memcg: make memcg's file mapped consistent with global VM
[sfrench/cifs-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  */
19
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/limits.h>
31 #include <linux/mutex.h>
32 #include <linux/rbtree.h>
33 #include <linux/slab.h>
34 #include <linux/swap.h>
35 #include <linux/spinlock.h>
36 #include <linux/fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mm_inline.h>
40 #include <linux/page_cgroup.h>
41 #include <linux/cpu.h>
42 #include "internal.h"
43
44 #include <asm/uaccess.h>
45
46 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
47 #define MEM_CGROUP_RECLAIM_RETRIES      5
48 struct mem_cgroup *root_mem_cgroup __read_mostly;
49
50 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
51 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
52 int do_swap_account __read_mostly;
53 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
54 #else
55 #define do_swap_account         (0)
56 #endif
57
58 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
59 #define SOFTLIMIT_EVENTS_THRESH (1000)
60
61 /*
62  * Statistics for memory cgroup.
63  */
64 enum mem_cgroup_stat_index {
65         /*
66          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
67          */
68         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
69         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
70         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
71         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
72         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
73         MEM_CGROUP_STAT_EVENTS, /* sum of pagein + pageout for internal use */
74         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
75
76         MEM_CGROUP_STAT_NSTATS,
77 };
78
79 struct mem_cgroup_stat_cpu {
80         s64 count[MEM_CGROUP_STAT_NSTATS];
81 } ____cacheline_aligned_in_smp;
82
83 struct mem_cgroup_stat {
84         struct mem_cgroup_stat_cpu cpustat[0];
85 };
86
87 static inline void
88 __mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
89                                 enum mem_cgroup_stat_index idx)
90 {
91         stat->count[idx] = 0;
92 }
93
94 static inline s64
95 __mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
96                                 enum mem_cgroup_stat_index idx)
97 {
98         return stat->count[idx];
99 }
100
101 /*
102  * For accounting under irq disable, no need for increment preempt count.
103  */
104 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
105                 enum mem_cgroup_stat_index idx, int val)
106 {
107         stat->count[idx] += val;
108 }
109
110 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
111                 enum mem_cgroup_stat_index idx)
112 {
113         int cpu;
114         s64 ret = 0;
115         for_each_possible_cpu(cpu)
116                 ret += stat->cpustat[cpu].count[idx];
117         return ret;
118 }
119
120 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
121 {
122         s64 ret;
123
124         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
125         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
126         return ret;
127 }
128
129 /*
130  * per-zone information in memory controller.
131  */
132 struct mem_cgroup_per_zone {
133         /*
134          * spin_lock to protect the per cgroup LRU
135          */
136         struct list_head        lists[NR_LRU_LISTS];
137         unsigned long           count[NR_LRU_LISTS];
138
139         struct zone_reclaim_stat reclaim_stat;
140         struct rb_node          tree_node;      /* RB tree node */
141         unsigned long long      usage_in_excess;/* Set to the value by which */
142                                                 /* the soft limit is exceeded*/
143         bool                    on_tree;
144         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
145                                                 /* use container_of        */
146 };
147 /* Macro for accessing counter */
148 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
149
150 struct mem_cgroup_per_node {
151         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
152 };
153
154 struct mem_cgroup_lru_info {
155         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
156 };
157
158 /*
159  * Cgroups above their limits are maintained in a RB-Tree, independent of
160  * their hierarchy representation
161  */
162
163 struct mem_cgroup_tree_per_zone {
164         struct rb_root rb_root;
165         spinlock_t lock;
166 };
167
168 struct mem_cgroup_tree_per_node {
169         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
170 };
171
172 struct mem_cgroup_tree {
173         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
174 };
175
176 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
177
178 /*
179  * The memory controller data structure. The memory controller controls both
180  * page cache and RSS per cgroup. We would eventually like to provide
181  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
182  * to help the administrator determine what knobs to tune.
183  *
184  * TODO: Add a water mark for the memory controller. Reclaim will begin when
185  * we hit the water mark. May be even add a low water mark, such that
186  * no reclaim occurs from a cgroup at it's low water mark, this is
187  * a feature that will be implemented much later in the future.
188  */
189 struct mem_cgroup {
190         struct cgroup_subsys_state css;
191         /*
192          * the counter to account for memory usage
193          */
194         struct res_counter res;
195         /*
196          * the counter to account for mem+swap usage.
197          */
198         struct res_counter memsw;
199         /*
200          * Per cgroup active and inactive list, similar to the
201          * per zone LRU lists.
202          */
203         struct mem_cgroup_lru_info info;
204
205         /*
206           protect against reclaim related member.
207         */
208         spinlock_t reclaim_param_lock;
209
210         int     prev_priority;  /* for recording reclaim priority */
211
212         /*
213          * While reclaiming in a hierarchy, we cache the last child we
214          * reclaimed from.
215          */
216         int last_scanned_child;
217         /*
218          * Should the accounting and control be hierarchical, per subtree?
219          */
220         bool use_hierarchy;
221         unsigned long   last_oom_jiffies;
222         atomic_t        refcnt;
223
224         unsigned int    swappiness;
225
226         /* set when res.limit == memsw.limit */
227         bool            memsw_is_minimum;
228
229         /*
230          * statistics. This must be placed at the end of memcg.
231          */
232         struct mem_cgroup_stat stat;
233 };
234
235 /*
236  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
237  * limit reclaim to prevent infinite loops, if they ever occur.
238  */
239 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
240 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
241
242 enum charge_type {
243         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
244         MEM_CGROUP_CHARGE_TYPE_MAPPED,
245         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
246         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
247         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
248         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
249         NR_CHARGE_TYPE,
250 };
251
252 /* only for here (for easy reading.) */
253 #define PCGF_CACHE      (1UL << PCG_CACHE)
254 #define PCGF_USED       (1UL << PCG_USED)
255 #define PCGF_LOCK       (1UL << PCG_LOCK)
256 /* Not used, but added here for completeness */
257 #define PCGF_ACCT       (1UL << PCG_ACCT)
258
259 /* for encoding cft->private value on file */
260 #define _MEM                    (0)
261 #define _MEMSWAP                (1)
262 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
263 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
264 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
265
266 /*
267  * Reclaim flags for mem_cgroup_hierarchical_reclaim
268  */
269 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
270 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
271 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
272 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
273 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
274 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
275
276 static void mem_cgroup_get(struct mem_cgroup *mem);
277 static void mem_cgroup_put(struct mem_cgroup *mem);
278 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
279 static void drain_all_stock_async(void);
280
281 static struct mem_cgroup_per_zone *
282 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
283 {
284         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
285 }
286
287 static struct mem_cgroup_per_zone *
288 page_cgroup_zoneinfo(struct page_cgroup *pc)
289 {
290         struct mem_cgroup *mem = pc->mem_cgroup;
291         int nid = page_cgroup_nid(pc);
292         int zid = page_cgroup_zid(pc);
293
294         if (!mem)
295                 return NULL;
296
297         return mem_cgroup_zoneinfo(mem, nid, zid);
298 }
299
300 static struct mem_cgroup_tree_per_zone *
301 soft_limit_tree_node_zone(int nid, int zid)
302 {
303         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
304 }
305
306 static struct mem_cgroup_tree_per_zone *
307 soft_limit_tree_from_page(struct page *page)
308 {
309         int nid = page_to_nid(page);
310         int zid = page_zonenum(page);
311
312         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
313 }
314
315 static void
316 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
317                                 struct mem_cgroup_per_zone *mz,
318                                 struct mem_cgroup_tree_per_zone *mctz,
319                                 unsigned long long new_usage_in_excess)
320 {
321         struct rb_node **p = &mctz->rb_root.rb_node;
322         struct rb_node *parent = NULL;
323         struct mem_cgroup_per_zone *mz_node;
324
325         if (mz->on_tree)
326                 return;
327
328         mz->usage_in_excess = new_usage_in_excess;
329         if (!mz->usage_in_excess)
330                 return;
331         while (*p) {
332                 parent = *p;
333                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
334                                         tree_node);
335                 if (mz->usage_in_excess < mz_node->usage_in_excess)
336                         p = &(*p)->rb_left;
337                 /*
338                  * We can't avoid mem cgroups that are over their soft
339                  * limit by the same amount
340                  */
341                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
342                         p = &(*p)->rb_right;
343         }
344         rb_link_node(&mz->tree_node, parent, p);
345         rb_insert_color(&mz->tree_node, &mctz->rb_root);
346         mz->on_tree = true;
347 }
348
349 static void
350 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
351                                 struct mem_cgroup_per_zone *mz,
352                                 struct mem_cgroup_tree_per_zone *mctz)
353 {
354         if (!mz->on_tree)
355                 return;
356         rb_erase(&mz->tree_node, &mctz->rb_root);
357         mz->on_tree = false;
358 }
359
360 static void
361 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
362                                 struct mem_cgroup_per_zone *mz,
363                                 struct mem_cgroup_tree_per_zone *mctz)
364 {
365         spin_lock(&mctz->lock);
366         __mem_cgroup_remove_exceeded(mem, mz, mctz);
367         spin_unlock(&mctz->lock);
368 }
369
370 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
371 {
372         bool ret = false;
373         int cpu;
374         s64 val;
375         struct mem_cgroup_stat_cpu *cpustat;
376
377         cpu = get_cpu();
378         cpustat = &mem->stat.cpustat[cpu];
379         val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
380         if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
381                 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
382                 ret = true;
383         }
384         put_cpu();
385         return ret;
386 }
387
388 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
389 {
390         unsigned long long excess;
391         struct mem_cgroup_per_zone *mz;
392         struct mem_cgroup_tree_per_zone *mctz;
393         int nid = page_to_nid(page);
394         int zid = page_zonenum(page);
395         mctz = soft_limit_tree_from_page(page);
396
397         /*
398          * Necessary to update all ancestors when hierarchy is used.
