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