page);
4015         }
4016         if (usage_count > 1) {
4017                 /*
4018                  * TODO: We don't move charges of shared(used by multiple
4019                  * processes) pages for now.
4020                  */
4021                 if (page)
4022                         put_page(page);
4023                 return 0;
4024         }
4025         if (page) {
4026                 pc = lookup_page_cgroup(page);
4027                 /*
4028                  * Do only loose check w/o page_cgroup lock.
4029                  * mem_cgroup_move_account() checks the pc is valid or not under
4030                  * the lock.
4031                  */
4032                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4033                         ret = MC_TARGET_PAGE;
4034                         if (target)
4035                                 target->page = page;
4036                 }
4037                 if (!ret || !target)
4038                         put_page(page);
4039         }
4040         /* throught */
4041         if (ent.val && do_swap_account && !ret &&
4042                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4043                 ret = MC_TARGET_SWAP;
4044                 if (target)
4045                         target->ent = ent;
4046         }
4047         return ret;
4048 }
4049
4050 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4051                                         unsigned long addr, unsigned long end,
4052                                         struct mm_walk *walk)
4053 {
4054         struct vm_area_struct *vma = walk->private;
4055         pte_t *pte;
4056         spinlock_t *ptl;
4057
4058         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4059         for (; addr != end; pte++, addr += PAGE_SIZE)
4060                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4061                         mc.precharge++; /* increment precharge temporarily */
4062         pte_unmap_unlock(pte - 1, ptl);
4063         cond_resched();
4064
4065         return 0;
4066 }
4067
4068 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4069 {
4070         unsigned long precharge;
4071         struct vm_area_struct *vma;
4072
4073         down_read(&mm->mmap_sem);
4074         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4075                 struct mm_walk mem_cgroup_count_precharge_walk = {
4076                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4077                         .mm = mm,
4078                         .private = vma,
4079                 };
4080                 if (is_vm_hugetlb_page(vma))
4081                         continue;
4082                 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4083                 if (vma->vm_flags & VM_SHARED)
4084                         continue;
4085                 walk_page_range(vma->vm_start, vma->vm_end,
4086                                         &mem_cgroup_count_precharge_walk);
4087         }
4088         up_read(&mm->mmap_sem);
4089
4090         precharge = mc.precharge;
4091         mc.precharge = 0;
4092
4093         return precharge;
4094 }
4095
4096 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4097 {
4098         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4099 }
4100
4101 static void mem_cgroup_clear_mc(void)
4102 {
4103         /* we must uncharge all the leftover precharges from mc.to */
4104         if (mc.precharge) {
4105                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4106                 mc.precharge = 0;
4107         }
4108         /*
4109          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4110          * we must uncharge here.
4111          */
4112         if (mc.moved_charge) {
4113                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4114                 mc.moved_charge = 0;
4115         }
4116         /* we must fixup refcnts and charges */
4117         if (mc.moved_swap) {
4118                 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4119                 /* uncharge swap account from the old cgroup */
4120                 if (!mem_cgroup_is_root(mc.from))
4121                         res_counter_uncharge(&mc.from->memsw,
4122                                                 PAGE_SIZE * mc.moved_swap);
4123                 __mem_cgroup_put(mc.from, mc.moved_swap);
4124
4125                 if (!mem_cgroup_is_root(mc.to)) {
4126                         /*
4127                          * we charged both to->res and to->memsw, so we should
4128                          * uncharge to->res.
4129                          */
4130                         res_counter_uncharge(&mc.to->res,
4131                                                 PAGE_SIZE * mc.moved_swap);
4132                         VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4133                         __css_put(&mc.to->css, mc.moved_swap);
4134                 }
4135                 /* we've already done mem_cgroup_get(mc.to) */
4136
4137                 mc.moved_swap = 0;
4138         }
4139         mc.from = NULL;
4140         mc.to = NULL;
4141         mc.moving_task = NULL;
4142         wake_up_all(&mc.waitq);
4143 }
4144
4145 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4146                                 struct cgroup *cgroup,
4147                                 struct task_struct *p,
4148                                 bool threadgroup)
4149 {
4150         int ret = 0;
4151         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4152
4153         if (mem->move_charge_at_immigrate) {
4154                 struct mm_struct *mm;
4155                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4156
4157                 VM_BUG_ON(from == mem);
4158
4159                 mm = get_task_mm(p);
4160                 if (!mm)
4161                         return 0;
4162                 /* We move charges only when we move a owner of the mm */
4163                 if (mm->owner == p) {
4164                         VM_BUG_ON(mc.from);
4165                         VM_BUG_ON(mc.to);
4166                         VM_BUG_ON(mc.precharge);
4167                         VM_BUG_ON(mc.moved_charge);
4168                         VM_BUG_ON(mc.moved_swap);
4169                         VM_BUG_ON(mc.moving_task);
4170                         mc.from = from;
4171                         mc.to = mem;
4172                         mc.precharge = 0;
4173                         mc.moved_charge = 0;
4174                         mc.moved_swap = 0;
4175                         mc.moving_task = current;
4176
4177                         ret = mem_cgroup_precharge_mc(mm);
4178                         if (ret)
4179                                 mem_cgroup_clear_mc();
4180                 }
4181                 mmput(mm);
4182         }