399          * because their event counter is not touched.
400          */
401         for (; mem; mem = parent_mem_cgroup(mem)) {
402                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
403                 excess = res_counter_soft_limit_excess(&mem->res);
404                 /*
405                  * We have to update the tree if mz is on RB-tree or
406                  * mem is over its softlimit.
407                  */
408                 if (excess || mz->on_tree) {
409                         spin_lock(&mctz->lock);
410                         /* if on-tree, remove it */
411                         if (mz->on_tree)
412                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
413                         /*
414                          * Insert again. mz->usage_in_excess will be updated.
415                          * If excess is 0, no tree ops.
416                          */
417                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
418                         spin_unlock(&mctz->lock);
419                 }
420         }
421 }
422
423 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
424 {
425         int node, zone;
426         struct mem_cgroup_per_zone *mz;
427         struct mem_cgroup_tree_per_zone *mctz;
428
429         for_each_node_state(node, N_POSSIBLE) {
430                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
431                         mz = mem_cgroup_zoneinfo(mem, node, zone);
432                         mctz = soft_limit_tree_node_zone(node, zone);
433                         mem_cgroup_remove_exceeded(mem, mz, mctz);
434                 }
435         }
436 }
437
438 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
439 {
440         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
441 }
442
443 static struct mem_cgroup_per_zone *
444 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
445 {
446         struct rb_node *rightmost = NULL;
447         struct mem_cgroup_per_zone *mz;
448
449 retry:
450         mz = NULL;
451         rightmost = rb_last(&mctz->rb_root);
452         if (!rightmost)
453                 goto done;              /* Nothing to reclaim from */
454
455         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
456         /*
457          * Remove the node now but someone else can add it back,
458          * we will to add it back at the end of reclaim to its correct
459          * position in the tree.
460          */
461         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
462         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
463                 !css_tryget(&mz->mem->css))
464                 goto retry;
465 done:
466         return mz;
467 }
468
469 static struct mem_cgroup_per_zone *
470 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
471 {
472         struct mem_cgroup_per_zone *mz;
473
474         spin_lock(&mctz->lock);
475         mz = __mem_cgroup_largest_soft_limit_node(mctz);
476         spin_unlock(&mctz->lock);
477         return mz;
478 }
479
480 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
481                                          bool charge)
482 {
483         int val = (charge) ? 1 : -1;
484         struct mem_cgroup_stat *stat = &mem->stat;
485         struct mem_cgroup_stat_cpu *cpustat;
486         int cpu = get_cpu();
487
488         cpustat = &stat->cpustat[cpu];
489         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
490         put_cpu();
491 }
492
493 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
494                                          struct page_cgroup *pc,
495                                          bool charge)
496 {
497         int val = (charge) ? 1 : -1;
498         struct mem_cgroup_stat *stat = &mem->stat;
499         struct mem_cgroup_stat_cpu *cpustat;
500         int cpu = get_cpu();
501
502         cpustat = &stat->cpustat[cpu];
503         if (PageCgroupCache(pc))
504                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
505         else
506                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
507
508         if (charge)
509                 __mem_cgroup_stat_add_safe(cpustat,
510                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
511         else
512                 __mem_cgroup_stat_add_safe(cpustat,
513                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
514         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1);
515         put_cpu();
516 }
517
518 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
519                                         enum lru_list idx)
520 {
521         int nid, zid;
522         struct mem_cgroup_per_zone *mz;
523         u64 total = 0;
524
525         for_each_online_node(nid)
526                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
527                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
528                         total += MEM_CGROUP_ZSTAT(mz, idx);
529                 }
530         return total;
531 }
532
533 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
534 {
535         return container_of(cgroup_subsys_state(cont,
536                                 mem_cgroup_subsys_id), struct mem_cgroup,
537                                 css);
538 }
539
540 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
541 {
542         /*
543          * mm_update_next_owner() may clear mm->owner to NULL
544          * if it races with swapoff, page migration, etc.
545          * So this can be called with p == NULL.
546          */
547         if (unlikely(!p))
548                 return NULL;
549
550         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
551                                 struct mem_cgroup, css);
552 }
553
554 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
555 {
556         struct mem_cgroup *mem = NULL;
557
558         if (!mm)
559                 return NULL;
560         /*
561          * Because we have no locks, mm->owner's may be being moved to other
562          * cgroup. We use css_tryget() here even if this looks
563          * pessimistic (rather than adding locks here).
564          */
565         rcu_read_lock();
566         do {
567                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
568                 if (unlikely(!mem))
569                         break;
570         } while (!css_tryget(&mem->css));
571         rcu_read_unlock();
572         return mem;
573 }
574
575 /*
576  * Call callback function against all cgroup under hierarchy tree.
577  */
578 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
579                           int (*func)(struct mem_cgroup *, void *))
580 {
581         int found, ret, nextid;
582         struct cgroup_subsys_state *css;
583         struct mem_cgroup *mem;
584
585         if (!root->use_hierarchy)
586                 return (*func)(root, data);
587
588         nextid = 1;
589         do {
590                 ret = 0;
591                 mem = NULL;
592
593                 rcu_read_lock();
594                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
595                                    &found);
596                 if (css && css_tryget(css))
597                         mem = container_of(css, struct mem_cgroup, css);
598                 rcu_read_unlock();
599
600                 if (mem) {
601                         ret = (*func)(mem, data);
602                         css_put(&mem->css);
603                 }
604                 nextid = found + 1;
605         } while (!ret && css);
606
607         return ret;
608 }
609
610 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
611 {
612         return (mem == root_mem_cgroup);
613 }
614
615 /*
616  * Following LRU functions are allowed to be used without PCG_LOCK.
617  * Operations are called by routine of global LRU independently from memcg.
618  * What we have to take care of here is validness of pc->mem_cgroup.
619  *
620  * Changes to pc->mem_cgroup happens when
621  * 1. charge
622  * 2. moving account
623  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
624  * It is added to LRU before charge.
625  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
626  * When moving account, the page is not on LRU. It's isolated.
627  */
628
629 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
630 {
631         struct page_cgroup *pc;
632         struct mem_cgroup_per_zone *mz;
633
634         if (mem_cgroup_disabled())
635                 return;
636         pc = lookup_page_cgroup(page);
637         /* can happen while we handle swapcache. */
638         if (!TestClearPageCgroupAcctLRU(pc))
639                 return;
640         VM_BUG_ON(!pc->mem_cgroup);
641         /*
642          * We don't check PCG_USED bit. It's cleared when the "page" is finally
643          * removed from global LRU.
