mm/memcontrol.c: fix "integer as NULL pointer" sparse warning
[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  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  */
19
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/limits.h>
31 #include <linux/mutex.h>
32 #include <linux/rbtree.h>
33 #include <linux/slab.h>
34 #include <linux/swap.h>
35 #include <linux/spinlock.h>
36 #include <linux/fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mm_inline.h>
40 #include <linux/page_cgroup.h>
41 #include <linux/cpu.h>
42 #include "internal.h"
43
44 #include <asm/uaccess.h>
45
46 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
47 #define MEM_CGROUP_RECLAIM_RETRIES      5
48 struct mem_cgroup *root_mem_cgroup __read_mostly;
49
50 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
51 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
52 int do_swap_account __read_mostly;
53 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
54 #else
55 #define do_swap_account         (0)
56 #endif
57
58 #define SOFTLIMIT_EVENTS_THRESH (1000)
59
60 /*
61  * Statistics for memory cgroup.
62  */
63 enum mem_cgroup_stat_index {
64         /*
65          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
66          */
67         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
68         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
69         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
70         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
71         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
72         MEM_CGROUP_STAT_EVENTS, /* sum of pagein + pageout for internal use */
73         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
74
75         MEM_CGROUP_STAT_NSTATS,
76 };
77
78 struct mem_cgroup_stat_cpu {
79         s64 count[MEM_CGROUP_STAT_NSTATS];
80 } ____cacheline_aligned_in_smp;
81
82 struct mem_cgroup_stat {
83         struct mem_cgroup_stat_cpu cpustat[0];
84 };
85
86 static inline void
87 __mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
88                                 enum mem_cgroup_stat_index idx)
89 {
90         stat->count[idx] = 0;
91 }
92
93 static inline s64
94 __mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
95                                 enum mem_cgroup_stat_index idx)
96 {
97         return stat->count[idx];
98 }
99
100 /*
101  * For accounting under irq disable, no need for increment preempt count.
102  */
103 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
104                 enum mem_cgroup_stat_index idx, int val)
105 {
106         stat->count[idx] += val;
107 }
108
109 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
110                 enum mem_cgroup_stat_index idx)
111 {
112         int cpu;
113         s64 ret = 0;
114         for_each_possible_cpu(cpu)
115                 ret += stat->cpustat[cpu].count[idx];
116         return ret;
117 }
118
119 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
120 {
121         s64 ret;
122
123         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
124         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
125         return ret;
126 }
127
128 /*
129  * per-zone information in memory controller.
130  */
131 struct mem_cgroup_per_zone {
132         /*
133          * spin_lock to protect the per cgroup LRU
134          */
135         struct list_head        lists[NR_LRU_LISTS];
136         unsigned long           count[NR_LRU_LISTS];
137
138         struct zone_reclaim_stat reclaim_stat;
139         struct rb_node          tree_node;      /* RB tree node */
140         unsigned long long      usage_in_excess;/* Set to the value by which */
141                                                 /* the soft limit is exceeded*/
142         bool                    on_tree;
143         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
144                                                 /* use container_of        */
145 };
146 /* Macro for accessing counter */
147 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
148
149 struct mem_cgroup_per_node {
150         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
151 };
152
153 struct mem_cgroup_lru_info {
154         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
155 };
156
157 /*
158  * Cgroups above their limits are maintained in a RB-Tree, independent of
159  * their hierarchy representation
160  */
161
162 struct mem_cgroup_tree_per_zone {
163         struct rb_root rb_root;
164         spinlock_t lock;
165 };
166
167 struct mem_cgroup_tree_per_node {
168         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
169 };
170
171 struct mem_cgroup_tree {
172         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
173 };
174
175 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176
177 /*
178  * The memory controller data structure. The memory controller controls both
179  * page cache and RSS per cgroup. We would eventually like to provide
180  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
181  * to help the administrator determine what knobs to tune.
182  *
183  * TODO: Add a water mark for the memory controller. Reclaim will begin when
184  * we hit the water mark. May be even add a low water mark, such that
185  * no reclaim occurs from a cgroup at it's low water mark, this is
186  * a feature that will be implemented much later in the future.
187  */
188 struct mem_cgroup {
189         struct cgroup_subsys_state css;
190         /*
191          * the counter to account for memory usage
192          */
193         struct res_counter res;
194         /*
195          * the counter to account for mem+swap usage.
196          */
197         struct res_counter memsw;
198         /*
199          * Per cgroup active and inactive list, similar to the
200          * per zone LRU lists.
201          */
202         struct mem_cgroup_lru_info info;
203
204         /*
205           protect against reclaim related member.
206         */
207         spinlock_t reclaim_param_lock;
208
209         int     prev_priority;  /* for recording reclaim priority */
210
211         /*
212          * While reclaiming in a hierarchy, we cache the last child we
213          * reclaimed from.
214          */
215         int last_scanned_child;
216         /*
217          * Should the accounting and control be hierarchical, per subtree?
218          */
219         bool use_hierarchy;
220         unsigned long   last_oom_jiffies;
221         atomic_t        refcnt;
222
223         unsigned int    swappiness;
224
225         /* set when res.limit == memsw.limit */
226         bool            memsw_is_minimum;
227
228         /*
229          * statistics. This must be placed at the end of memcg.
230          */
231         struct mem_cgroup_stat stat;
232 };
233
234 /*
235  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
236  * limit reclaim to prevent infinite loops, if they ever occur.
237  */
238 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
239 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
240
241 enum charge_type {
242         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
243         MEM_CGROUP_CHARGE_TYPE_MAPPED,
244         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
245         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
246         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
247         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
248         NR_CHARGE_TYPE,
249 };
250
251 /* only for here (for easy reading.) */
252 #define PCGF_CACHE      (1UL << PCG_CACHE)
253 #define PCGF_USED       (1UL << PCG_USED)
254 #define PCGF_LOCK       (1UL << PCG_LOCK)
255 /* Not used, but added here for completeness */
256 #define PCGF_ACCT       (1UL << PCG_ACCT)
257
258 /* for encoding cft->private value on file */
259 #define _MEM                    (0)
260 #define _MEMSWAP                (1)
261 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
262 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
263 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
264
265 /*
266  * Reclaim flags for mem_cgroup_hierarchical_reclaim
267  */
268 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
269 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
270 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
271 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
272 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
273 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
274
275 static void mem_cgroup_get(struct mem_cgroup *mem);
276 static void mem_cgroup_put(struct mem_cgroup *mem);
277 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
278 static void drain_all_stock_async(void);
279
280 static struct mem_cgroup_per_zone *
281 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
282 {
283         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
284 }
285
286 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
287 {
288         return &mem->css;
289 }
290
291 static struct mem_cgroup_per_zone *
292 page_cgroup_zoneinfo(struct page_cgroup *pc)
293 {
294         struct mem_cgroup *mem = pc->mem_cgroup;
295         int nid = page_cgroup_nid(pc);
296         int zid = page_cgroup_zid(pc);
297
298         if (!mem)
299                 return NULL;
300
301         return mem_cgroup_zoneinfo(mem, nid, zid);
302 }
303
304 static struct mem_cgroup_tree_per_zone *
305 soft_limit_tree_node_zone(int nid, int zid)
306 {
307         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
308 }
309
310 static struct mem_cgroup_tree_per_zone *
311 soft_limit_tree_from_page(struct page *page)
312 {
313         int nid = page_to_nid(page);
314         int zid = page_zonenum(page);
315
316         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
317 }
318
319 static void
320 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
321                                 struct mem_cgroup_per_zone *mz,
322                                 struct mem_cgroup_tree_per_zone *mctz,
323                                 unsigned long long new_usage_in_excess)
324 {
325         struct rb_node **p = &mctz->rb_root.rb_node;
326         struct rb_node *parent = NULL;
327         struct mem_cgroup_per_zone *mz_node;
328
329         if (mz->on_tree)
330                 return;
331
332         mz->usage_in_excess = new_usage_in_excess;
333         if (!mz->usage_in_excess)
334                 return;
335         while (*p) {
336                 parent = *p;
337                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
338                                         tree_node);
339                 if (mz->usage_in_excess < mz_node->usage_in_excess)
340                         p = &(*p)->rb_left;
341                 /*
342                  * We can't avoid mem cgroups that are over their soft
343                  * limit by the same amount
344                  */
345                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
346                         p = &(*p)->rb_right;
347         }
348         rb_link_node(&mz->tree_node, parent, p);
349         rb_insert_color(&mz->tree_node, &mctz->rb_root);
350         mz->on_tree = true;
351 }
352
353 static void
354 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
355                                 struct mem_cgroup_per_zone *mz,
356                                 struct mem_cgroup_tree_per_zone *mctz)
357 {
358         if (!mz->on_tree)
359                 return;
360         rb_erase(&mz->tree_node, &mctz->rb_root);
361         mz->on_tree = false;
362 }
363
364 static void
365 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
366                                 struct mem_cgroup_per_zone *mz,
367                                 struct mem_cgroup_tree_per_zone *mctz)
368 {
369         spin_lock(&mctz->lock);
370         __mem_cgroup_remove_exceeded(mem, mz, mctz);
371         spin_unlock(&mctz->lock);
372 }
373
374 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
375 {
376         bool ret = false;
377         int cpu;
378         s64 val;
379         struct mem_cgroup_stat_cpu *cpustat;
380
381         cpu = get_cpu();
382         cpustat = &mem->stat.cpustat[cpu];
383         val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
384         if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
385                 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
386                 ret = true;
387         }
388         put_cpu();
389         return ret;
390 }
391
392 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
393 {
394         unsigned long long excess;
395         struct mem_cgroup_per_zone *mz;
396         struct mem_cgroup_tree_per_zone *mctz;
397         int nid = page_to_nid(page);
398         int zid = page_zonenum(page);
399         mctz = soft_limit_tree_from_page(page);
400
401         /*
402          * Necessary to update all ancestors when hierarchy is used.
403          * because their event counter is not touched.
404          */
405         for (; mem; mem = parent_mem_cgroup(mem)) {
406                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
407                 excess = res_counter_soft_limit_excess(&mem->res);
408                 /*
409                  * We have to update the tree if mz is on RB-tree or
410                  * mem is over its softlimit.
411                  */
412                 if (excess || mz->on_tree) {
413                         spin_lock(&mctz->lock);
414                         /* if on-tree, remove it */
415                         if (mz->on_tree)
416                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
417                         /*
418                          * Insert again. mz->usage_in_excess will be updated.
419                          * If excess is 0, no tree ops.
420                          */
421                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
422                         spin_unlock(&mctz->lock);
423                 }
424         }
425 }
426
427 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
428 {
429         int node, zone;
430         struct mem_cgroup_per_zone *mz;
431         struct mem_cgroup_tree_per_zone *mctz;
432
433         for_each_node_state(node, N_POSSIBLE) {
434                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
435                         mz = mem_cgroup_zoneinfo(mem, node, zone);
436                         mctz = soft_limit_tree_node_zone(node, zone);
437                         mem_cgroup_remove_exceeded(mem, mz, mctz);
438                 }
439         }
440 }
441
442 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
443 {
444         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
445 }
446
447 static struct mem_cgroup_per_zone *
448 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
449 {
450         struct rb_node *rightmost = NULL;
451         struct mem_cgroup_per_zone *mz;
452
453 retry:
454         mz = NULL;
455         rightmost = rb_last(&mctz->rb_root);
456         if (!rightmost)
457                 goto done;              /* Nothing to reclaim from */
458
459         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
460         /*
461          * Remove the node now but someone else can add it back,
462          * we will to add it back at the end of reclaim to its correct
463          * position in the tree.
