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