644          */
645         mz = page_cgroup_zoneinfo(pc);
646         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
647         if (mem_cgroup_is_root(pc->mem_cgroup))
648                 return;
649         VM_BUG_ON(list_empty(&pc->lru));
650         list_del_init(&pc->lru);
651         return;
652 }
653
654 void mem_cgroup_del_lru(struct page *page)
655 {
656         mem_cgroup_del_lru_list(page, page_lru(page));
657 }
658
659 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
660 {
661         struct mem_cgroup_per_zone *mz;
662         struct page_cgroup *pc;
663
664         if (mem_cgroup_disabled())
665                 return;
666
667         pc = lookup_page_cgroup(page);
668         /*
669          * Used bit is set without atomic ops but after smp_wmb().
670          * For making pc->mem_cgroup visible, insert smp_rmb() here.
671          */
672         smp_rmb();
673         /* unused or root page is not rotated. */
674         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
675                 return;
676         mz = page_cgroup_zoneinfo(pc);
677         list_move(&pc->lru, &mz->lists[lru]);
678 }
679
680 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
681 {
682         struct page_cgroup *pc;
683         struct mem_cgroup_per_zone *mz;
684
685         if (mem_cgroup_disabled())
686                 return;
687         pc = lookup_page_cgroup(page);
688         VM_BUG_ON(PageCgroupAcctLRU(pc));
689         /*
690          * Used bit is set without atomic ops but after smp_wmb().
691          * For making pc->mem_cgroup visible, insert smp_rmb() here.
692          */
693         smp_rmb();
694         if (!PageCgroupUsed(pc))
695                 return;
696
697         mz = page_cgroup_zoneinfo(pc);
698         MEM_CGROUP_ZSTAT(mz, lru) += 1;
699         SetPageCgroupAcctLRU(pc);
700         if (mem_cgroup_is_root(pc->mem_cgroup))
701                 return;
702         list_add(&pc->lru, &mz->lists[lru]);
703 }
704
705 /*
706  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
707  * lru because the page may.be reused after it's fully uncharged (because of
708  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
709  * it again. This function is only used to charge SwapCache. It's done under
710  * lock_page and expected that zone->lru_lock is never held.
711  */
712 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
713 {
714         unsigned long flags;
715         struct zone *zone = page_zone(page);
716         struct page_cgroup *pc = lookup_page_cgroup(page);
717
718         spin_lock_irqsave(&zone->lru_lock, flags);
719         /*
720          * Forget old LRU when this page_cgroup is *not* used. This Used bit
721          * is guarded by lock_page() because the page is SwapCache.
722          */
723         if (!PageCgroupUsed(pc))
724                 mem_cgroup_del_lru_list(page, page_lru(page));
725         spin_unlock_irqrestore(&zone->lru_lock, flags);
726 }
727
728 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
729 {
730         unsigned long flags;
731         struct zone *zone = page_zone(page);
732         struct page_cgroup *pc = lookup_page_cgroup(page);
733
734         spin_lock_irqsave(&zone->lru_lock, flags);
735         /* link when the page is linked to LRU but page_cgroup isn't */
736         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
737                 mem_cgroup_add_lru_list(page, page_lru(page));
738         spin_unlock_irqrestore(&zone->lru_lock, flags);
739 }
740
741
742 void mem_cgroup_move_lists(struct page *page,
743                            enum lru_list from, enum lru_list to)
744 {
745         if (mem_cgroup_disabled())
746                 return;
747         mem_cgroup_del_lru_list(page, from);
748         mem_cgroup_add_lru_list(page, to);
749 }
750
751 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
752 {
753         int ret;
754         struct mem_cgroup *curr = NULL;
755
756         task_lock(task);
757         rcu_read_lock();
758         curr = try_get_mem_cgroup_from_mm(task->mm);
759         rcu_read_unlock();
760         task_unlock(task);
761         if (!curr)
762                 return 0;
763         if (curr->use_hierarchy)
764                 ret = css_is_ancestor(&curr->css, &mem->css);
765         else
766                 ret = (curr == mem);
767         css_put(&curr->css);
768         return ret;
769 }
770
771 /*
772  * prev_priority control...this will be used in memory reclaim path.
773  */
774 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
775 {
776         int prev_priority;
777
778         spin_lock(&mem->reclaim_param_lock);
779         prev_priority = mem->prev_priority;
780         spin_unlock(&mem->reclaim_param_lock);
781
782         return prev_priority;
783 }
784
785 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
786 {
787         spin_lock(&mem->reclaim_param_lock);
788         if (priority < mem->prev_priority)
789                 mem->prev_priority = priority;
790         spin_unlock(&mem->reclaim_param_lock);
791 }
792
793 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
794 {
795         spin_lock(&mem->reclaim_param_lock);
796         mem->prev_priority = priority;
797         spin_unlock(&mem->reclaim_param_lock);
798 }
799
800 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
801 {
802         unsigned long active;
803         unsigned long inactive;
804         unsigned long gb;
805         unsigned long inactive_ratio;
806
807         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
808         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
809
810         gb = (inactive + active) >> (30 - PAGE_SHIFT);
811         if (gb)
812                 inactive_ratio = int_sqrt(10 * gb);
813         else
814                 inactive_ratio = 1;
815
816         if (present_pages) {
817                 present_pages[0] = inactive;
818                 present_pages[1] = active;
819         }
820
821         return inactive_ratio;
822 }
823
824 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
825 {
826         unsigned long active;
827         unsigned long inactive;
828         unsigned long present_pages[2];
829         unsigned long inactive_ratio;
830
831         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
832
833         inactive = present_pages[0];
834         active = present_pages[1];
835
836         if (inactive * inactive_ratio < active)
837                 return 1;
838
839         return 0;
840 }
841
842 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
843 {
844         unsigned long active;
845         unsigned long inactive;
846
847         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
848         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
849
850         return (active > inactive);
851 }
852
853 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
854                                        struct zone *zone,
855                                        enum lru_list lru)
856 {
857         int nid = zone->zone_pgdat->node_id;
858         int zid = zone_idx(zone);
859         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
860
861         return MEM_CGROUP_ZSTAT(mz, lru);
862 }
863
864 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
865                                                       struct zone *zone)
866 {
867         int nid = zone->zone_pgdat->node_id;
868         int zid = zone_idx(zone);
869         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
870
871         return &mz->reclaim_stat;
872 }
873
874 struct zone_reclaim_stat *
875 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
876 {
877         struct page_cgroup *pc;
878         struct mem_cgroup_per_zone *mz;
879
880         if (mem_cgroup_disabled())
881                 return NULL;
882
883         pc = lookup_page_cgroup(page);
884         /*
885          * Used bit is set without atomic ops but after smp_wmb().
886          * For making pc->mem_cgroup visible, insert smp_rmb() here.