464          */
465         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
466         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
467                 !css_tryget(&mz->mem->css))
468                 goto retry;
469 done:
470         return mz;
471 }
472
473 static struct mem_cgroup_per_zone *
474 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
475 {
476         struct mem_cgroup_per_zone *mz;
477
478         spin_lock(&mctz->lock);
479         mz = __mem_cgroup_largest_soft_limit_node(mctz);
480         spin_unlock(&mctz->lock);
481         return mz;
482 }
483
484 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
485                                          bool charge)
486 {
487         int val = (charge) ? 1 : -1;
488         struct mem_cgroup_stat *stat = &mem->stat;
489         struct mem_cgroup_stat_cpu *cpustat;
490         int cpu = get_cpu();
491
492         cpustat = &stat->cpustat[cpu];
493         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
494         put_cpu();
495 }
496
497 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
498                                          struct page_cgroup *pc,
499                                          bool charge)
500 {
501         int val = (charge) ? 1 : -1;
502         struct mem_cgroup_stat *stat = &mem->stat;
503         struct mem_cgroup_stat_cpu *cpustat;
504         int cpu = get_cpu();
505
506         cpustat = &stat->cpustat[cpu];
507         if (PageCgroupCache(pc))
508                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
509         else
510                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
511
512         if (charge)
513                 __mem_cgroup_stat_add_safe(cpustat,
514                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
515         else
516                 __mem_cgroup_stat_add_safe(cpustat,
517                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
518         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1);
519         put_cpu();
520 }
521
522 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
523                                         enum lru_list idx)
524 {
525         int nid, zid;
526         struct mem_cgroup_per_zone *mz;
527         u64 total = 0;
528
529         for_each_online_node(nid)
530                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
531                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
532                         total += MEM_CGROUP_ZSTAT(mz, idx);
533                 }
534         return total;
535 }
536
537 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
538 {
539         return container_of(cgroup_subsys_state(cont,
540                                 mem_cgroup_subsys_id), struct mem_cgroup,
541                                 css);
542 }
543
544 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
545 {
546         /*
547          * mm_update_next_owner() may clear mm->owner to NULL
548          * if it races with swapoff, page migration, etc.
549          * So this can be called with p == NULL.
550          */
551         if (unlikely(!p))
552                 return NULL;
553
554         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
555                                 struct mem_cgroup, css);
556 }
557
558 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
559 {
560         struct mem_cgroup *mem = NULL;
561
562         if (!mm)
563                 return NULL;
564         /*
565          * Because we have no locks, mm->owner's may be being moved to other
566          * cgroup. We use css_tryget() here even if this looks
567          * pessimistic (rather than adding locks here).
568          */
569         rcu_read_lock();
570         do {
571                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
572                 if (unlikely(!mem))
573                         break;
574         } while (!css_tryget(&mem->css));
575         rcu_read_unlock();
576         return mem;
577 }
578
579 /*
580  * Call callback function against all cgroup under hierarchy tree.
581  */
582 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
583                           int (*func)(struct mem_cgroup *, void *))
584 {
585         int found, ret, nextid;
586         struct cgroup_subsys_state *css;
587         struct mem_cgroup *mem;
588
589         if (!root->use_hierarchy)
590                 return (*func)(root, data);
591
592         nextid = 1;
593         do {
594                 ret = 0;
595                 mem = NULL;
596
597                 rcu_read_lock();
598                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
599                                    &found);
600                 if (css && css_tryget(css))
601                         mem = container_of(css, struct mem_cgroup, css);
602                 rcu_read_unlock();
603
604                 if (mem) {
605                         ret = (*func)(mem, data);
606                         css_put(&mem->css);
607                 }
608                 nextid = found + 1;
609         } while (!ret && css);
610
611         return ret;
612 }
613
614 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
615 {
616         return (mem == root_mem_cgroup);
617 }
618
619 /*
620  * Following LRU functions are allowed to be used without PCG_LOCK.
621  * Operations are called by routine of global LRU independently from memcg.
622  * What we have to take care of here is validness of pc->mem_cgroup.
623  *
624  * Changes to pc->mem_cgroup happens when
625  * 1. charge
626  * 2. moving account
627  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
628  * It is added to LRU before charge.
629  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
630  * When moving account, the page is not on LRU. It's isolated.
631  */
632
633 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
634 {
635         struct page_cgroup *pc;
636         struct mem_cgroup_per_zone *mz;
637
638         if (mem_cgroup_disabled())
639                 return;
640         pc = lookup_page_cgroup(page);
641         /* can happen while we handle swapcache. */
642         if (!TestClearPageCgroupAcctLRU(pc))
643                 return;
644         VM_BUG_ON(!pc->mem_cgroup);
645         /*
646          * We don't check PCG_USED bit. It's cleared when the "page" is finally
647          * removed from global LRU.
648          */
649         mz = page_cgroup_zoneinfo(pc);
650         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
651         if (mem_cgroup_is_root(pc->mem_cgroup))
652                 return;
653         VM_BUG_ON(list_empty(&pc->lru));
654         list_del_init(&pc->lru);
655         return;
656 }
657
658 void mem_cgroup_del_lru(struct page *page)
659 {
660         mem_cgroup_del_lru_list(page, page_lru(page));
661 }
662
663 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
664 {
665         struct mem_cgroup_per_zone *mz;
666         struct page_cgroup *pc;
667
668         if (mem_cgroup_disabled())
669                 return;
670
671         pc = lookup_page_cgroup(page);
672         /*
673          * Used bit is set without atomic ops but after smp_wmb().
674          * For making pc->mem_cgroup visible, insert smp_rmb() here.
675          */
676         smp_rmb();
677         /* unused or root page is not rotated. */
678         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
679                 return;
680         mz = page_cgroup_zoneinfo(pc);
681         list_move(&pc->lru, &mz->lists[lru]);
682 }
683
684 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
685 {
686         struct page_cgroup *pc;
687         struct mem_cgroup_per_zone *mz;
688
689         if (mem_cgroup_disabled())
690                 return;
691         pc = lookup_page_cgroup(page);
692         VM_BUG_ON(PageCgroupAcctLRU(pc));
693         /*
694          * Used bit is set without atomic ops but after smp_wmb().
695          * For making pc->mem_cgroup visible, insert smp_rmb() here.
696          */
697         smp_rmb();
698         if (!PageCgroupUsed(pc))
699                 return;
700
701         mz = page_cgroup_zoneinfo(pc);
702         MEM_CGROUP_ZSTAT(mz, lru) += 1;
703         SetPageCgroupAcctLRU(pc);
704         if (mem_cgroup_is_root(pc->mem_cgroup))
705                 return;
706         list_add(&pc->lru, &mz->lists[lru]);
707 }
708
709 /*
710  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
711  * lru because the page may.be reused after it's fully uncharged (because of
712  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
713  * it again. This function is only used to charge SwapCache. It's done under
714  * lock_page and expected that zone->lru_lock is never held.
715  */
716 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
717 {
718         unsigned long flags;
719         struct zone *zone = page_zone(page);
720         struct page_cgroup *pc = lookup_page_cgroup(page);
721
722         spin_lock_irqsave(&zone->lru_lock, flags);
723         /*
724          * Forget old LRU when this page_cgroup is *not* used. This Used bit
725          * is guarded by lock_page() because the page is SwapCache.
726          */
727         if (!PageCgroupUsed(pc))
728                 mem_cgroup_del_lru_list(page, page_lru(page));
729         spin_unlock_irqrestore(&zone->lru_lock, flags);
730 }
731
732 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
733 {
734         unsigned long flags;
735         struct zone *zone = page_zone(page);
736         struct page_cgroup *pc = lookup_page_cgroup(page);
737
738         spin_lock_irqsave(&zone->lru_lock, flags);
739         /* link when the page is linked to LRU but page_cgroup isn't */
740         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
741                 mem_cgroup_add_lru_list(page, page_lru(page));
742         spin_unlock_irqrestore(&zone->lru_lock, flags);
743 }
744
745
746 void mem_cgroup_move_lists(struct page *page,
747                            enum lru_list from, enum lru_list to)
748 {
749         if (mem_cgroup_disabled())
750                 return;
751         mem_cgroup_del_lru_list(page, from);
752         mem_cgroup_add_lru_list(page, to);
753 }
754
755 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
756 {
757         int ret;
758         struct mem_cgroup *curr = NULL;
759
760         task_lock(task);
761         rcu_read_lock();
762         curr = try_get_mem_cgroup_from_mm(task->mm);
763         rcu_read_unlock();
764         task_unlock(task);
765         if (!curr)
766                 return 0;
767         /*
768          * We should check use_hierarchy of "mem" not "curr". Because checking
769          * use_hierarchy of "curr" here make this function true if hierarchy is
770          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
771          * hierarchy(even if use_hierarchy is disabled in "mem").
772          */
773         if (mem->use_hierarchy)
774                 ret = css_is_ancestor(&curr->css, &mem->css);
775         else
776                 ret = (curr == mem);
777         css_put(&curr->css);
778         return ret;
779 }
780
781 /*
782  * prev_priority control...this will be used in memory reclaim path.
783  */
784 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
785 {
786         int prev_priority;
787
788         spin_lock(&mem->reclaim_param_lock);
789         prev_priority = mem->prev_priority;
790         spin_unlock(&mem->reclaim_param_lock);
791
792         return prev_priority;
793 }
794
795 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
796 {
797         spin_lock(&mem->reclaim_param_lock);
798         if (priority < mem->prev_priority)
799                 mem->prev_priority = priority;
800         spin_unlock(&mem->reclaim_param_lock);
801 }
802
803 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
804 {
805         spin_lock(&mem->reclaim_param_lock);
806         mem->prev_priority = priority;
807         spin_unlock(&mem->reclaim_param_lock);
808 }
809
810 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
811 {
812         unsigned long active;
813         unsigned long inactive;
814         unsigned long gb;
815         unsigned long inactive_ratio;
816
817         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
818         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
819
820         gb = (inactive + active) >> (30 - PAGE_SHIFT);
821         if (gb)
822                 inactive_ratio = int_sqrt(10 * gb);
823         else
824                 inactive_ratio = 1;
825
826         if (present_pages) {
827                 present_pages[0] = inactive;
828                 present_pages[1] = active;
829         }
830
831         return inactive_ratio;
832 }
833
834 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
835 {
836         unsigned long active;
837         unsigned long inactive;
838         unsigned long present_pages[2];
839         unsigned long inactive_ratio;
840
841         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
842
843         inactive = present_pages[0];
844         active = present_pages[1];
845
846         if (inactive * inactive_ratio < active)
847                 return 1;
848
849         return 0;
850 }
851
852 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
853 {
854         unsigned long active;
855         unsigned long inactive;
856
857         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
858         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
859
860         return (active > inactive);
861 }
862
863 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
864                                        struct zone *zone,
865                                        enum lru_list lru)
866 {
867         int nid = zone->zone_pgdat->node_id;
868         int zid = zone_idx(zone);
869         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
870
871         return MEM_CGROUP_ZSTAT(mz, lru);
872 }
873
874 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
875                                                       struct zone *zone)
876 {
877         int nid = zone->zone_pgdat->node_id;
878         int zid = zone_idx(zone);
879         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
880
881         return &mz->reclaim_stat;
882 }
883
884 struct zone_reclaim_stat *
885 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
886 {
887         struct page_cgroup *pc;
888         struct mem_cgroup_per_zone *mz;
889
890         if (mem_cgroup_disabled())
891                 return NULL;
892
893         pc = lookup_page_cgroup(page);
894         /*
895          * Used bit is set without atomic ops but after smp_wmb().
896          * For making pc->mem_cgroup visible, insert smp_rmb() here.
