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