887          */
888         smp_rmb();
889         if (!PageCgroupUsed(pc))
890                 return NULL;
891
892         mz = page_cgroup_zoneinfo(pc);
893         if (!mz)
894                 return NULL;
895
896         return &mz->reclaim_stat;
897 }
898
899 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
900                                         struct list_head *dst,
901                                         unsigned long *scanned, int order,
902                                         int mode, struct zone *z,
903                                         struct mem_cgroup *mem_cont,
904                                         int active, int file)
905 {
906         unsigned long nr_taken = 0;
907         struct page *page;
908         unsigned long scan;
909         LIST_HEAD(pc_list);
910         struct list_head *src;
911         struct page_cgroup *pc, *tmp;
912         int nid = z->zone_pgdat->node_id;
913         int zid = zone_idx(z);
914         struct mem_cgroup_per_zone *mz;
915         int lru = LRU_FILE * file + active;
916         int ret;
917
918         BUG_ON(!mem_cont);
919         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
920         src = &mz->lists[lru];
921
922         scan = 0;
923         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
924                 if (scan >= nr_to_scan)
925                         break;
926
927                 page = pc->page;
928                 if (unlikely(!PageCgroupUsed(pc)))
929                         continue;
930                 if (unlikely(!PageLRU(page)))
931                         continue;
932
933                 scan++;
934                 ret = __isolate_lru_page(page, mode, file);
935                 switch (ret) {
936                 case 0:
937                         list_move(&page->lru, dst);
938                         mem_cgroup_del_lru(page);
939                         nr_taken++;
940                         break;
941                 case -EBUSY:
942                         /* we don't affect global LRU but rotate in our LRU */
943                         mem_cgroup_rotate_lru_list(page, page_lru(page));
944                         break;
945                 default:
946                         break;
947                 }
948         }
949
950         *scanned = scan;
951         return nr_taken;
952 }
953
954 #define mem_cgroup_from_res_counter(counter, member)    \
955         container_of(counter, struct mem_cgroup, member)
956
957 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
958 {
959         if (do_swap_account) {
960                 if (res_counter_check_under_limit(&mem->res) &&
961                         res_counter_check_under_limit(&mem->memsw))
962                         return true;
963         } else
964                 if (res_counter_check_under_limit(&mem->res))
965                         return true;
966         return false;
967 }
968
969 static unsigned int get_swappiness(struct mem_cgroup *memcg)
970 {
971         struct cgroup *cgrp = memcg->css.cgroup;
972         unsigned int swappiness;
973
974         /* root ? */
975         if (cgrp->parent == NULL)
976                 return vm_swappiness;
977
978         spin_lock(&memcg->reclaim_param_lock);
979         swappiness = memcg->swappiness;
980         spin_unlock(&memcg->reclaim_param_lock);
981
982         return swappiness;
983 }
984
985 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
986 {
987         int *val = data;
988         (*val)++;
989         return 0;
990 }
991
992 /**
993  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
994  * @memcg: The memory cgroup that went over limit
995  * @p: Task that is going to be killed
996  *
997  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
998  * enabled
999  */
1000 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1001 {
1002         struct cgroup *task_cgrp;
1003         struct cgroup *mem_cgrp;
1004         /*
1005          * Need a buffer in BSS, can't rely on allocations. The code relies
1006          * on the assumption that OOM is serialized for memory controller.
1007          * If this assumption is broken, revisit this code.
1008          */
1009         static char memcg_name[PATH_MAX];
1010         int ret;
1011
1012         if (!memcg)
1013                 return;
1014
1015
1016         rcu_read_lock();
1017
1018         mem_cgrp = memcg->css.cgroup;
1019         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1020
1021         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1022         if (ret < 0) {
1023                 /*
1024                  * Unfortunately, we are unable to convert to a useful name
1025                  * But we'll still print out the usage information
1026                  */
1027                 rcu_read_unlock();
1028                 goto done;
1029         }
1030         rcu_read_unlock();
1031
1032         printk(KERN_INFO "Task in %s killed", memcg_name);
1033
1034         rcu_read_lock();
1035         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1036         if (ret < 0) {
1037                 rcu_read_unlock();
1038                 goto done;
1039         }
1040         rcu_read_unlock();
1041
1042         /*
1043          * Continues from above, so we don't need an KERN_ level
1044          */
1045         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1046 done:
1047
1048         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1049                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1050                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1051                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1052         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1053                 "failcnt %llu\n",
1054                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1055                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1056                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1057 }
1058
1059 /*
1060  * This function returns the number of memcg under hierarchy tree. Returns
1061  * 1(self count) if no children.
1062  */
1063 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1064 {
1065         int num = 0;
1066         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1067         return num;
1068 }
1069
1070 /*
1071  * Visit the first child (need not be the first child as per the ordering
1072  * of the cgroup list, since we track last_scanned_child) of @mem and use
1073  * that to reclaim free pages from.
1074  */
1075 static struct mem_cgroup *
1076 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1077 {
1078         struct mem_cgroup *ret = NULL;
1079         struct cgroup_subsys_state *css;
1080         int nextid, found;
1081
1082         if (!root_mem->use_hierarchy) {
1083                 css_get(&root_mem->css);
1084                 ret = root_mem;
1085         }
1086
1087         while (!ret) {
1088                 rcu_read_lock();
1089                 nextid = root_mem->last_scanned_child + 1;
1090                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1091                                    &found);
1092                 if (css && css_tryget(css))
1093                         ret = container_of(css, struct mem_cgroup, css);
1094
1095                 rcu_read_unlock();
1096                 /* Updates scanning parameter */
1097                 spin_lock(&root_mem->reclaim_param_lock);
1098                 if (!css) {
1099                         /* this means start scan from ID:1 */
1100                         root_mem->last_scanned_child = 0;
1101                 } else
1102                         root_mem->last_scanned_child = found;
1103                 spin_unlock(&root_mem->reclaim_param_lock);
1104         }
1105
1106         return ret;
1107 }
1108
1109 /*
1110  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1111  * we reclaimed from, so that we don't end up penalizing one child extensively
1112  * based on its position in the children list.
1113  *
1114  * root_mem is the original ancestor that we've been reclaim from.
1115  *
1116  * We give up and return to the caller when we visit root_mem twice.
1117  * (other groups can be removed while we're walking....)
1118  *
1119  * If shrink==true, for avoiding to free too much, this returns immedieately.
1120  */
1121 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1122                                                 struct zone *zone,
1123                                                 gfp_t gfp_mask,
1124                                                 unsigned long reclaim_options)
1125 {
1126         struct mem_cgroup *victim;
1127         int ret, total = 0;
1128         int loop = 0;
1129         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1130         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1131         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1132         unsigned long excess = mem_cgroup_get_excess(root_mem);
1133
1134         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1135         if (root_mem->memsw_is_minimum)
1136                 noswap = true;
1137
1138         while (1) {
1139                 victim = mem_cgroup_select_victim(root_mem);
1140                 if (victim == root_mem) {
1141                         loop++;
1142                         if (loop >= 1)
1143                                 drain_all_stock_async();
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_file_mapped(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         pc = lookup_page_cgroup(page);
1239         if (unlikely(!pc))
1240                 return;
1241
1242         lock_page_cgroup(pc);
1243         mem = pc->mem_cgroup;
1244         if (!mem)
1245                 goto done;
1246
1247         if (!PageCgroupUsed(pc))
1248                 goto done;
1249
1250         /*
1251          * Preemption is already disabled, we don't need get_cpu()
1252          */
1253         cpu = smp_processor_id();
1254         stat = &mem->stat;
1255         cpustat = &stat->cpustat[cpu];
1256
1257         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED, val);
1258 done:
1259         unlock_page_cgroup(pc);
1260 }
1261
1262 /*
1263  * size of first charge trial. "32" comes from vmscan.c's magic value.
1264  * TODO: maybe necessary to use big numbers in big irons.
1265  */
1266 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1267 struct memcg_stock_pcp {
1268         struct mem_cgroup *cached; /* this never be root cgroup */
1269         int charge;
1270         struct work_struct work;
1271 };
1272 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1273 static atomic_t memcg_drain_count;
1274
1275 /*
1276  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1277  * from local stock and true is returned. If the stock is 0 or charges from a
1278  * cgroup which is not current target, returns false. This stock will be
1279  * refilled.
1280  */
1281 static bool consume_stock(struct mem_cgroup *mem)
1282 {
1283         struct memcg_stock_pcp *stock;
1284         bool ret = true;
1285
1286         stock = &get_cpu_var(memcg_stock);
1287         if (mem == stock->cached && stock->charge)
1288                 stock->charge -= PAGE_SIZE;
1289         else /* need to call res_counter_charge */
1290                 ret = false;
1291         put_cpu_var(memcg_stock);
1292         return ret;
1293 }
1294
1295 /*
1296  * Returns stocks cached in percpu to res_counter and reset cached information.
1297  */
1298 static void drain_stock(struct memcg_stock_pcp *stock)
1299 {
1300         struct mem_cgroup *old = stock->cached;
1301
1302         if (stock->charge) {
1303                 res_counter_uncharge(&old->res, stock->charge);
1304                 if (do_swap_account)
1305                         res_counter_uncharge(&old->memsw, stock->charge);
1306         }
1307         stock->cached = NULL;
1308         stock->charge = 0;
1309 }
1310
1311 /*
1312  * This must be called under preempt disabled or must be called by
1313  * a thread which is pinned to local cpu.