897          */
898         smp_rmb();
899         if (!PageCgroupUsed(pc))
900                 return NULL;
901
902         mz = page_cgroup_zoneinfo(pc);
903         if (!mz)
904                 return NULL;
905
906         return &mz->reclaim_stat;
907 }
908
909 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
910                                         struct list_head *dst,
911                                         unsigned long *scanned, int order,
912                                         int mode, struct zone *z,
913                                         struct mem_cgroup *mem_cont,
914                                         int active, int file)
915 {
916         unsigned long nr_taken = 0;
917         struct page *page;
918         unsigned long scan;
919         LIST_HEAD(pc_list);
920         struct list_head *src;
921         struct page_cgroup *pc, *tmp;
922         int nid = z->zone_pgdat->node_id;
923         int zid = zone_idx(z);
924         struct mem_cgroup_per_zone *mz;
925         int lru = LRU_FILE * file + active;
926         int ret;
927
928         BUG_ON(!mem_cont);
929         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
930         src = &mz->lists[lru];
931
932         scan = 0;
933         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
934                 if (scan >= nr_to_scan)
935                         break;
936
937                 page = pc->page;
938                 if (unlikely(!PageCgroupUsed(pc)))
939                         continue;
940                 if (unlikely(!PageLRU(page)))
941                         continue;
942
943                 scan++;
944                 ret = __isolate_lru_page(page, mode, file);
945                 switch (ret) {
946                 case 0:
947                         list_move(&page->lru, dst);
948                         mem_cgroup_del_lru(page);
949                         nr_taken++;
950                         break;
951                 case -EBUSY:
952                         /* we don't affect global LRU but rotate in our LRU */
953                         mem_cgroup_rotate_lru_list(page, page_lru(page));
954                         break;
955                 default:
956                         break;
957                 }
958         }
959
960         *scanned = scan;
961         return nr_taken;
962 }
963
964 #define mem_cgroup_from_res_counter(counter, member)    \
965         container_of(counter, struct mem_cgroup, member)
966
967 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
968 {
969         if (do_swap_account) {
970                 if (res_counter_check_under_limit(&mem->res) &&
971                         res_counter_check_under_limit(&mem->memsw))
972                         return true;
973         } else
974                 if (res_counter_check_under_limit(&mem->res))
975                         return true;
976         return false;
977 }
978
979 static unsigned int get_swappiness(struct mem_cgroup *memcg)
980 {
981         struct cgroup *cgrp = memcg->css.cgroup;
982         unsigned int swappiness;
983
984         /* root ? */
985         if (cgrp->parent == NULL)
986                 return vm_swappiness;
987
988         spin_lock(&memcg->reclaim_param_lock);
989         swappiness = memcg->swappiness;
990         spin_unlock(&memcg->reclaim_param_lock);
991
992         return swappiness;
993 }
994
995 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
996 {
997         int *val = data;
998         (*val)++;
999         return 0;
1000 }
1001
1002 /**
1003  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
1004  * @memcg: The memory cgroup that went over limit
1005  * @p: Task that is going to be killed
1006  *
1007  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1008  * enabled
1009  */
1010 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1011 {
1012         struct cgroup *task_cgrp;
1013         struct cgroup *mem_cgrp;
1014         /*
1015          * Need a buffer in BSS, can't rely on allocations. The code relies
1016          * on the assumption that OOM is serialized for memory controller.
1017          * If this assumption is broken, revisit this code.
1018          */
1019         static char memcg_name[PATH_MAX];
1020         int ret;
1021
1022         if (!memcg || !p)
1023                 return;
1024
1025
1026         rcu_read_lock();
1027
1028         mem_cgrp = memcg->css.cgroup;
1029         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1030
1031         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1032         if (ret < 0) {
1033                 /*
1034                  * Unfortunately, we are unable to convert to a useful name
1035                  * But we'll still print out the usage information
1036                  */
1037                 rcu_read_unlock();
1038                 goto done;
1039         }
1040         rcu_read_unlock();
1041
1042         printk(KERN_INFO "Task in %s killed", memcg_name);
1043
1044         rcu_read_lock();
1045         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1046         if (ret < 0) {
1047                 rcu_read_unlock();
1048                 goto done;
1049         }
1050         rcu_read_unlock();
1051
1052         /*
1053          * Continues from above, so we don't need an KERN_ level
1054          */
1055         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1056 done:
1057
1058         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1059                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1060                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1061                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1062         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1063                 "failcnt %llu\n",
1064                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1065                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1066                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1067 }
1068
1069 /*
1070  * This function returns the number of memcg under hierarchy tree. Returns
1071  * 1(self count) if no children.
1072  */
1073 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1074 {
1075         int num = 0;
1076         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1077         return num;
1078 }
1079
1080 /*
1081  * Visit the first child (need not be the first child as per the ordering
1082  * of the cgroup list, since we track last_scanned_child) of @mem and use
1083  * that to reclaim free pages from.
1084  */
1085 static struct mem_cgroup *
1086 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1087 {
1088         struct mem_cgroup *ret = NULL;
1089         struct cgroup_subsys_state *css;
1090         int nextid, found;
1091
1092         if (!root_mem->use_hierarchy) {
1093                 css_get(&root_mem->css);
1094                 ret = root_mem;
1095         }
1096
1097         while (!ret) {
1098                 rcu_read_lock();
1099                 nextid = root_mem->last_scanned_child + 1;
1100                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1101                                    &found);
1102                 if (css && css_tryget(css))
1103                         ret = container_of(css, struct mem_cgroup, css);
1104
1105                 rcu_read_unlock();
1106                 /* Updates scanning parameter */
1107                 spin_lock(&root_mem->reclaim_param_lock);
1108                 if (!css) {
1109                         /* this means start scan from ID:1 */
1110                         root_mem->last_scanned_child = 0;
1111                 } else
1112                         root_mem->last_scanned_child = found;
1113                 spin_unlock(&root_mem->reclaim_param_lock);
1114         }
1115
1116         return ret;
1117 }
1118
1119 /*
1120  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1121  * we reclaimed from, so that we don't end up penalizing one child extensively
1122  * based on its position in the children list.
1123  *
1124  * root_mem is the original ancestor that we've been reclaim from.
1125  *
1126  * We give up and return to the caller when we visit root_mem twice.
1127  * (other groups can be removed while we're walking....)
1128  *
1129  * If shrink==true, for avoiding to free too much, this returns immedieately.
1130  */
1131 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1132                                                 struct zone *zone,
1133                                                 gfp_t gfp_mask,
1134                                                 unsigned long reclaim_options)
1135 {
1136         struct mem_cgroup *victim;
1137         int ret, total = 0;
1138         int loop = 0;
1139         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1140         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1141         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1142         unsigned long excess = mem_cgroup_get_excess(root_mem);
1143
1144         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1145         if (root_mem->memsw_is_minimum)
1146                 noswap = true;
1147
1148         while (1) {
1149                 victim = mem_cgroup_select_victim(root_mem);
1150                 if (victim == root_mem) {
1151                         loop++;
1152                         if (loop >= 1)
1153                                 drain_all_stock_async();
1154                         if (loop >= 2) {
1155                                 /*
1156                                  * If we have not been able to reclaim
1157                                  * anything, it might because there are
1158                                  * no reclaimable pages under this hierarchy
1159                                  */
1160                                 if (!check_soft || !total) {
1161                                         css_put(&victim->css);
1162                                         break;
1163                                 }
1164                                 /*
1165                                  * We want to do more targetted reclaim.
1166                                  * excess >> 2 is not to excessive so as to
1167                                  * reclaim too much, nor too less that we keep
1168                                  * coming back to reclaim from this cgroup
1169                                  */
1170                                 if (total >= (excess >> 2) ||
1171                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1172                                         css_put(&victim->css);
1173                                         break;
1174                                 }
1175                         }
1176                 }
1177                 if (!mem_cgroup_local_usage(&victim->stat)) {
1178                         /* this cgroup's local usage == 0 */
1179                         css_put(&victim->css);
1180                         continue;
1181                 }
1182                 /* we use swappiness of local cgroup */
1183                 if (check_soft)
1184                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1185                                 noswap, get_swappiness(victim), zone,
1186                                 zone->zone_pgdat->node_id);
1187                 else
1188                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1189                                                 noswap, get_swappiness(victim));
1190                 css_put(&victim->css);
1191                 /*
1192                  * At shrinking usage, we can't check we should stop here or
1193                  * reclaim more. It's depends on callers. last_scanned_child
1194                  * will work enough for keeping fairness under tree.
1195                  */
1196                 if (shrink)
1197                         return ret;
1198                 total += ret;
1199                 if (check_soft) {
1200                         if (res_counter_check_under_soft_limit(&root_mem->res))
1201                                 return total;
1202                 } else if (mem_cgroup_check_under_limit(root_mem))
1203                         return 1 + total;
1204         }
1205         return total;
1206 }
1207
1208 bool mem_cgroup_oom_called(struct task_struct *task)
1209 {
1210         bool ret = false;
1211         struct mem_cgroup *mem;
1212         struct mm_struct *mm;
1213
1214         rcu_read_lock();
1215         mm = task->mm;
1216         if (!mm)
1217                 mm = &init_mm;
1218         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1219         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1220                 ret = true;
1221         rcu_read_unlock();
1222         return ret;
1223 }
1224
1225 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
1226 {
1227         mem->last_oom_jiffies = jiffies;
1228         return 0;
1229 }
1230
1231 static void record_last_oom(struct mem_cgroup *mem)
1232 {
1233         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
1234 }
1235
1236 /*
1237  * Currently used to update mapped file statistics, but the routine can be
1238  * generalized to update other statistics as well.
1239  */
1240 void mem_cgroup_update_file_mapped(struct page *page, int val)
1241 {
1242         struct mem_cgroup *mem;
1243         struct mem_cgroup_stat *stat;
1244         struct mem_cgroup_stat_cpu *cpustat;
1245         int cpu;
1246         struct page_cgroup *pc;
1247
1248         pc = lookup_page_cgroup(page);
1249         if (unlikely(!pc))
1250                 return;
1251
1252         lock_page_cgroup(pc);
1253         mem = pc->mem_cgroup;
1254         if (!mem)
1255                 goto done;
1256
1257         if (!PageCgroupUsed(pc))
1258                 goto done;
1259
1260         /*
1261          * Preemption is already disabled, we don't need get_cpu()
1262          */
1263         cpu = smp_processor_id();
1264         stat = &mem->stat;
1265         cpustat = &stat->cpustat[cpu];
1266
1267         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED, val);
1268 done:
1269         unlock_page_cgroup(pc);
1270 }
1271
1272 /*
1273  * size of first charge trial. "32" comes from vmscan.c's magic value.
1274  * TODO: maybe necessary to use big numbers in big irons.
1275  */
1276 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1277 struct memcg_stock_pcp {
1278         struct mem_cgroup *cached; /* this never be root cgroup */
1279         int charge;
1280         struct work_struct work;
1281 };
1282 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1283 static atomic_t memcg_drain_count;
1284
1285 /*
1286  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1287  * from local stock and true is returned. If the stock is 0 or charges from a
1288  * cgroup which is not current target, returns false. This stock will be
1289  * refilled.
1290  */
1291 static bool consume_stock(struct mem_cgroup *mem)
1292 {
1293         struct memcg_stock_pcp *stock;
1294         bool ret = true;
1295
1296         stock = &get_cpu_var(memcg_stock);
1297         if (mem == stock->cached && stock->charge)
1298                 stock->charge -= PAGE_SIZE;
1299         else /* need to call res_counter_charge */
1300                 ret = false;
1301         put_cpu_var(memcg_stock);
1302         return ret;
1303 }
1304
1305 /*
1306  * Returns stocks cached in percpu to res_counter and reset cached information.
1307  */
1308 static void drain_stock(struct memcg_stock_pcp *stock)
1309 {
1310         struct mem_cgroup *old = stock->cached;
1311
1312         if (stock->charge) {
1313                 res_counter_uncharge(&old->res, stock->charge);
1314                 if (do_swap_account)
1315                         res_counter_uncharge(&old->memsw, stock->charge);
1316         }
1317         stock->cached = NULL;
1318         stock->charge = 0;
1319 }
1320
1321 /*
1322  * This must be called under preempt disabled or must be called by
1323  * a thread which is pinned to local cpu.
1324  */
1325 static void drain_local_stock(struct work_struct *dummy)
1326 {
1327         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1328         drain_stock(stock);
1329 }
1330
1331 /*
1332  * Cache charges(val) which is from res_counter, to local per_cpu area.
1333  * This will be consumed by consumt_stock() function, later.
1334  */
1335 static void refill_stock(struct mem_cgroup *mem, int val)
1336 {
1337         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1338
1339         if (stock->cached != mem) { /* reset if necessary */
1340                 drain_stock(stock);
1341                 stock->cached = mem;
1342         }
1343         stock->charge += val;
1344         put_cpu_var(memcg_stock);
1345 }
1346
1347 /*
1348  * Tries to drain stocked charges in other cpus. This function is asynchronous
1349  * and just put a work per cpu for draining localy on each cpu. Caller can
1350  * expects some charges will be back to res_counter later but cannot wait for
1351  * it.
1352  */
1353 static void drain_all_stock_async(void)
1354 {
1355         int cpu;
1356         /* This function is for scheduling "drain" in asynchronous way.
1357          * The result of "drain" is not directly handled by callers. Then,
1358          * if someone is calling drain, we don't have to call drain more.
1359          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1360          * there is a race. We just do loose check here.