4183         return ret;
4184 }
4185
4186 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4187                                 struct cgroup *cgroup,
4188                                 struct task_struct *p,
4189                                 bool threadgroup)
4190 {
4191         mem_cgroup_clear_mc();
4192 }
4193
4194 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4195                                 unsigned long addr, unsigned long end,
4196                                 struct mm_walk *walk)
4197 {
4198         int ret = 0;
4199         struct vm_area_struct *vma = walk->private;
4200         pte_t *pte;
4201         spinlock_t *ptl;
4202
4203 retry:
4204         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4205         for (; addr != end; addr += PAGE_SIZE) {
4206                 pte_t ptent = *(pte++);
4207                 union mc_target target;
4208                 int type;
4209                 struct page *page;
4210                 struct page_cgroup *pc;
4211                 swp_entry_t ent;
4212
4213                 if (!mc.precharge)
4214                         break;
4215
4216                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4217                 switch (type) {
4218                 case MC_TARGET_PAGE:
4219                         page = target.page;
4220                         if (isolate_lru_page(page))
4221                                 goto put;
4222                         pc = lookup_page_cgroup(page);
4223                         if (!mem_cgroup_move_account(pc,
4224                                                 mc.from, mc.to, false)) {
4225                                 mc.precharge--;
4226                                 /* we uncharge from mc.from later. */
4227                                 mc.moved_charge++;
4228                         }
4229                         putback_lru_page(page);
4230 put:                    /* is_target_pte_for_mc() gets the page */
4231                         put_page(page);
4232                         break;
4233                 case MC_TARGET_SWAP:
4234                         ent = target.ent;
4235                         if (!mem_cgroup_move_swap_account(ent,
4236                                                 mc.from, mc.to, false)) {
4237                                 mc.precharge--;
4238                                 /* we fixup refcnts and charges later. */
4239                                 mc.moved_swap++;
4240                         }
4241                         break;
4242                 default:
4243                         break;
4244                 }
4245         }
4246         pte_unmap_unlock(pte - 1, ptl);
4247         cond_resched();
4248
4249         if (addr != end) {
4250                 /*
4251                  * We have consumed all precharges we got in can_attach().
4252                  * We try charge one by one, but don't do any additional
4253                  * charges to mc.to if we have failed in charge once in attach()
4254                  * phase.
4255                  */
4256                 ret = mem_cgroup_do_precharge(1);
4257                 if (!ret)
4258                         goto retry;
4259         }
4260
4261         return ret;
4262 }
4263
4264 static void mem_cgroup_move_charge(struct mm_struct *mm)
4265 {
4266         struct vm_area_struct *vma;
4267
4268         lru_add_drain_all();
4269         down_read(&mm->mmap_sem);
4270         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4271                 int ret;
4272                 struct mm_walk mem_cgroup_move_charge_walk = {
4273                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4274                         .mm = mm,
4275                         .private = vma,
4276                 };
4277                 if (is_vm_hugetlb_page(vma))
4278                         continue;
4279                 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4280                 if (vma->vm_flags & VM_SHARED)
4281                         continue;
4282                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4283                                                 &mem_cgroup_move_charge_walk);
4284                 if (ret)
4285                         /*
4286                          * means we have consumed all precharges and failed in
4287                          * doing additional charge. Just abandon here.
4288                          */
4289                         break;
4290         }
4291         up_read(&mm->mmap_sem);
4292 }
4293
4294 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4295                                 struct cgroup *cont,
4296                                 struct cgroup *old_cont,
4297                                 struct task_struct *p,
4298                                 bool threadgroup)
4299 {
4300         struct mm_struct *mm;
4301
4302         if (!mc.to)
4303                 /* no need to move charge */
4304                 return;
4305
4306         mm = get_task_mm(p);
4307         if (mm) {
4308                 mem_cgroup_move_charge(mm);
4309                 mmput(mm);
4310         }
4311         mem_cgroup_clear_mc();
4312 }
4313 #else   /* !CONFIG_MMU */
4314 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4315                                 struct cgroup *cgroup,
4316                                 struct task_struct *p,
4317                                 bool threadgroup)
4318 {
4319         return 0;
4320 }
4321 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4322                                 struct cgroup *cgroup,
4323                                 struct task_struct *p,
4324                                 bool threadgroup)
4325 {
4326 }
4327 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4328                                 struct cgroup *cont,
4329                                 struct cgroup *old_cont,
4330                                 struct task_struct *p,
4331                                 bool threadgroup)
4332 {
4333 }
4334 #endif
4335
4336 struct cgroup_subsys mem_cgroup_subsys = {
4337         .name = "memory",
4338         .subsys_id = mem_cgroup_subsys_id,
4339         .create = mem_cgroup_create,
4340         .pre_destroy = mem_cgroup_pre_destroy,
4341         .destroy = mem_cgroup_destroy,
4342         .populate = mem_cgroup_populate,
4343         .can_attach = mem_cgroup_can_attach,
4344         .cancel_attach = mem_cgroup_cancel_attach,
4345         .attach = mem_cgroup_move_task,
4346         .early_init = 0,
4347         .use_id = 1,
4348 };
4349
4350 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4351
4352 static int __init disable_swap_account(char *s)
4353 {
4354         really_do_swap_account = 0;
4355         return 1;
4356 }
4357 __setup("noswapaccount", disable_swap_account);
4358 #endif