1314  */
1315 static void drain_local_stock(struct work_struct *dummy)
1316 {
1317         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1318         drain_stock(stock);
1319 }
1320
1321 /*
1322  * Cache charges(val) which is from res_counter, to local per_cpu area.
1323  * This will be consumed by consumt_stock() function, later.
1324  */
1325 static void refill_stock(struct mem_cgroup *mem, int val)
1326 {
1327         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1328
1329         if (stock->cached != mem) { /* reset if necessary */
1330                 drain_stock(stock);
1331                 stock->cached = mem;
1332         }
1333         stock->charge += val;
1334         put_cpu_var(memcg_stock);
1335 }
1336
1337 /*
1338  * Tries to drain stocked charges in other cpus. This function is asynchronous
1339  * and just put a work per cpu for draining localy on each cpu. Caller can
1340  * expects some charges will be back to res_counter later but cannot wait for
1341  * it.
1342  */
1343 static void drain_all_stock_async(void)
1344 {
1345         int cpu;
1346         /* This function is for scheduling "drain" in asynchronous way.
1347          * The result of "drain" is not directly handled by callers. Then,
1348          * if someone is calling drain, we don't have to call drain more.
1349          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1350          * there is a race. We just do loose check here.
1351          */
1352         if (atomic_read(&memcg_drain_count))
1353                 return;
1354         /* Notify other cpus that system-wide "drain" is running */
1355         atomic_inc(&memcg_drain_count);
1356         get_online_cpus();
1357         for_each_online_cpu(cpu) {
1358                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1359                 schedule_work_on(cpu, &stock->work);
1360         }
1361         put_online_cpus();
1362         atomic_dec(&memcg_drain_count);
1363         /* We don't wait for flush_work */
1364 }
1365
1366 /* This is a synchronous drain interface. */
1367 static void drain_all_stock_sync(void)
1368 {
1369         /* called when force_empty is called */
1370         atomic_inc(&memcg_drain_count);
1371         schedule_on_each_cpu(drain_local_stock);
1372         atomic_dec(&memcg_drain_count);
1373 }
1374
1375 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1376                                         unsigned long action,
1377                                         void *hcpu)
1378 {
1379         int cpu = (unsigned long)hcpu;
1380         struct memcg_stock_pcp *stock;
1381
1382         if (action != CPU_DEAD)
1383                 return NOTIFY_OK;
1384         stock = &per_cpu(memcg_stock, cpu);
1385         drain_stock(stock);
1386         return NOTIFY_OK;
1387 }
1388
1389 /*
1390  * Unlike exported interface, "oom" parameter is added. if oom==true,
1391  * oom-killer can be invoked.
1392  */
1393 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1394                         gfp_t gfp_mask, struct mem_cgroup **memcg,
1395                         bool oom, struct page *page)
1396 {
1397         struct mem_cgroup *mem, *mem_over_limit;
1398         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1399         struct res_counter *fail_res;
1400         int csize = CHARGE_SIZE;
1401
1402         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1403                 /* Don't account this! */
1404                 *memcg = NULL;
1405                 return 0;
1406         }
1407
1408         /*
1409          * We always charge the cgroup the mm_struct belongs to.
1410          * The mm_struct's mem_cgroup changes on task migration if the
1411          * thread group leader migrates. It's possible that mm is not
1412          * set, if so charge the init_mm (happens for pagecache usage).
1413          */
1414         mem = *memcg;
1415         if (likely(!mem)) {
1416                 mem = try_get_mem_cgroup_from_mm(mm);
1417                 *memcg = mem;
1418         } else {
1419                 css_get(&mem->css);
1420         }
1421         if (unlikely(!mem))
1422                 return 0;
1423
1424         VM_BUG_ON(css_is_removed(&mem->css));
1425         if (mem_cgroup_is_root(mem))
1426                 goto done;
1427
1428         while (1) {
1429                 int ret = 0;
1430                 unsigned long flags = 0;
1431
1432                 if (consume_stock(mem))
1433                         goto charged;
1434
1435                 ret = res_counter_charge(&mem->res, csize, &fail_res);
1436                 if (likely(!ret)) {
1437                         if (!do_swap_account)
1438                                 break;
1439                         ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1440                         if (likely(!ret))
1441                                 break;
1442                         /* mem+swap counter fails */
1443                         res_counter_uncharge(&mem->res, csize);
1444                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1445                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1446                                                                         memsw);
1447                 } else
1448                         /* mem counter fails */
1449                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1450                                                                         res);
1451
1452                 /* reduce request size and retry */
1453                 if (csize > PAGE_SIZE) {
1454                         csize = PAGE_SIZE;
1455                         continue;
1456                 }
1457                 if (!(gfp_mask & __GFP_WAIT))
1458                         goto nomem;
1459
1460                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1461                                                 gfp_mask, flags);
1462                 if (ret)
1463                         continue;
1464
1465                 /*
1466                  * try_to_free_mem_cgroup_pages() might not give us a full
1467                  * picture of reclaim. Some pages are reclaimed and might be
1468                  * moved to swap cache or just unmapped from the cgroup.
1469                  * Check the limit again to see if the reclaim reduced the
1470                  * current usage of the cgroup before giving up
1471                  *
1472                  */
1473                 if (mem_cgroup_check_under_limit(mem_over_limit))
1474                         continue;
1475
1476                 if (!nr_retries--) {
1477                         if (oom) {
1478                                 mutex_lock(&memcg_tasklist);
1479                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1480                                 mutex_unlock(&memcg_tasklist);
1481                                 record_last_oom(mem_over_limit);
1482                         }
1483                         goto nomem;
1484                 }
1485         }
1486         if (csize > PAGE_SIZE)
1487                 refill_stock(mem, csize - PAGE_SIZE);
1488 charged:
1489         /*
1490          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1491          * if they exceeds softlimit.
1492          */
1493         if (mem_cgroup_soft_limit_check(mem))
1494                 mem_cgroup_update_tree(mem, page);
1495 done:
1496         return 0;
1497 nomem:
1498         css_put(&mem->css);
1499         return -ENOMEM;
1500 }
1501
1502 /*
1503  * A helper function to get mem_cgroup from ID. must be called under
1504  * rcu_read_lock(). The caller must check css_is_removed() or some if
1505  * it's concern. (dropping refcnt from swap can be called against removed
1506  * memcg.)
1507  */
1508 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1509 {
1510         struct cgroup_subsys_state *css;
1511
1512         /* ID 0 is unused ID */
1513         if (!id)
1514                 return NULL;
1515         css = css_lookup(&mem_cgroup_subsys, id);
1516         if (!css)
1517                 return NULL;
1518         return container_of(css, struct mem_cgroup, css);
1519 }
1520
1521 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1522 {
1523         struct mem_cgroup *mem;
1524         struct page_cgroup *pc;
1525         unsigned short id;
1526         swp_entry_t ent;
1527
1528         VM_BUG_ON(!PageLocked(page));
1529
1530         if (!PageSwapCache(page))
1531                 return NULL;
1532
1533         pc = lookup_page_cgroup(page);
1534         lock_page_cgroup(pc);
1535         if (PageCgroupUsed(pc)) {
1536                 mem = pc->mem_cgroup;
1537                 if (mem && !css_tryget(&mem->css))
1538                         mem = NULL;
1539         } else {
1540                 ent.val = page_private(page);
1541                 id = lookup_swap_cgroup(ent);
1542                 rcu_read_lock();
1543                 mem = mem_cgroup_lookup(id);
1544                 if (mem && !css_tryget(&mem->css))
1545                         mem = NULL;
1546                 rcu_read_unlock();
1547         }
1548         unlock_page_cgroup(pc);
1549         return mem;
1550 }
1551
1552 /*
1553  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1554  * USED state. If already USED, uncharge and return.