1361          */
1362         if (atomic_read(&memcg_drain_count))
1363                 return;
1364         /* Notify other cpus that system-wide "drain" is running */
1365         atomic_inc(&memcg_drain_count);
1366         get_online_cpus();
1367         for_each_online_cpu(cpu) {
1368                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1369                 schedule_work_on(cpu, &stock->work);
1370         }
1371         put_online_cpus();
1372         atomic_dec(&memcg_drain_count);
1373         /* We don't wait for flush_work */
1374 }
1375
1376 /* This is a synchronous drain interface. */
1377 static void drain_all_stock_sync(void)
1378 {
1379         /* called when force_empty is called */
1380         atomic_inc(&memcg_drain_count);
1381         schedule_on_each_cpu(drain_local_stock);
1382         atomic_dec(&memcg_drain_count);
1383 }
1384
1385 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1386                                         unsigned long action,
1387                                         void *hcpu)
1388 {
1389         int cpu = (unsigned long)hcpu;
1390         struct memcg_stock_pcp *stock;
1391
1392         if (action != CPU_DEAD)
1393                 return NOTIFY_OK;
1394         stock = &per_cpu(memcg_stock, cpu);
1395         drain_stock(stock);
1396         return NOTIFY_OK;
1397 }
1398
1399 /*
1400  * Unlike exported interface, "oom" parameter is added. if oom==true,
1401  * oom-killer can be invoked.
1402  */
1403 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1404                         gfp_t gfp_mask, struct mem_cgroup **memcg,
1405                         bool oom, struct page *page)
1406 {
1407         struct mem_cgroup *mem, *mem_over_limit;
1408         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1409         struct res_counter *fail_res;
1410         int csize = CHARGE_SIZE;
1411
1412         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1413                 /* Don't account this! */
1414                 *memcg = NULL;
1415                 return 0;
1416         }
1417
1418         /*
1419          * We always charge the cgroup the mm_struct belongs to.
1420          * The mm_struct's mem_cgroup changes on task migration if the
1421          * thread group leader migrates. It's possible that mm is not
1422          * set, if so charge the init_mm (happens for pagecache usage).
1423          */
1424         mem = *memcg;
1425         if (likely(!mem)) {
1426                 mem = try_get_mem_cgroup_from_mm(mm);
1427                 *memcg = mem;
1428         } else {
1429                 css_get(&mem->css);
1430         }
1431         if (unlikely(!mem))
1432                 return 0;
1433
1434         VM_BUG_ON(css_is_removed(&mem->css));
1435         if (mem_cgroup_is_root(mem))
1436                 goto done;
1437
1438         while (1) {
1439                 int ret = 0;
1440                 unsigned long flags = 0;
1441
1442                 if (consume_stock(mem))
1443                         goto charged;
1444
1445                 ret = res_counter_charge(&mem->res, csize, &fail_res);
1446                 if (likely(!ret)) {
1447                         if (!do_swap_account)
1448                                 break;
1449                         ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1450                         if (likely(!ret))
1451                                 break;
1452                         /* mem+swap counter fails */
1453                         res_counter_uncharge(&mem->res, csize);
1454                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1455                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1456                                                                         memsw);
1457                 } else
1458                         /* mem counter fails */
1459                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1460                                                                         res);
1461
1462                 /* reduce request size and retry */
1463                 if (csize > PAGE_SIZE) {
1464                         csize = PAGE_SIZE;
1465                         continue;
1466                 }
1467                 if (!(gfp_mask & __GFP_WAIT))
1468                         goto nomem;
1469
1470                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1471                                                 gfp_mask, flags);
1472                 if (ret)
1473                         continue;
1474
1475                 /*
1476                  * try_to_free_mem_cgroup_pages() might not give us a full
1477                  * picture of reclaim. Some pages are reclaimed and might be
1478                  * moved to swap cache or just unmapped from the cgroup.
1479                  * Check the limit again to see if the reclaim reduced the
1480                  * current usage of the cgroup before giving up
1481                  *
1482                  */
1483                 if (mem_cgroup_check_under_limit(mem_over_limit))
1484                         continue;
1485
1486                 if (!nr_retries--) {
1487                         if (oom) {
1488                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1489                                 record_last_oom(mem_over_limit);
1490                         }
1491                         goto nomem;
1492                 }
1493         }
1494         if (csize > PAGE_SIZE)
1495                 refill_stock(mem, csize - PAGE_SIZE);
1496 charged:
1497         /*
1498          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1499          * if they exceeds softlimit.
1500          */
1501         if (mem_cgroup_soft_limit_check(mem))
1502                 mem_cgroup_update_tree(mem, page);
1503 done:
1504         return 0;
1505 nomem:
1506         css_put(&mem->css);
1507         return -ENOMEM;
1508 }
1509
1510 /*
1511  * Somemtimes we have to undo a charge we got by try_charge().
1512  * This function is for that and do uncharge, put css's refcnt.
1513  * gotten by try_charge().
1514  */
1515 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1516 {
1517         if (!mem_cgroup_is_root(mem)) {
1518                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1519                 if (do_swap_account)
1520                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1521         }
1522         css_put(&mem->css);
1523 }
1524
1525 /*
1526  * A helper function to get mem_cgroup from ID. must be called under
1527  * rcu_read_lock(). The caller must check css_is_removed() or some if
1528  * it's concern. (dropping refcnt from swap can be called against removed
1529  * memcg.)
1530  */
1531 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1532 {
1533         struct cgroup_subsys_state *css;
1534
1535         /* ID 0 is unused ID */
1536         if (!id)
1537                 return NULL;
1538         css = css_lookup(&mem_cgroup_subsys, id);
1539         if (!css)
1540                 return NULL;
1541         return container_of(css, struct mem_cgroup, css);
1542 }
1543
1544 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1545 {
1546         struct mem_cgroup *mem = NULL;
1547         struct page_cgroup *pc;
1548         unsigned short id;
1549         swp_entry_t ent;
1550
1551         VM_BUG_ON(!PageLocked(page));
1552
1553         pc = lookup_page_cgroup(page);
1554         lock_page_cgroup(pc);
1555         if (PageCgroupUsed(pc)) {
1556                 mem = pc->mem_cgroup;
1557                 if (mem && !css_tryget(&mem->css))
1558                         mem = NULL;
1559         } else if (PageSwapCache(page)) {
1560                 ent.val = page_private(page);
1561                 id = lookup_swap_cgroup(ent);
1562                 rcu_read_lock();
1563                 mem = mem_cgroup_lookup(id);
1564                 if (mem && !css_tryget(&mem->css))
1565                         mem = NULL;
1566                 rcu_read_unlock();
1567         }
1568         unlock_page_cgroup(pc);
1569         return mem;
1570 }
1571
1572 /*
1573  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1574  * USED state. If already USED, uncharge and return.
1575  */
1576
1577 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1578                                      struct page_cgroup *pc,
1579                                      enum charge_type ctype)
1580 {
1581         /* try_charge() can return NULL to *memcg, taking care of it. */
1582         if (!mem)
1583                 return;
1584
1585         lock_page_cgroup(pc);
1586         if (unlikely(PageCgroupUsed(pc))) {
1587                 unlock_page_cgroup(pc);
1588                 mem_cgroup_cancel_charge(mem);
1589                 return;
1590         }
1591
1592         pc->mem_cgroup = mem;
1593         /*
1594          * We access a page_cgroup asynchronously without lock_page_cgroup().
1595          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1596          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1597          * before USED bit, we need memory barrier here.
1598          * See mem_cgroup_add_lru_list(), etc.
1599          */
1600         smp_wmb();
1601         switch (ctype) {
1602         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1603         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1604                 SetPageCgroupCache(pc);
1605                 SetPageCgroupUsed(pc);
1606                 break;
1607         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1608                 ClearPageCgroupCache(pc);
1609                 SetPageCgroupUsed(pc);
1610                 break;
1611         default:
1612                 break;
1613         }
1614
1615         mem_cgroup_charge_statistics(mem, pc, true);
1616
1617         unlock_page_cgroup(pc);
1618 }
1619
1620 /**
1621  * __mem_cgroup_move_account - move account of the page
1622  * @pc: page_cgroup of the page.
1623  * @from: mem_cgroup which the page is moved from.
1624  * @to: mem_cgroup which the page is moved to. @from != @to.
1625  *
1626  * The caller must confirm following.
1627  * - page is not on LRU (isolate_page() is useful.)
1628  * - the pc is locked, used, and ->mem_cgroup points to @from.
1629  *
1630  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1631  * new cgroup. It should be done by a caller.
1632  */
1633
1634 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1635         struct mem_cgroup *from, struct mem_cgroup *to)
1636 {
1637         struct page *page;
1638         int cpu;
1639         struct mem_cgroup_stat *stat;
1640         struct mem_cgroup_stat_cpu *cpustat;
1641
1642         VM_BUG_ON(from == to);
1643         VM_BUG_ON(PageLRU(pc->page));
1644         VM_BUG_ON(!PageCgroupLocked(pc));
1645         VM_BUG_ON(!PageCgroupUsed(pc));
1646         VM_BUG_ON(pc->mem_cgroup != from);
1647
1648         if (!mem_cgroup_is_root(from))
1649                 res_counter_uncharge(&from->res, PAGE_SIZE);
1650         mem_cgroup_charge_statistics(from, pc, false);
1651
1652         page = pc->page;
1653         if (page_mapped(page) && !PageAnon(page)) {
1654                 cpu = smp_processor_id();
1655                 /* Update mapped_file data for mem_cgroup "from" */
1656                 stat = &from->stat;
1657                 cpustat = &stat->cpustat[cpu];
1658                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
1659                                                 -1);
1660
1661                 /* Update mapped_file data for mem_cgroup "to" */
1662                 stat = &to->stat;
1663                 cpustat = &stat->cpustat[cpu];
1664                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
1665                                                 1);
1666         }
1667
1668         if (do_swap_account && !mem_cgroup_is_root(from))
1669                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1670         css_put(&from->css);
1671
1672         css_get(&to->css);
1673         pc->mem_cgroup = to;
1674         mem_cgroup_charge_statistics(to, pc, true);
1675         /*
1676          * We charges against "to" which may not have any tasks. Then, "to"
1677          * can be under rmdir(). But in current implementation, caller of
1678          * this function is just force_empty() and it's garanteed that
1679          * "to" is never removed. So, we don't check rmdir status here.
1680          */
1681 }
1682
1683 /*
1684  * check whether the @pc is valid for moving account and call
1685  * __mem_cgroup_move_account()
1686  */
1687 static int mem_cgroup_move_account(struct page_cgroup *pc,
1688                                 struct mem_cgroup *from, struct mem_cgroup *to)
1689 {
1690         int ret = -EINVAL;
1691         lock_page_cgroup(pc);
1692         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1693                 __mem_cgroup_move_account(pc, from, to);
1694                 ret = 0;
1695         }
1696         unlock_page_cgroup(pc);
1697         return ret;
1698 }
1699
1700 /*
1701  * move charges to its parent.
1702  */
1703
1704 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1705                                   struct mem_cgroup *child,
1706                                   gfp_t gfp_mask)
1707 {
1708         struct page *page = pc->page;
1709         struct cgroup *cg = child->css.cgroup;
1710         struct cgroup *pcg = cg->parent;
1711         struct mem_cgroup *parent;
1712         int ret;
1713
1714         /* Is ROOT ? */
1715         if (!pcg)
1716                 return -EINVAL;
1717
1718         ret = -EBUSY;
1719         if (!get_page_unless_zero(page))
1720                 goto out;
1721         if (isolate_lru_page(page))
1722                 goto put;
1723
1724         parent = mem_cgroup_from_cont(pcg);
1725         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page);
1726         if (ret || !parent)
1727                 goto put_back;
1728
1729         ret = mem_cgroup_move_account(pc, child, parent);
1730         if (!ret)
1731                 css_put(&parent->css);  /* drop extra refcnt by try_charge() */
1732         else
1733                 mem_cgroup_cancel_charge(parent);       /* does css_put */
1734 put_back:
1735         putback_lru_page(page);
1736 put:
1737         put_page(page);
1738 out:
1739         return ret;
1740 }
1741
1742 /*
1743  * Charge the memory controller for page usage.
1744  * Return
1745  * 0 if the charge was successful
1746  * < 0 if the cgroup is over its limit
1747  */
1748 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1749                                 gfp_t gfp_mask, enum charge_type ctype,
1750                                 struct mem_cgroup *memcg)
1751 {
1752         struct mem_cgroup *mem;
1753         struct page_cgroup *pc;
1754         int ret;
1755
1756         pc = lookup_page_cgroup(page);
1757         /* can happen at boot */
1758         if (unlikely(!pc))
1759                 return 0;
1760         prefetchw(pc);
1761
1762         mem = memcg;
1763         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page);
1764         if (ret || !mem)
1765                 return ret;
1766
1767         __mem_cgroup_commit_charge(mem, pc, ctype);
1768         return 0;
1769 }
1770
1771 int mem_cgroup_newpage_charge(struct page *page,
1772                               struct mm_struct *mm, gfp_t gfp_mask)
1773 {
1774         if (mem_cgroup_disabled())
1775                 return 0;
1776         if (PageCompound(page))
1777                 return 0;
1778         /*
1779          * If already mapped, we don't have to account.