1555  */
1556
1557 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1558                                      struct page_cgroup *pc,
1559                                      enum charge_type ctype)
1560 {
1561         /* try_charge() can return NULL to *memcg, taking care of it. */
1562         if (!mem)
1563                 return;
1564
1565         lock_page_cgroup(pc);
1566         if (unlikely(PageCgroupUsed(pc))) {
1567                 unlock_page_cgroup(pc);
1568                 if (!mem_cgroup_is_root(mem)) {
1569                         res_counter_uncharge(&mem->res, PAGE_SIZE);
1570                         if (do_swap_account)
1571                                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1572                 }
1573                 css_put(&mem->css);
1574                 return;
1575         }
1576
1577         pc->mem_cgroup = mem;
1578         /*
1579          * We access a page_cgroup asynchronously without lock_page_cgroup().
1580          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1581          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1582          * before USED bit, we need memory barrier here.
1583          * See mem_cgroup_add_lru_list(), etc.
1584          */
1585         smp_wmb();
1586         switch (ctype) {
1587         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1588         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1589                 SetPageCgroupCache(pc);
1590                 SetPageCgroupUsed(pc);
1591                 break;
1592         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1593                 ClearPageCgroupCache(pc);
1594                 SetPageCgroupUsed(pc);
1595                 break;
1596         default:
1597                 break;
1598         }
1599
1600         mem_cgroup_charge_statistics(mem, pc, true);
1601
1602         unlock_page_cgroup(pc);
1603 }
1604
1605 /**
1606  * mem_cgroup_move_account - move account of the page
1607  * @pc: page_cgroup of the page.
1608  * @from: mem_cgroup which the page is moved from.
1609  * @to: mem_cgroup which the page is moved to. @from != @to.
1610  *
1611  * The caller must confirm following.
1612  * - page is not on LRU (isolate_page() is useful.)
1613  *
1614  * returns 0 at success,
1615  * returns -EBUSY when lock is busy or "pc" is unstable.
1616  *
1617  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1618  * new cgroup. It should be done by a caller.
1619  */
1620
1621 static int mem_cgroup_move_account(struct page_cgroup *pc,
1622         struct mem_cgroup *from, struct mem_cgroup *to)
1623 {
1624         struct mem_cgroup_per_zone *from_mz, *to_mz;
1625         int nid, zid;
1626         int ret = -EBUSY;
1627         struct page *page;
1628         int cpu;
1629         struct mem_cgroup_stat *stat;
1630         struct mem_cgroup_stat_cpu *cpustat;
1631
1632         VM_BUG_ON(from == to);
1633         VM_BUG_ON(PageLRU(pc->page));
1634
1635         nid = page_cgroup_nid(pc);
1636         zid = page_cgroup_zid(pc);
1637         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1638         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1639
1640         if (!trylock_page_cgroup(pc))
1641                 return ret;
1642
1643         if (!PageCgroupUsed(pc))
1644                 goto out;
1645
1646         if (pc->mem_cgroup != from)
1647                 goto out;
1648
1649         if (!mem_cgroup_is_root(from))
1650                 res_counter_uncharge(&from->res, PAGE_SIZE);
1651         mem_cgroup_charge_statistics(from, pc, false);
1652
1653         page = pc->page;
1654         if (page_mapped(page) && !PageAnon(page)) {
1655                 cpu = smp_processor_id();
1656                 /* Update mapped_file data for mem_cgroup "from" */
1657                 stat = &from->stat;
1658                 cpustat = &stat->cpustat[cpu];
1659                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
1660                                                 -1);
1661
1662                 /* Update mapped_file data for mem_cgroup "to" */
1663                 stat = &to->stat;
1664                 cpustat = &stat->cpustat[cpu];
1665                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
1666                                                 1);
1667         }
1668
1669         if (do_swap_account && !mem_cgroup_is_root(from))
1670                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1671         css_put(&from->css);
1672
1673         css_get(&to->css);
1674         pc->mem_cgroup = to;
1675         mem_cgroup_charge_statistics(to, pc, true);
1676         ret = 0;
1677 out:
1678         unlock_page_cgroup(pc);
1679         /*
1680          * We charges against "to" which may not have any tasks. Then, "to"
1681          * can be under rmdir(). But in current implementation, caller of
1682          * this function is just force_empty() and it's garanteed that
1683          * "to" is never removed. So, we don't check rmdir status here.
1684          */
1685         return ret;
1686 }
1687
1688 /*
1689  * move charges to its parent.
1690  */
1691
1692 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1693                                   struct mem_cgroup *child,
1694                                   gfp_t gfp_mask)
1695 {
1696         struct page *page = pc->page;
1697         struct cgroup *cg = child->css.cgroup;
1698         struct cgroup *pcg = cg->parent;
1699         struct mem_cgroup *parent;
1700         int ret;
1701
1702         /* Is ROOT ? */
1703         if (!pcg)
1704                 return -EINVAL;
1705
1706
1707         parent = mem_cgroup_from_cont(pcg);
1708
1709
1710         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page);
1711         if (ret || !parent)
1712                 return ret;
1713
1714         if (!get_page_unless_zero(page)) {
1715                 ret = -EBUSY;
1716                 goto uncharge;
1717         }
1718
1719         ret = isolate_lru_page(page);
1720
1721         if (ret)
1722                 goto cancel;
1723
1724         ret = mem_cgroup_move_account(pc, child, parent);
1725
1726         putback_lru_page(page);
1727         if (!ret) {
1728                 put_page(page);
1729                 /* drop extra refcnt by try_charge() */
1730                 css_put(&parent->css);
1731                 return 0;
1732         }
1733
1734 cancel:
1735         put_page(page);
1736 uncharge:
1737         /* drop extra refcnt by try_charge() */
1738         css_put(&parent->css);
1739         /* uncharge if move fails */
1740         if (!mem_cgroup_is_root(parent)) {
1741                 res_counter_uncharge(&parent->res, PAGE_SIZE);
1742                 if (do_swap_account)
1743                         res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1744         }
1745         return ret;
1746 }
1747
1748 /*
1749  * Charge the memory controller for page usage.
1750  * Return
1751  * 0 if the charge was successful
1752  * < 0 if the cgroup is over its limit
1753  */
1754 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1755                                 gfp_t gfp_mask, enum charge_type ctype,
1756                                 struct mem_cgroup *memcg)
1757 {
1758         struct mem_cgroup *mem;
1759         struct page_cgroup *pc;
1760         int ret;
1761
1762         pc = lookup_page_cgroup(page);
1763         /* can happen at boot */
1764         if (unlikely(!pc))
1765                 return 0;
1766         prefetchw(pc);
1767
1768         mem = memcg;
1769         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page);
1770         if (ret || !mem)
1771                 return ret;
1772
1773         __mem_cgroup_commit_charge(mem, pc, ctype);
1774         return 0;
1775 }
1776
1777 int mem_cgroup_newpage_charge(struct page *page,
1778                               struct mm_struct *mm, gfp_t gfp_mask)
1779 {
1780         if (mem_cgroup_disabled())
1781                 return 0;
1782         if (PageCompound(page))
1783                 return 0;
1784         /*
1785          * If already mapped, we don't have to account.
1786          * If page cache, page->mapping has address_space.
1787          * But page->mapping may have out-of-use anon_vma pointer,
1788          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1789          * is NULL.
1790          */
1791         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1792                 return 0;
1793         if (unlikely(!mm))
1794                 mm = &init_mm;
1795         return mem_cgroup_charge_common(page, mm, gfp_mask,
1796                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1797 }
1798
1799 static void
1800 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1801                                         enum charge_type ctype);
1802
1803 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1804                                 gfp_t gfp_mask)
1805 {
1806         struct mem_cgroup *mem = NULL;
1807         int ret;
1808
1809         if (mem_cgroup_disabled())
1810                 return 0;
1811         if (PageCompound(page))
1812                 return 0;
1813         /*
1814          * Corner case handling. This is called from add_to_page_cache()
1815          * in usual. But some FS (shmem) precharges this page before calling it
1816          * and call add_to_page_cache() with GFP_NOWAIT.
1817          *
1818          * For GFP_NOWAIT case, the page may be pre-charged before calling
1819          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1820          * charge twice. (It works but has to pay a bit larger cost.)
1821          * And when the page is SwapCache, it should take swap information
1822          * into account. This is under lock_page() now.