1780          * If page cache, page->mapping has address_space.
1781          * But page->mapping may have out-of-use anon_vma pointer,
1782          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1783          * is NULL.
1784          */
1785         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1786                 return 0;
1787         if (unlikely(!mm))
1788                 mm = &init_mm;
1789         return mem_cgroup_charge_common(page, mm, gfp_mask,
1790                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1791 }
1792
1793 static void
1794 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1795                                         enum charge_type ctype);
1796
1797 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1798                                 gfp_t gfp_mask)
1799 {
1800         struct mem_cgroup *mem = NULL;
1801         int ret;
1802
1803         if (mem_cgroup_disabled())
1804                 return 0;
1805         if (PageCompound(page))
1806                 return 0;
1807         /*
1808          * Corner case handling. This is called from add_to_page_cache()
1809          * in usual. But some FS (shmem) precharges this page before calling it
1810          * and call add_to_page_cache() with GFP_NOWAIT.
1811          *
1812          * For GFP_NOWAIT case, the page may be pre-charged before calling
1813          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1814          * charge twice. (It works but has to pay a bit larger cost.)
1815          * And when the page is SwapCache, it should take swap information
1816          * into account. This is under lock_page() now.
1817          */
1818         if (!(gfp_mask & __GFP_WAIT)) {
1819                 struct page_cgroup *pc;
1820
1821
1822                 pc = lookup_page_cgroup(page);
1823                 if (!pc)
1824                         return 0;
1825                 lock_page_cgroup(pc);
1826                 if (PageCgroupUsed(pc)) {
1827                         unlock_page_cgroup(pc);
1828                         return 0;
1829                 }
1830                 unlock_page_cgroup(pc);
1831         }
1832
1833         if (unlikely(!mm && !mem))
1834                 mm = &init_mm;
1835
1836         if (page_is_file_cache(page))
1837                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1838                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1839
1840         /* shmem */
1841         if (PageSwapCache(page)) {
1842                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1843                 if (!ret)
1844                         __mem_cgroup_commit_charge_swapin(page, mem,
1845                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1846         } else
1847                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1848                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1849
1850         return ret;
1851 }
1852
1853 /*
1854  * While swap-in, try_charge -> commit or cancel, the page is locked.
1855  * And when try_charge() successfully returns, one refcnt to memcg without
1856  * struct page_cgroup is acquired. This refcnt will be consumed by
1857  * "commit()" or removed by "cancel()"
1858  */
1859 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1860                                  struct page *page,
1861                                  gfp_t mask, struct mem_cgroup **ptr)
1862 {
1863         struct mem_cgroup *mem;
1864         int ret;
1865
1866         if (mem_cgroup_disabled())
1867                 return 0;
1868
1869         if (!do_swap_account)
1870                 goto charge_cur_mm;
1871         /*
1872          * A racing thread's fault, or swapoff, may have already updated
1873          * the pte, and even removed page from swap cache: in those cases
1874          * do_swap_page()'s pte_same() test will fail; but there's also a
1875          * KSM case which does need to charge the page.
1876          */
1877         if (!PageSwapCache(page))
1878                 goto charge_cur_mm;
1879         mem = try_get_mem_cgroup_from_page(page);
1880         if (!mem)
1881                 goto charge_cur_mm;
1882         *ptr = mem;
1883         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page);
1884         /* drop extra refcnt from tryget */
1885         css_put(&mem->css);
1886         return ret;
1887 charge_cur_mm:
1888         if (unlikely(!mm))
1889                 mm = &init_mm;
1890         return __mem_cgroup_try_charge(mm, mask, ptr, true, page);
1891 }
1892
1893 static void
1894 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1895                                         enum charge_type ctype)
1896 {
1897         struct page_cgroup *pc;
1898
1899         if (mem_cgroup_disabled())
1900                 return;
1901         if (!ptr)
1902                 return;
1903         cgroup_exclude_rmdir(&ptr->css);
1904         pc = lookup_page_cgroup(page);
1905         mem_cgroup_lru_del_before_commit_swapcache(page);
1906         __mem_cgroup_commit_charge(ptr, pc, ctype);
1907         mem_cgroup_lru_add_after_commit_swapcache(page);
1908         /*
1909          * Now swap is on-memory. This means this page may be
1910          * counted both as mem and swap....double count.
1911          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1912          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1913          * may call delete_from_swap_cache() before reach here.
1914          */
1915         if (do_swap_account && PageSwapCache(page)) {
1916                 swp_entry_t ent = {.val = page_private(page)};
1917                 unsigned short id;
1918                 struct mem_cgroup *memcg;
1919
1920                 id = swap_cgroup_record(ent, 0);
1921                 rcu_read_lock();
1922                 memcg = mem_cgroup_lookup(id);
1923                 if (memcg) {
1924                         /*
1925                          * This recorded memcg can be obsolete one. So, avoid
1926                          * calling css_tryget
1927                          */
1928                         if (!mem_cgroup_is_root(memcg))
1929                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1930                         mem_cgroup_swap_statistics(memcg, false);
1931                         mem_cgroup_put(memcg);
1932                 }
1933                 rcu_read_unlock();
1934         }
1935         /*
1936          * At swapin, we may charge account against cgroup which has no tasks.
1937          * So, rmdir()->pre_destroy() can be called while we do this charge.
1938          * In that case, we need to call pre_destroy() again. check it here.
1939          */
1940         cgroup_release_and_wakeup_rmdir(&ptr->css);
1941 }
1942
1943 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1944 {
1945         __mem_cgroup_commit_charge_swapin(page, ptr,
1946                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1947 }
1948
1949 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1950 {
1951         if (mem_cgroup_disabled())
1952                 return;
1953         if (!mem)
1954                 return;
1955         mem_cgroup_cancel_charge(mem);
1956 }
1957
1958 static void
1959 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
1960 {
1961         struct memcg_batch_info *batch = NULL;
1962         bool uncharge_memsw = true;
1963         /* If swapout, usage of swap doesn't decrease */
1964         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1965                 uncharge_memsw = false;
1966         /*
1967          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
1968          * In those cases, all pages freed continously can be expected to be in
1969          * the same cgroup and we have chance to coalesce uncharges.
1970          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
1971          * because we want to do uncharge as soon as possible.
1972          */
1973         if (!current->memcg_batch.do_batch || test_thread_flag(TIF_MEMDIE))
1974                 goto direct_uncharge;
1975
1976         batch = &current->memcg_batch;
1977         /*
1978          * In usual, we do css_get() when we remember memcg pointer.
1979          * But in this case, we keep res->usage until end of a series of
1980          * uncharges. Then, it's ok to ignore memcg's refcnt.
1981          */
1982         if (!batch->memcg)
1983                 batch->memcg = mem;
1984         /*
1985          * In typical case, batch->memcg == mem. This means we can
1986          * merge a series of uncharges to an uncharge of res_counter.
1987          * If not, we uncharge res_counter ony by one.
1988          */
1989         if (batch->memcg != mem)
1990                 goto direct_uncharge;
1991         /* remember freed charge and uncharge it later */
1992         batch->bytes += PAGE_SIZE;
1993         if (uncharge_memsw)
1994                 batch->memsw_bytes += PAGE_SIZE;
1995         return;
1996 direct_uncharge:
1997         res_counter_uncharge(&mem->res, PAGE_SIZE);
1998         if (uncharge_memsw)
1999                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2000         return;
2001 }
2002
2003 /*
2004  * uncharge if !page_mapped(page)
2005  */
2006 static struct mem_cgroup *
2007 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2008 {
2009         struct page_cgroup *pc;
2010         struct mem_cgroup *mem = NULL;
2011         struct mem_cgroup_per_zone *mz;
2012
2013         if (mem_cgroup_disabled())
2014                 return NULL;
2015
2016         if (PageSwapCache(page))
2017                 return NULL;
2018
2019         /*
2020          * Check if our page_cgroup is valid
2021          */
2022         pc = lookup_page_cgroup(page);
2023         if (unlikely(!pc || !PageCgroupUsed(pc)))
2024                 return NULL;
2025
2026         lock_page_cgroup(pc);
2027
2028         mem = pc->mem_cgroup;
2029
2030         if (!PageCgroupUsed(pc))
2031                 goto unlock_out;
2032
2033         switch (ctype) {
2034         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2035         case MEM_CGROUP_CHARGE_TYPE_DROP:
2036                 if (page_mapped(page))
2037                         goto unlock_out;
2038                 break;
2039         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2040                 if (!PageAnon(page)) {  /* Shared memory */
2041                         if (page->mapping && !page_is_file_cache(page))
2042                                 goto unlock_out;
2043                 } else if (page_mapped(page)) /* Anon */
2044                                 goto unlock_out;
2045                 break;
2046         default:
2047                 break;
2048         }
2049
2050         if (!mem_cgroup_is_root(mem))
2051                 __do_uncharge(mem, ctype);
2052         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2053                 mem_cgroup_swap_statistics(mem, true);
2054         mem_cgroup_charge_statistics(mem, pc, false);
2055
2056         ClearPageCgroupUsed(pc);
2057         /*
2058          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2059          * freed from LRU. This is safe because uncharged page is expected not
2060          * to be reused (freed soon). Exception is SwapCache, it's handled by
2061          * special functions.
2062          */
2063
2064         mz = page_cgroup_zoneinfo(pc);
2065         unlock_page_cgroup(pc);
2066
2067         if (mem_cgroup_soft_limit_check(mem))
2068                 mem_cgroup_update_tree(mem, page);
2069         /* at swapout, this memcg will be accessed to record to swap */
2070         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2071                 css_put(&mem->css);
2072
2073         return mem;
2074
2075 unlock_out:
2076         unlock_page_cgroup(pc);
2077         return NULL;
2078 }
2079
2080 void mem_cgroup_uncharge_page(struct page *page)
2081 {
2082         /* early check. */
2083         if (page_mapped(page))
2084                 return;
2085         if (page->mapping && !PageAnon(page))
2086                 return;
2087         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2088 }
2089
2090 void mem_cgroup_uncharge_cache_page(struct page *page)
2091 {
2092         VM_BUG_ON(page_mapped(page));
2093         VM_BUG_ON(page->mapping);
2094         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2095 }
2096
2097 /*
2098  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2099  * In that cases, pages are freed continuously and we can expect pages
2100  * are in the same memcg. All these calls itself limits the number of
2101  * pages freed at once, then uncharge_start/end() is called properly.
2102  * This may be called prural(2) times in a context,
2103  */
2104
2105 void mem_cgroup_uncharge_start(void)
2106 {
2107         current->memcg_batch.do_batch++;
2108         /* We can do nest. */
2109         if (current->memcg_batch.do_batch == 1) {
2110                 current->memcg_batch.memcg = NULL;
2111                 current->memcg_batch.bytes = 0;
2112                 current->memcg_batch.memsw_bytes = 0;
2113         }
2114 }
2115
2116 void mem_cgroup_uncharge_end(void)
2117 {
2118         struct memcg_batch_info *batch = &current->memcg_batch;
2119
2120         if (!batch->do_batch)
2121                 return;
2122
2123         batch->do_batch--;
2124         if (batch->do_batch) /* If stacked, do nothing. */
2125                 return;
2126
2127         if (!batch->memcg)
2128                 return;
2129         /*
2130          * This "batch->memcg" is valid without any css_get/put etc...
2131          * bacause we hide charges behind us.
2132          */
2133         if (batch->bytes)
2134                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2135         if (batch->memsw_bytes)
2136                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2137         /* forget this pointer (for sanity check) */
2138         batch->memcg = NULL;
2139 }
2140
2141 #ifdef CONFIG_SWAP
2142 /*
2143  * called after __delete_from_swap_cache() and drop "page" account.