1823          */
1824         if (!(gfp_mask & __GFP_WAIT)) {
1825                 struct page_cgroup *pc;
1826
1827
1828                 pc = lookup_page_cgroup(page);
1829                 if (!pc)
1830                         return 0;
1831                 lock_page_cgroup(pc);
1832                 if (PageCgroupUsed(pc)) {
1833                         unlock_page_cgroup(pc);
1834                         return 0;
1835                 }
1836                 unlock_page_cgroup(pc);
1837         }
1838
1839         if (unlikely(!mm && !mem))
1840                 mm = &init_mm;
1841
1842         if (page_is_file_cache(page))
1843                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1844                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1845
1846         /* shmem */
1847         if (PageSwapCache(page)) {
1848                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1849                 if (!ret)
1850                         __mem_cgroup_commit_charge_swapin(page, mem,
1851                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1852         } else
1853                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1854                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1855
1856         return ret;
1857 }
1858
1859 /*
1860  * While swap-in, try_charge -> commit or cancel, the page is locked.
1861  * And when try_charge() successfully returns, one refcnt to memcg without
1862  * struct page_cgroup is acquired. This refcnt will be consumed by
1863  * "commit()" or removed by "cancel()"
1864  */
1865 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1866                                  struct page *page,
1867                                  gfp_t mask, struct mem_cgroup **ptr)
1868 {
1869         struct mem_cgroup *mem;
1870         int ret;
1871
1872         if (mem_cgroup_disabled())
1873                 return 0;
1874
1875         if (!do_swap_account)
1876                 goto charge_cur_mm;
1877         /*
1878          * A racing thread's fault, or swapoff, may have already updated
1879          * the pte, and even removed page from swap cache: in those cases
1880          * do_swap_page()'s pte_same() test will fail; but there's also a
1881          * KSM case which does need to charge the page.
1882          */
1883         if (!PageSwapCache(page))
1884                 goto charge_cur_mm;
1885         mem = try_get_mem_cgroup_from_swapcache(page);
1886         if (!mem)
1887                 goto charge_cur_mm;
1888         *ptr = mem;
1889         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page);
1890         /* drop extra refcnt from tryget */
1891         css_put(&mem->css);
1892         return ret;
1893 charge_cur_mm:
1894         if (unlikely(!mm))
1895                 mm = &init_mm;
1896         return __mem_cgroup_try_charge(mm, mask, ptr, true, page);
1897 }
1898
1899 static void
1900 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1901                                         enum charge_type ctype)
1902 {
1903         struct page_cgroup *pc;
1904
1905         if (mem_cgroup_disabled())
1906                 return;
1907         if (!ptr)
1908                 return;
1909         cgroup_exclude_rmdir(&ptr->css);
1910         pc = lookup_page_cgroup(page);
1911         mem_cgroup_lru_del_before_commit_swapcache(page);
1912         __mem_cgroup_commit_charge(ptr, pc, ctype);
1913         mem_cgroup_lru_add_after_commit_swapcache(page);
1914         /*
1915          * Now swap is on-memory. This means this page may be
1916          * counted both as mem and swap....double count.
1917          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1918          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1919          * may call delete_from_swap_cache() before reach here.
1920          */
1921         if (do_swap_account && PageSwapCache(page)) {
1922                 swp_entry_t ent = {.val = page_private(page)};
1923                 unsigned short id;
1924                 struct mem_cgroup *memcg;
1925
1926                 id = swap_cgroup_record(ent, 0);
1927                 rcu_read_lock();
1928                 memcg = mem_cgroup_lookup(id);
1929                 if (memcg) {
1930                         /*
1931                          * This recorded memcg can be obsolete one. So, avoid
1932                          * calling css_tryget
1933                          */
1934                         if (!mem_cgroup_is_root(memcg))
1935                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1936                         mem_cgroup_swap_statistics(memcg, false);
1937                         mem_cgroup_put(memcg);
1938                 }
1939                 rcu_read_unlock();
1940         }
1941         /*
1942          * At swapin, we may charge account against cgroup which has no tasks.
1943          * So, rmdir()->pre_destroy() can be called while we do this charge.
1944          * In that case, we need to call pre_destroy() again. check it here.
1945          */
1946         cgroup_release_and_wakeup_rmdir(&ptr->css);
1947 }
1948
1949 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1950 {
1951         __mem_cgroup_commit_charge_swapin(page, ptr,
1952                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1953 }
1954
1955 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1956 {
1957         if (mem_cgroup_disabled())
1958                 return;
1959         if (!mem)
1960                 return;
1961         if (!mem_cgroup_is_root(mem)) {
1962                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1963                 if (do_swap_account)
1964                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1965         }
1966         css_put(&mem->css);
1967 }
1968
1969 static void
1970 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
1971 {
1972         struct memcg_batch_info *batch = NULL;
1973         bool uncharge_memsw = true;
1974         /* If swapout, usage of swap doesn't decrease */
1975         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1976                 uncharge_memsw = false;
1977         /*
1978          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
1979          * In those cases, all pages freed continously can be expected to be in
1980          * the same cgroup and we have chance to coalesce uncharges.
1981          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
1982          * because we want to do uncharge as soon as possible.
1983          */
1984         if (!current->memcg_batch.do_batch || test_thread_flag(TIF_MEMDIE))
1985                 goto direct_uncharge;
1986
1987         batch = &current->memcg_batch;
1988         /*
1989          * In usual, we do css_get() when we remember memcg pointer.
1990          * But in this case, we keep res->usage until end of a series of
1991          * uncharges. Then, it's ok to ignore memcg's refcnt.
1992          */
1993         if (!batch->memcg)
1994                 batch->memcg = mem;
1995         /*
1996          * In typical case, batch->memcg == mem. This means we can
1997          * merge a series of uncharges to an uncharge of res_counter.
1998          * If not, we uncharge res_counter ony by one.
1999          */
2000         if (batch->memcg != mem)
2001                 goto direct_uncharge;
2002         /* remember freed charge and uncharge it later */
2003         batch->bytes += PAGE_SIZE;
2004         if (uncharge_memsw)
2005                 batch->memsw_bytes += PAGE_SIZE;
2006         return;
2007 direct_uncharge:
2008         res_counter_uncharge(&mem->res, PAGE_SIZE);
2009         if (uncharge_memsw)
2010                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2011         return;
2012 }
2013
2014 /*
2015  * uncharge if !page_mapped(page)
2016  */
2017 static struct mem_cgroup *
2018 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2019 {
2020         struct page_cgroup *pc;
2021         struct mem_cgroup *mem = NULL;
2022         struct mem_cgroup_per_zone *mz;
2023
2024         if (mem_cgroup_disabled())
2025                 return NULL;
2026
2027         if (PageSwapCache(page))
2028                 return NULL;
2029
2030         /*
2031          * Check if our page_cgroup is valid
2032          */
2033         pc = lookup_page_cgroup(page);
2034         if (unlikely(!pc || !PageCgroupUsed(pc)))
2035                 return NULL;
2036
2037         lock_page_cgroup(pc);
2038
2039         mem = pc->mem_cgroup;
2040
2041         if (!PageCgroupUsed(pc))
2042                 goto unlock_out;
2043
2044         switch (ctype) {
2045         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2046         case MEM_CGROUP_CHARGE_TYPE_DROP:
2047                 if (page_mapped(page))
2048                         goto unlock_out;
2049                 break;
2050         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2051                 if (!PageAnon(page)) {  /* Shared memory */
2052                         if (page->mapping && !page_is_file_cache(page))
2053                                 goto unlock_out;
2054                 } else if (page_mapped(page)) /* Anon */
2055                                 goto unlock_out;
2056                 break;
2057         default:
2058                 break;
2059         }
2060
2061         if (!mem_cgroup_is_root(mem))
2062                 __do_uncharge(mem, ctype);
2063         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2064                 mem_cgroup_swap_statistics(mem, true);
2065         mem_cgroup_charge_statistics(mem, pc, false);
2066
2067         ClearPageCgroupUsed(pc);
2068         /*
2069          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2070          * freed from LRU. This is safe because uncharged page is expected not
2071          * to be reused (freed soon). Exception is SwapCache, it's handled by
2072          * special functions.