2144  * memcg information is recorded to swap_cgroup of "ent"
2145  */
2146 void
2147 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2148 {
2149         struct mem_cgroup *memcg;
2150         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2151
2152         if (!swapout) /* this was a swap cache but the swap is unused ! */
2153                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2154
2155         memcg = __mem_cgroup_uncharge_common(page, ctype);
2156
2157         /* record memcg information */
2158         if (do_swap_account && swapout && memcg) {
2159                 swap_cgroup_record(ent, css_id(&memcg->css));
2160                 mem_cgroup_get(memcg);
2161         }
2162         if (swapout && memcg)
2163                 css_put(&memcg->css);
2164 }
2165 #endif
2166
2167 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2168 /*
2169  * called from swap_entry_free(). remove record in swap_cgroup and
2170  * uncharge "memsw" account.
2171  */
2172 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2173 {
2174         struct mem_cgroup *memcg;
2175         unsigned short id;
2176
2177         if (!do_swap_account)
2178                 return;
2179
2180         id = swap_cgroup_record(ent, 0);
2181         rcu_read_lock();
2182         memcg = mem_cgroup_lookup(id);
2183         if (memcg) {
2184                 /*
2185                  * We uncharge this because swap is freed.
2186                  * This memcg can be obsolete one. We avoid calling css_tryget
2187                  */
2188                 if (!mem_cgroup_is_root(memcg))
2189                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2190                 mem_cgroup_swap_statistics(memcg, false);
2191                 mem_cgroup_put(memcg);
2192         }
2193         rcu_read_unlock();
2194 }
2195 #endif
2196
2197 /*
2198  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2199  * page belongs to.
2200  */
2201 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2202 {
2203         struct page_cgroup *pc;
2204         struct mem_cgroup *mem = NULL;
2205         int ret = 0;
2206
2207         if (mem_cgroup_disabled())
2208                 return 0;
2209
2210         pc = lookup_page_cgroup(page);
2211         lock_page_cgroup(pc);
2212         if (PageCgroupUsed(pc)) {
2213                 mem = pc->mem_cgroup;
2214                 css_get(&mem->css);
2215         }
2216         unlock_page_cgroup(pc);
2217
2218         if (mem) {
2219                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
2220                                                 page);
2221                 css_put(&mem->css);
2222         }
2223         *ptr = mem;
2224         return ret;
2225 }
2226
2227 /* remove redundant charge if migration failed*/
2228 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2229                 struct page *oldpage, struct page *newpage)
2230 {
2231         struct page *target, *unused;
2232         struct page_cgroup *pc;
2233         enum charge_type ctype;
2234
2235         if (!mem)
2236                 return;
2237         cgroup_exclude_rmdir(&mem->css);
2238         /* at migration success, oldpage->mapping is NULL. */
2239         if (oldpage->mapping) {
2240                 target = oldpage;
2241                 unused = NULL;
2242         } else {
2243                 target = newpage;
2244                 unused = oldpage;
2245         }
2246
2247         if (PageAnon(target))
2248                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2249         else if (page_is_file_cache(target))
2250                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2251         else
2252                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2253
2254         /* unused page is not on radix-tree now. */
2255         if (unused)
2256                 __mem_cgroup_uncharge_common(unused, ctype);
2257
2258         pc = lookup_page_cgroup(target);
2259         /*
2260          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2261          * So, double-counting is effectively avoided.
2262          */
2263         __mem_cgroup_commit_charge(mem, pc, ctype);
2264
2265         /*
2266          * Both of oldpage and newpage are still under lock_page().
2267          * Then, we don't have to care about race in radix-tree.
2268          * But we have to be careful that this page is unmapped or not.
2269          *
2270          * There is a case for !page_mapped(). At the start of
2271          * migration, oldpage was mapped. But now, it's zapped.
2272          * But we know *target* page is not freed/reused under us.
2273          * mem_cgroup_uncharge_page() does all necessary checks.
2274          */
2275         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2276                 mem_cgroup_uncharge_page(target);
2277         /*
2278          * At migration, we may charge account against cgroup which has no tasks
2279          * So, rmdir()->pre_destroy() can be called while we do this charge.
2280          * In that case, we need to call pre_destroy() again. check it here.
2281          */
2282         cgroup_release_and_wakeup_rmdir(&mem->css);
2283 }
2284
2285 /*
2286  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2287  * Calling hierarchical_reclaim is not enough because we should update
2288  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2289  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2290  * not from the memcg which this page would be charged to.
2291  * try_charge_swapin does all of these works properly.
2292  */
2293 int mem_cgroup_shmem_charge_fallback(struct page *page,
2294                             struct mm_struct *mm,
2295                             gfp_t gfp_mask)
2296 {
2297         struct mem_cgroup *mem = NULL;
2298         int ret;
2299
2300         if (mem_cgroup_disabled())
2301                 return 0;
2302
2303         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2304         if (!ret)
2305                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2306
2307         return ret;
2308 }
2309
2310 static DEFINE_MUTEX(set_limit_mutex);
2311
2312 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2313                                 unsigned long long val)
2314 {
2315         int retry_count;
2316         u64 memswlimit;
2317         int ret = 0;
2318         int children = mem_cgroup_count_children(memcg);
2319         u64 curusage, oldusage;
2320
2321         /*
2322          * For keeping hierarchical_reclaim simple, how long we should retry
2323          * is depends on callers. We set our retry-count to be function
2324          * of # of children which we should visit in this loop.
2325          */
2326         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2327
2328         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2329
2330         while (retry_count) {
2331                 if (signal_pending(current)) {
2332                         ret = -EINTR;
2333                         break;
2334                 }
2335                 /*
2336                  * Rather than hide all in some function, I do this in
2337                  * open coded manner. You see what this really does.
2338                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2339                  */
2340                 mutex_lock(&set_limit_mutex);
2341                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2342                 if (memswlimit < val) {
2343                         ret = -EINVAL;
2344                         mutex_unlock(&set_limit_mutex);
2345                         break;
2346                 }
2347                 ret = res_counter_set_limit(&memcg->res, val);
2348                 if (!ret) {
2349                         if (memswlimit == val)
2350                                 memcg->memsw_is_minimum = true;
2351                         else
2352                                 memcg->memsw_is_minimum = false;
2353                 }
2354                 mutex_unlock(&set_limit_mutex);
2355
2356                 if (!ret)
2357                         break;
2358
2359                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2360                                                 MEM_CGROUP_RECLAIM_SHRINK);
2361                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2362                 /* Usage is reduced ? */
2363                 if (curusage >= oldusage)
2364                         retry_count--;
2365                 else
2366                         oldusage = curusage;
2367         }
2368
2369         return ret;
2370 }
2371
2372 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2373                                         unsigned long long val)
2374 {
2375         int retry_count;
2376         u64 memlimit, oldusage, curusage;
2377         int children = mem_cgroup_count_children(memcg);
2378         int ret = -EBUSY;
2379
2380         /* see mem_cgroup_resize_res_limit */
2381         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2382         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2383         while (retry_count) {
2384                 if (signal_pending(current)) {
2385                         ret = -EINTR;
2386                         break;
2387                 }
2388                 /*
2389                  * Rather than hide all in some function, I do this in
2390                  * open coded manner. You see what this really does.
2391                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2392                  */
2393                 mutex_lock(&set_limit_mutex);
2394                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2395                 if (memlimit > val) {
2396                         ret = -EINVAL;
2397                         mutex_unlock(&set_limit_mutex);
2398                         break;
2399                 }
2400                 ret = res_counter_set_limit(&memcg->memsw, val);
2401                 if (!ret) {
2402                         if (memlimit == val)
2403                                 memcg->memsw_is_minimum = true;
2404                         else
2405                                 memcg->memsw_is_minimum = false;
2406                 }
2407                 mutex_unlock(&set_limit_mutex);
2408
2409                 if (!ret)
2410                         break;
2411
2412                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2413                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2414                                                 MEM_CGROUP_RECLAIM_SHRINK);
2415                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2416                 /* Usage is reduced ? */
2417                 if (curusage >= oldusage)
2418                         retry_count--;
2419                 else
2420                         oldusage = curusage;
2421         }
2422         return ret;
2423 }
2424
2425 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2426                                                 gfp_t gfp_mask, int nid,
2427                                                 int zid)
2428 {
2429         unsigned long nr_reclaimed = 0;
2430         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2431         unsigned long reclaimed;
2432         int loop = 0;
2433         struct mem_cgroup_tree_per_zone *mctz;
2434         unsigned long long excess;
2435
2436         if (order > 0)
2437                 return 0;
2438
2439         mctz = soft_limit_tree_node_zone(nid, zid);
2440         /*
2441          * This loop can run a while, specially if mem_cgroup's continuously
2442          * keep exceeding their soft limit and putting the system under
2443          * pressure
2444          */
2445         do {
2446                 if (next_mz)
2447                         mz = next_mz;
2448                 else
2449                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2450                 if (!mz)
2451                         break;
2452
2453                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2454                                                 gfp_mask,
2455                                                 MEM_CGROUP_RECLAIM_SOFT);
2456                 nr_reclaimed += reclaimed;
2457                 spin_lock(&mctz->lock);
2458
2459                 /*
2460                  * If we failed to reclaim anything from this memory cgroup
2461                  * it is time to move on to the next cgroup
2462                  */
2463                 next_mz = NULL;
2464                 if (!reclaimed) {
2465                         do {
2466                                 /*
2467                                  * Loop until we find yet another one.
2468                                  *
2469                                  * By the time we get the soft_limit lock
2470                                  * again, someone might have aded the
2471                                  * group back on the RB tree. Iterate to
2472                                  * make sure we get a different mem.
2473                                  * mem_cgroup_largest_soft_limit_node returns
2474                                  * NULL if no other cgroup is present on
2475                                  * the tree
2476                                  */
2477                                 next_mz =
2478                                 __mem_cgroup_largest_soft_limit_node(mctz);
2479                                 if (next_mz == mz) {
2480                                         css_put(&next_mz->mem->css);
2481                                         next_mz = NULL;
2482                                 } else /* next_mz == NULL or other memcg */
2483                                         break;
2484                         } while (1);
2485                 }
2486                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2487                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2488                 /*
2489                  * One school of thought says that we should not add
2490                  * back the node to the tree if reclaim returns 0.
2491                  * But our reclaim could return 0, simply because due
2492                  * to priority we are exposing a smaller subset of
2493                  * memory to reclaim from. Consider this as a longer
2494                  * term TODO.
2495                  */
2496                 /* If excess == 0, no tree ops */
2497                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2498                 spin_unlock(&mctz->lock);
2499                 css_put(&mz->mem->css);
2500                 loop++;
2501                 /*
2502                  * Could not reclaim anything and there are no more
2503                  * mem cgroups to try or we seem to be looping without
2504                  * reclaiming anything.
2505                  */
2506                 if (!nr_reclaimed &&
2507                         (next_mz == NULL ||
2508                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2509                         break;
2510         } while (!nr_reclaimed);
2511         if (next_mz)
2512                 css_put(&next_mz->mem->css);
2513         return nr_reclaimed;
2514 }
2515
2516 /*
2517  * This routine traverse page_cgroup in given list and drop them all.
2518  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2519  */
2520 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2521                                 int node, int zid, enum lru_list lru)
2522 {
2523         struct zone *zone;
2524         struct mem_cgroup_per_zone *mz;
2525         struct page_cgroup *pc, *busy;
2526         unsigned long flags, loop;
2527         struct list_head *list;
2528         int ret = 0;
2529
2530         zone = &NODE_DATA(node)->node_zones[zid];
2531         mz = mem_cgroup_zoneinfo(mem, node, zid);
2532         list = &mz->lists[lru];
2533
2534         loop = MEM_CGROUP_ZSTAT(mz, lru);
2535         /* give some margin against EBUSY etc...*/
2536         loop += 256;
2537         busy = NULL;
2538         while (loop--) {
2539                 ret = 0;
2540                 spin_lock_irqsave(&zone->lru_lock, flags);
2541                 if (list_empty(list)) {
2542                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2543                         break;
2544                 }
2545                 pc = list_entry(list->prev, struct page_cgroup, lru);
2546                 if (busy == pc) {
2547                         list_move(&pc->lru, list);
2548                         busy = NULL;
2549                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2550                         continue;
2551                 }
2552                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2553
2554                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2555                 if (ret == -ENOMEM)
2556                         break;
2557
2558                 if (ret == -EBUSY || ret == -EINVAL) {
2559                         /* found lock contention or "pc" is obsolete. */
2560                         busy = pc;
2561                         cond_resched();
2562                 } else
2563                         busy = NULL;
2564         }
2565
2566         if (!ret && !list_empty(list))
2567                 return -EBUSY;
2568         return ret;
2569 }
2570
2571 /*
2572  * make mem_cgroup's charge to be 0 if there is no task.
2573  * This enables deleting this mem_cgroup.
2574  */
2575 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2576 {
2577         int ret;
2578         int node, zid, shrink;
2579         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2580         struct cgroup *cgrp = mem->css.cgroup;
2581
2582         css_get(&mem->css);
2583
2584         shrink = 0;
2585         /* should free all ? */
2586         if (free_all)
2587                 goto try_to_free;
2588 move_account:
2589         do {
2590                 ret = -EBUSY;
2591                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2592                         goto out;
2593                 ret = -EINTR;
2594                 if (signal_pending(current))
2595                         goto out;
2596                 /* This is for making all *used* pages to be on LRU. */
2597                 lru_add_drain_all();
2598                 drain_all_stock_sync();
2599                 ret = 0;
2600                 for_each_node_state(node, N_HIGH_MEMORY) {
2601                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2602                                 enum lru_list l;
2603                                 for_each_lru(l) {
2604                                         ret = mem_cgroup_force_empty_list(mem,
2605                                                         node, zid, l);
2606                                         if (ret)
2607                                                 break;
2608                                 }
2609                         }
2610                         if (ret)
2611                                 break;
2612                 }
2613                 /* it seems parent cgroup doesn't have enough mem */
2614                 if (ret == -ENOMEM)
2615                         goto try_to_free;
2616                 cond_resched();
2617         /* "ret" should also be checked to ensure all lists are empty. */
2618         } while (mem->res.usage > 0 || ret);
2619 out:
2620         css_put(&mem->css);
2621         return ret;
2622
2623 try_to_free:
2624         /* returns EBUSY if there is a task or if we come here twice. */
2625         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2626                 ret = -EBUSY;
2627                 goto out;
2628         }
2629         /* we call try-to-free pages for make this cgroup empty */
2630         lru_add_drain_all();
2631         /* try to free all pages in this cgroup */
2632         shrink = 1;
2633         while (nr_retries && mem->res.usage > 0) {
2634                 int progress;
2635
2636                 if (signal_pending(current)) {
2637                         ret = -EINTR;
2638                         goto out;
2639                 }
2640                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2641                                                 false, get_swappiness(mem));
2642                 if (!progress) {
2643                         nr_retries--;
2644                         /* maybe some writeback is necessary */
2645                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2646                 }
2647
2648         }
2649         lru_add_drain();
2650         /* try move_account...there may be some *locked* pages. */
2651         goto move_account;
2652 }
2653
2654 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2655 {
2656         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2657 }
2658
2659
2660 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2661 {
2662         return mem_cgroup_from_cont(cont)->use_hierarchy;
2663 }
2664
2665 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2666                                         u64 val)
2667 {
2668         int retval = 0;
2669         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2670         struct cgroup *parent = cont->parent;
2671         struct mem_cgroup *parent_mem = NULL;
2672
2673         if (parent)
2674                 parent_mem = mem_cgroup_from_cont(parent);
2675
2676         cgroup_lock();
2677         /*
2678          * If parent's use_hierarchy is set, we can't make any modifications
2679          * in the child subtrees. If it is unset, then the change can
2680          * occur, provided the current cgroup has no children.
2681          *
2682          * For the root cgroup, parent_mem is NULL, we allow value to be
2683          * set if there are no children.
2684          */
2685         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2686                                 (val == 1 || val == 0)) {
2687                 if (list_empty(&cont->children))
2688                         mem->use_hierarchy = val;
2689                 else
2690                         retval = -EBUSY;
2691         } else
2692                 retval = -EINVAL;
2693         cgroup_unlock();
2694
2695         return retval;
2696 }
2697
2698 struct mem_cgroup_idx_data {
2699         s64 val;
2700         enum mem_cgroup_stat_index idx;
2701 };
2702
2703 static int
2704 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2705 {
2706         struct mem_cgroup_idx_data *d = data;
2707         d->val += mem_cgroup_read_stat(&mem->stat, d->idx);
2708         return 0;
2709 }
2710
2711 static void
2712 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2713                                 enum mem_cgroup_stat_index idx, s64 *val)
2714 {
2715         struct mem_cgroup_idx_data d;
2716         d.idx = idx;
2717         d.val = 0;
2718         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2719         *val = d.val;
2720 }
2721
2722 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2723 {
2724         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2725         u64 idx_val, val;
2726         int type, name;
2727
2728         type = MEMFILE_TYPE(cft->private);
2729         name = MEMFILE_ATTR(cft->private);
2730         switch (type) {
2731         case _MEM:
2732                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2733                         mem_cgroup_get_recursive_idx_stat(mem,
2734                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2735                         val = idx_val;
2736                         mem_cgroup_get_recursive_idx_stat(mem,
2737                                 MEM_CGROUP_STAT_RSS, &idx_val);
2738                         val += idx_val;
2739                         val <<= PAGE_SHIFT;
2740                 } else
2741                         val = res_counter_read_u64(&mem->res, name);
2742                 break;
2743         case _MEMSWAP:
2744                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2745                         mem_cgroup_get_recursive_idx_stat(mem,
2746                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2747                         val = idx_val;
2748                         mem_cgroup_get_recursive_idx_stat(mem,
2749                                 MEM_CGROUP_STAT_RSS, &idx_val);
2750                         val += idx_val;
2751                         mem_cgroup_get_recursive_idx_stat(mem,
2752                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2753                         val += idx_val;
2754                         val <<= PAGE_SHIFT;
2755                 } else
2756                         val = res_counter_read_u64(&mem->memsw, name);
2757                 break;
2758         default:
2759                 BUG();
2760                 break;
2761         }
2762         return val;
2763 }
2764 /*
2765  * The user of this function is...
2766  * RES_LIMIT.
2767  */
2768 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2769                             const char *buffer)
2770 {
2771         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2772         int type, name;
2773         unsigned long long val;
2774         int ret;
2775
2776         type = MEMFILE_TYPE(cft->private);
2777         name = MEMFILE_ATTR(cft->private);
2778         switch (name) {
2779         case RES_LIMIT:
2780                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2781                         ret = -EINVAL;
2782                         break;
2783                 }
2784                 /* This function does all necessary parse...reuse it */
2785                 ret = res_counter_memparse_write_strategy(buffer, &val);
2786                 if (ret)
2787                         break;
2788                 if (type == _MEM)
2789                         ret = mem_cgroup_resize_limit(memcg, val);
2790                 else
2791                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2792                 break;
2793         case RES_SOFT_LIMIT:
2794                 ret = res_counter_memparse_write_strategy(buffer, &val);
2795                 if (ret)
2796                         break;
2797                 /*
2798                  * For memsw, soft limits are hard to implement in terms
2799                  * of semantics, for now, we support soft limits for
2800                  * control without swap
2801                  */
2802                 if (type == _MEM)
2803                         ret = res_counter_set_soft_limit(&memcg->res, val);
2804                 else
2805                         ret = -EINVAL;
2806                 break;
2807         default:
2808                 ret = -EINVAL; /* should be BUG() ? */
2809                 break;
2810         }
2811         return ret;
2812 }
2813
2814 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2815                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2816 {
2817         struct cgroup *cgroup;
2818         unsigned long long min_limit, min_memsw_limit, tmp;
2819
2820         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2821         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2822         cgroup = memcg->css.cgroup;
2823         if (!memcg->use_hierarchy)
2824                 goto out;
2825
2826         while (cgroup->parent) {
2827                 cgroup = cgroup->parent;
2828                 memcg = mem_cgroup_from_cont(cgroup);
2829                 if (!memcg->use_hierarchy)
2830                         break;
2831                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2832                 min_limit = min(min_limit, tmp);
2833                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2834                 min_memsw_limit = min(min_memsw_limit, tmp);
2835         }
2836 out:
2837         *mem_limit = min_limit;
2838         *memsw_limit = min_memsw_limit;
2839         return;
2840 }
2841
2842 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2843 {
2844         struct mem_cgroup *mem;
2845         int type, name;
2846
2847         mem = mem_cgroup_from_cont(cont);
2848         type = MEMFILE_TYPE(event);
2849         name = MEMFILE_ATTR(event);
2850         switch (name) {
2851         case RES_MAX_USAGE:
2852                 if (type == _MEM)
2853                         res_counter_reset_max(&mem->res);
2854                 else
2855                         res_counter_reset_max(&mem->memsw);
2856                 break;
2857         case RES_FAILCNT:
2858                 if (type == _MEM)
2859                         res_counter_reset_failcnt(&mem->res);
2860                 else
2861                         res_counter_reset_failcnt(&mem->memsw);
2862                 break;
2863         }
2864
2865         return 0;
2866 }
2867
2868
2869 /* For read statistics */
2870 enum {
2871         MCS_CACHE,
2872         MCS_RSS,
2873         MCS_FILE_MAPPED,
2874         MCS_PGPGIN,
2875         MCS_PGPGOUT,
2876         MCS_SWAP,
2877         MCS_INACTIVE_ANON,
2878         MCS_ACTIVE_ANON,
2879         MCS_INACTIVE_FILE,
2880         MCS_ACTIVE_FILE,
2881         MCS_UNEVICTABLE,
2882         NR_MCS_STAT,
2883 };
2884
2885 struct mcs_total_stat {
2886         s64 stat[NR_MCS_STAT];
2887 };
2888
2889 struct {
2890         char *local_name;
2891         char *total_name;
2892 } memcg_stat_strings[NR_MCS_STAT] = {
2893         {"cache", "total_cache"},
2894         {"rss", "total_rss"},
2895         {"mapped_file", "total_mapped_file"},
2896         {"pgpgin", "total_pgpgin"},
2897         {"pgpgout", "total_pgpgout"},
2898         {"swap", "total_swap"},
2899         {"inactive_anon", "total_inactive_anon"},
2900         {"active_anon", "total_active_anon"},
2901         {"inactive_file", "total_inactive_file"},
2902         {"active_file", "total_active_file"},
2903         {"unevictable", "total_unevictable"}
2904 };
2905
2906
2907 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2908 {
2909         struct mcs_total_stat *s = data;
2910         s64 val;
2911
2912         /* per cpu stat */
2913         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2914         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2915         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2916         s->stat[MCS_RSS] += val * PAGE_SIZE;
2917         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_FILE_MAPPED);
2918         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
2919         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2920         s->stat[MCS_PGPGIN] += val;
2921         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2922         s->stat[MCS_PGPGOUT] += val;
2923         if (do_swap_account) {
2924                 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT);
2925                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
2926         }
2927
2928         /* per zone stat */
2929         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2930         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2931         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2932         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2933         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2934         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2935         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2936         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2937         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2938         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2939         return 0;
2940 }
2941
2942 static void
2943 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2944 {
2945         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2946 }
2947
2948 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2949                                  struct cgroup_map_cb *cb)
2950 {
2951         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2952         struct mcs_total_stat mystat;
2953         int i;
2954
2955         memset(&mystat, 0, sizeof(mystat));
2956         mem_cgroup_get_local_stat(mem_cont, &mystat);
2957
2958         for (i = 0; i < NR_MCS_STAT; i++) {
2959                 if (i == MCS_SWAP && !do_swap_account)
2960                         continue;
2961                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2962         }
2963
2964         /* Hierarchical information */
2965         {
2966                 unsigned long long limit, memsw_limit;
2967                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2968                 cb->fill(cb, "hierarchical_memory_limit", limit);
2969                 if (do_swap_account)
2970                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2971         }
2972
2973         memset(&mystat, 0, sizeof(mystat));
2974         mem_cgroup_get_total_stat(mem_cont, &mystat);
2975         for (i = 0; i < NR_MCS_STAT; i++) {
2976                 if (i == MCS_SWAP && !do_swap_account)
2977                         continue;
2978                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2979         }
2980
2981 #ifdef CONFIG_DEBUG_VM
2982         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2983
2984         {
2985                 int nid, zid;
2986                 struct mem_cgroup_per_zone *mz;
2987                 unsigned long recent_rotated[2] = {0, 0};
2988                 unsigned long recent_scanned[2] = {0, 0};
2989
2990                 for_each_online_node(nid)
2991                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2992                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2993
2994                                 recent_rotated[0] +=
2995                                         mz->reclaim_stat.recent_rotated[0];
2996                                 recent_rotated[1] +=
2997                                         mz->reclaim_stat.recent_rotated[1];
2998                                 recent_scanned[0] +=
2999                                         mz->reclaim_stat.recent_scanned[0];
3000                                 recent_scanned[1] +=
3001                                         mz->reclaim_stat.recent_scanned[1];
3002                         }
3003                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3004                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3005                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3006                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3007         }
3008 #endif
3009
3010         return 0;
3011 }
3012
3013 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3014 {
3015         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3016
3017         return get_swappiness(memcg);
3018 }
3019
3020 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3021                                        u64 val)
3022 {
3023         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3024         struct mem_cgroup *parent;
3025
3026         if (val > 100)
3027                 return -EINVAL;
3028
3029         if (cgrp->parent == NULL)
3030                 return -EINVAL;
3031
3032         parent = mem_cgroup_from_cont(cgrp->parent);
3033
3034         cgroup_lock();
3035
3036         /* If under hierarchy, only empty-root can set this value */
3037         if ((parent->use_hierarchy) ||
3038             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3039                 cgroup_unlock();
3040                 return -EINVAL;
3041         }
3042
3043         spin_lock(&memcg->reclaim_param_lock);
3044         memcg->swappiness = val;
3045         spin_unlock(&memcg->reclaim_param_lock);
3046
3047         cgroup_unlock();
3048
3049         return 0;
3050 }
3051
3052
3053 static struct cftype mem_cgroup_files[] = {
3054         {
3055                 .name = "usage_in_bytes",
3056                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3057                 .read_u64 = mem_cgroup_read,
3058         },
3059         {
3060                 .name = "max_usage_in_bytes",
3061                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3062                 .trigger = mem_cgroup_reset,
3063                 .read_u64 = mem_cgroup_read,
3064         },
3065         {
3066                 .name = "limit_in_bytes",
3067                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3068                 .write_string = mem_cgroup_write,
3069                 .read_u64 = mem_cgroup_read,
3070         },
3071         {
3072                 .name = "soft_limit_in_bytes",
3073                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3074                 .write_string = mem_cgroup_write,
3075                 .read_u64 = mem_cgroup_read,
3076         },
3077         {
3078                 .name = "failcnt",
3079                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3080                 .trigger = mem_cgroup_reset,
3081                 .read_u64 = mem_cgroup_read,
3082         },
3083         {
3084                 .name = "stat",
3085                 .read_map = mem_control_stat_show,
3086         },
3087         {
3088                 .name = "force_empty",
3089                 .trigger = mem_cgroup_force_empty_write,
3090         },
3091         {
3092                 .name = "use_hierarchy",
3093                 .write_u64 = mem_cgroup_hierarchy_write,
3094                 .read_u64 = mem_cgroup_hierarchy_read,
3095         },
3096         {
3097                 .name = "swappiness",
3098                 .read_u64 = mem_cgroup_swappiness_read,
3099                 .write_u64 = mem_cgroup_swappiness_write,
3100         },
3101 };
3102
3103 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3104 static struct cftype memsw_cgroup_files[] = {
3105         {
3106                 .name = "memsw.usage_in_bytes",
3107                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3108                 .read_u64 = mem_cgroup_read,
3109         },
3110         {
3111                 .name = "memsw.max_usage_in_bytes",
3112                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3113                 .trigger = mem_cgroup_reset,
3114                 .read_u64 = mem_cgroup_read,
3115         },
3116         {
3117                 .name = "memsw.limit_in_bytes",
3118                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3119                 .write_string = mem_cgroup_write,
3120                 .read_u64 = mem_cgroup_read,
3121         },
3122         {
3123                 .name = "memsw.failcnt",
3124                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3125                 .trigger = mem_cgroup_reset,
3126                 .read_u64 = mem_cgroup_read,
3127         },
3128 };
3129
3130 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3131 {
3132         if (!do_swap_account)
3133                 return 0;
3134         return cgroup_add_files(cont, ss, memsw_cgroup_files,
3135                                 ARRAY_SIZE(memsw_cgroup_files));
3136 };
3137 #else
3138 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3139 {
3140         return 0;
3141 }
3142 #endif
3143
3144 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3145 {
3146         struct mem_cgroup_per_node *pn;
3147         struct mem_cgroup_per_zone *mz;
3148         enum lru_list l;
3149         int zone, tmp = node;
3150         /*
3151          * This routine is called against possible nodes.
3152          * But it's BUG to call kmalloc() against offline node.
3153          *
3154          * TODO: this routine can waste much memory for nodes which will
3155          *       never be onlined. It's better to use memory hotplug callback
3156          *       function.
3157          */
3158         if (!node_state(node, N_NORMAL_MEMORY))
3159                 tmp = -1;
3160         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3161         if (!pn)
3162                 return 1;
3163
3164         mem->info.nodeinfo[node] = pn;
3165         memset(pn, 0, sizeof(*pn));
3166
3167         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3168                 mz = &pn->zoneinfo[zone];
3169                 for_each_lru(l)
3170                         INIT_LIST_HEAD(&mz->lists[l]);
3171                 mz->usage_in_excess = 0;
3172                 mz->on_tree = false;
3173                 mz->mem = mem;
3174         }
3175         return 0;
3176 }
3177
3178 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3179 {
3180         kfree(mem->info.nodeinfo[node]);
3181 }
3182
3183 static int mem_cgroup_size(void)
3184 {
3185         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
3186         return sizeof(struct mem_cgroup) + cpustat_size;
3187 }
3188
3189 static struct mem_cgroup *mem_cgroup_alloc(void)
3190 {
3191         struct mem_cgroup *mem;
3192         int size = mem_cgroup_size();
3193
3194         if (size < PAGE_SIZE)
3195                 mem = kmalloc(size, GFP_KERNEL);
3196         else
3197                 mem = vmalloc(size);
3198
3199         if (mem)
3200                 memset(mem, 0, size);
3201         return mem;
3202 }
3203
3204 /*
3205  * At destroying mem_cgroup, references from swap_cgroup can remain.
3206  * (scanning all at force_empty is too costly...)
3207  *
3208  * Instead of clearing all references at force_empty, we remember
3209  * the number of reference from swap_cgroup and free mem_cgroup when
3210  * it goes down to 0.
3211  *
3212  * Removal of cgroup itself succeeds regardless of refs from swap.
3213  */
3214
3215 static void __mem_cgroup_free(struct mem_cgroup *mem)
3216 {
3217         int node;
3218
3219         mem_cgroup_remove_from_trees(mem);
3220         free_css_id(&mem_cgroup_subsys, &mem->css);
3221
3222         for_each_node_state(node, N_POSSIBLE)
3223                 free_mem_cgroup_per_zone_info(mem, node);
3224
3225         if (mem_cgroup_size() < PAGE_SIZE)
3226                 kfree(mem);
3227         else
3228                 vfree(mem);
3229 }
3230
3231 static void mem_cgroup_get(struct mem_cgroup *mem)
3232 {
3233         atomic_inc(&mem->refcnt);
3234 }
3235
3236 static void mem_cgroup_put(struct mem_cgroup *mem)
3237 {
3238         if (atomic_dec_and_test(&mem->refcnt)) {
3239                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3240                 __mem_cgroup_free(mem);
3241                 if (parent)
3242                         mem_cgroup_put(parent);
3243         }
3244 }
3245
3246 /*
3247  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3248  */
3249 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3250 {
3251         if (!mem->res.parent)
3252                 return NULL;
3253         return mem_cgroup_from_res_counter(mem->res.parent, res);
3254 }
3255
3256 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3257 static void __init enable_swap_cgroup(void)
3258 {
3259         if (!mem_cgroup_disabled() && really_do_swap_account)
3260                 do_swap_account = 1;
3261 }
3262 #else
3263 static void __init enable_swap_cgroup(void)
3264 {
3265 }
3266 #endif
3267
3268 static int mem_cgroup_soft_limit_tree_init(void)
3269 {
3270         struct mem_cgroup_tree_per_node *rtpn;
3271         struct mem_cgroup_tree_per_zone *rtpz;
3272         int tmp, node, zone;
3273
3274         for_each_node_state(node, N_POSSIBLE) {
3275                 tmp = node;
3276                 if (!node_state(node, N_NORMAL_MEMORY))
3277                         tmp = -1;
3278                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3279                 if (!rtpn)
3280                         return 1;
3281
3282                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3283
3284                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3285                         rtpz = &rtpn->rb_tree_per_zone[zone];
3286                         rtpz->rb_root = RB_ROOT;
3287                         spin_lock_init(&rtpz->lock);
3288                 }
3289         }
3290         return 0;
3291 }
3292
3293 static struct cgroup_subsys_state * __ref
3294 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3295 {
3296         struct mem_cgroup *mem, *parent;
3297         long error = -ENOMEM;
3298         int node;
3299
3300         mem = mem_cgroup_alloc();
3301         if (!mem)
3302                 return ERR_PTR(error);
3303
3304         for_each_node_state(node, N_POSSIBLE)
3305                 if (alloc_mem_cgroup_per_zone_info(mem, node))
3306                         goto free_out;
3307
3308         /* root ? */
3309         if (cont->parent == NULL) {
3310                 int cpu;
3311                 enable_swap_cgroup();
3312                 parent = NULL;
3313                 root_mem_cgroup = mem;
3314                 if (mem_cgroup_soft_limit_tree_init())
3315                         goto free_out;
3316                 for_each_possible_cpu(cpu) {
3317                         struct memcg_stock_pcp *stock =
3318                                                 &per_cpu(memcg_stock, cpu);
3319                         INIT_WORK(&stock->work, drain_local_stock);
3320                 }
3321                 hotcpu_notifier(memcg_stock_cpu_callback, 0);
3322
3323         } else {
3324                 parent = mem_cgroup_from_cont(cont->parent);
3325                 mem->use_hierarchy = parent->use_hierarchy;
3326         }
3327
3328         if (parent && parent->use_hierarchy) {
3329                 res_counter_init(&mem->res, &parent->res);
3330                 res_counter_init(&mem->memsw, &parent->memsw);
3331                 /*
3332                  * We increment refcnt of the parent to ensure that we can
3333                  * safely access it on res_counter_charge/uncharge.
3334                  * This refcnt will be decremented when freeing this
3335                  * mem_cgroup(see mem_cgroup_put).
3336                  */
3337                 mem_cgroup_get(parent);
3338         } else {
3339                 res_counter_init(&mem->res, NULL);
3340                 res_counter_init(&mem->memsw, NULL);
3341         }
3342         mem->last_scanned_child = 0;
3343         spin_lock_init(&mem->reclaim_param_lock);
3344
3345         if (parent)
3346                 mem->swappiness = get_swappiness(parent);
3347         atomic_set(&mem->refcnt, 1);
3348         return &mem->css;
3349 free_out:
3350         __mem_cgroup_free(mem);
3351         root_mem_cgroup = NULL;
3352         return ERR_PTR(error);
3353 }
3354
3355 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
3356                                         struct cgroup *cont)
3357 {
3358         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3359
3360         return mem_cgroup_force_empty(mem, false);
3361 }
3362
3363 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
3364                                 struct cgroup *cont)
3365 {
3366         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3367
3368         mem_cgroup_put(mem);
3369 }
3370
3371 static int mem_cgroup_populate(struct cgroup_subsys *ss,
3372                                 struct cgroup *cont)
3373 {
3374         int ret;
3375
3376         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
3377                                 ARRAY_SIZE(mem_cgroup_files));
3378
3379         if (!ret)
3380                 ret = register_memsw_files(cont, ss);
3381         return ret;
3382 }
3383
3384 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
3385                                 struct cgroup *cont,
3386                                 struct cgroup *old_cont,
3387                                 struct task_struct *p,
3388                                 bool threadgroup)
3389 {
3390         /*
3391          * FIXME: It's better to move charges of this process from old
3392          * memcg to new memcg. But it's just on TODO-List now.
3393          */
3394 }
3395
3396 struct cgroup_subsys mem_cgroup_subsys = {
3397         .name = "memory",
3398         .subsys_id = mem_cgroup_subsys_id,
3399         .create = mem_cgroup_create,
3400         .pre_destroy = mem_cgroup_pre_destroy,
3401         .destroy = mem_cgroup_destroy,
3402         .populate = mem_cgroup_populate,
3403         .attach = mem_cgroup_move_task,
3404         .early_init = 0,
3405         .use_id = 1,
3406 };
3407
3408 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3409
3410 static int __init disable_swap_account(char *s)
3411 {
3412         really_do_swap_account = 0;
3413         return 1;
3414 }
3415 __setup("noswapaccount", disable_swap_account);
3416 #endif