2073          */
2074
2075         mz = page_cgroup_zoneinfo(pc);
2076         unlock_page_cgroup(pc);
2077
2078         if (mem_cgroup_soft_limit_check(mem))
2079                 mem_cgroup_update_tree(mem, page);
2080         /* at swapout, this memcg will be accessed to record to swap */
2081         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2082                 css_put(&mem->css);
2083
2084         return mem;
2085
2086 unlock_out:
2087         unlock_page_cgroup(pc);
2088         return NULL;
2089 }
2090
2091 void mem_cgroup_uncharge_page(struct page *page)
2092 {
2093         /* early check. */
2094         if (page_mapped(page))
2095                 return;
2096         if (page->mapping && !PageAnon(page))
2097                 return;
2098         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2099 }
2100
2101 void mem_cgroup_uncharge_cache_page(struct page *page)
2102 {
2103         VM_BUG_ON(page_mapped(page));
2104         VM_BUG_ON(page->mapping);
2105         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2106 }
2107
2108 /*
2109  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2110  * In that cases, pages are freed continuously and we can expect pages
2111  * are in the same memcg. All these calls itself limits the number of
2112  * pages freed at once, then uncharge_start/end() is called properly.
2113  * This may be called prural(2) times in a context,
2114  */
2115
2116 void mem_cgroup_uncharge_start(void)
2117 {
2118         current->memcg_batch.do_batch++;
2119         /* We can do nest. */
2120         if (current->memcg_batch.do_batch == 1) {
2121                 current->memcg_batch.memcg = NULL;
2122                 current->memcg_batch.bytes = 0;
2123                 current->memcg_batch.memsw_bytes = 0;
2124         }
2125 }
2126
2127 void mem_cgroup_uncharge_end(void)
2128 {
2129         struct memcg_batch_info *batch = &current->memcg_batch;
2130
2131         if (!batch->do_batch)
2132                 return;
2133
2134         batch->do_batch--;
2135         if (batch->do_batch) /* If stacked, do nothing. */
2136                 return;
2137
2138         if (!batch->memcg)
2139                 return;
2140         /*
2141          * This "batch->memcg" is valid without any css_get/put etc...
2142          * bacause we hide charges behind us.
2143          */
2144         if (batch->bytes)
2145                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2146         if (batch->memsw_bytes)
2147                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2148         /* forget this pointer (for sanity check) */
2149         batch->memcg = NULL;
2150 }
2151
2152 #ifdef CONFIG_SWAP
2153 /*
2154  * called after __delete_from_swap_cache() and drop "page" account.
2155  * memcg information is recorded to swap_cgroup of "ent"
2156  */
2157 void
2158 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2159 {
2160         struct mem_cgroup *memcg;
2161         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2162
2163         if (!swapout) /* this was a swap cache but the swap is unused ! */
2164                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2165
2166         memcg = __mem_cgroup_uncharge_common(page, ctype);
2167
2168         /* record memcg information */
2169         if (do_swap_account && swapout && memcg) {
2170                 swap_cgroup_record(ent, css_id(&memcg->css));
2171                 mem_cgroup_get(memcg);
2172         }
2173         if (swapout && memcg)
2174                 css_put(&memcg->css);
2175 }
2176 #endif
2177
2178 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2179 /*
2180  * called from swap_entry_free(). remove record in swap_cgroup and
2181  * uncharge "memsw" account.
2182  */
2183 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2184 {
2185         struct mem_cgroup *memcg;
2186         unsigned short id;
2187
2188         if (!do_swap_account)
2189                 return;
2190
2191         id = swap_cgroup_record(ent, 0);
2192         rcu_read_lock();
2193         memcg = mem_cgroup_lookup(id);
2194         if (memcg) {
2195                 /*
2196                  * We uncharge this because swap is freed.
2197                  * This memcg can be obsolete one. We avoid calling css_tryget
2198                  */
2199                 if (!mem_cgroup_is_root(memcg))
2200                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2201                 mem_cgroup_swap_statistics(memcg, false);
2202                 mem_cgroup_put(memcg);
2203         }
2204         rcu_read_unlock();
2205 }
2206 #endif
2207
2208 /*
2209  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2210  * page belongs to.
2211  */
2212 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2213 {
2214         struct page_cgroup *pc;
2215         struct mem_cgroup *mem = NULL;
2216         int ret = 0;
2217
2218         if (mem_cgroup_disabled())
2219                 return 0;
2220
2221         pc = lookup_page_cgroup(page);
2222         lock_page_cgroup(pc);
2223         if (PageCgroupUsed(pc)) {
2224                 mem = pc->mem_cgroup;
2225                 css_get(&mem->css);
2226         }
2227         unlock_page_cgroup(pc);
2228
2229         if (mem) {
2230                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
2231                                                 page);
2232                 css_put(&mem->css);
2233         }
2234         *ptr = mem;
2235         return ret;
2236 }
2237
2238 /* remove redundant charge if migration failed*/
2239 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2240                 struct page *oldpage, struct page *newpage)
2241 {
2242         struct page *target, *unused;
2243         struct page_cgroup *pc;
2244         enum charge_type ctype;
2245
2246         if (!mem)
2247                 return;
2248         cgroup_exclude_rmdir(&mem->css);
2249         /* at migration success, oldpage->mapping is NULL. */
2250         if (oldpage->mapping) {
2251                 target = oldpage;
2252                 unused = NULL;
2253         } else {
2254                 target = newpage;
2255                 unused = oldpage;
2256         }
2257
2258         if (PageAnon(target))
2259                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2260         else if (page_is_file_cache(target))
2261                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2262         else
2263                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2264
2265         /* unused page is not on radix-tree now. */
2266         if (unused)
2267                 __mem_cgroup_uncharge_common(unused, ctype);
2268
2269         pc = lookup_page_cgroup(target);
2270         /*
2271          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2272          * So, double-counting is effectively avoided.
2273          */
2274         __mem_cgroup_commit_charge(mem, pc, ctype);
2275
2276         /*
2277          * Both of oldpage and newpage are still under lock_page().
2278          * Then, we don't have to care about race in radix-tree.
2279          * But we have to be careful that this page is unmapped or not.
2280          *
2281          * There is a case for !page_mapped(). At the start of
2282          * migration, oldpage was mapped. But now, it's zapped.
2283          * But we know *target* page is not freed/reused under us.
2284          * mem_cgroup_uncharge_page() does all necessary checks.
2285          */
2286         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2287                 mem_cgroup_uncharge_page(target);
2288         /*
2289          * At migration, we may charge account against cgroup which has no tasks
2290          * So, rmdir()->pre_destroy() can be called while we do this charge.
2291          * In that case, we need to call pre_destroy() again. check it here.
2292          */
2293         cgroup_release_and_wakeup_rmdir(&mem->css);
2294 }
2295
2296 /*
2297  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2298  * Calling hierarchical_reclaim is not enough because we should update
2299  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2300  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2301  * not from the memcg which this page would be charged to.
2302  * try_charge_swapin does all of these works properly.
2303  */
2304 int mem_cgroup_shmem_charge_fallback(struct page *page,
2305                             struct mm_struct *mm,
2306                             gfp_t gfp_mask)
2307 {
2308         struct mem_cgroup *mem = NULL;
2309         int ret;
2310
2311         if (mem_cgroup_disabled())
2312                 return 0;
2313
2314         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2315         if (!ret)
2316                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2317
2318         return ret;
2319 }
2320
2321 static DEFINE_MUTEX(set_limit_mutex);
2322
2323 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2324                                 unsigned long long val)
2325 {
2326         int retry_count;
2327         int progress;
2328         u64 memswlimit;
2329         int ret = 0;
2330         int children = mem_cgroup_count_children(memcg);
2331         u64 curusage, oldusage;
2332
2333         /*
2334          * For keeping hierarchical_reclaim simple, how long we should retry
2335          * is depends on callers. We set our retry-count to be function
2336          * of # of children which we should visit in this loop.
2337          */
2338         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2339
2340         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2341
2342         while (retry_count) {
2343                 if (signal_pending(current)) {
2344                         ret = -EINTR;
2345                         break;
2346                 }
2347                 /*
2348                  * Rather than hide all in some function, I do this in
2349                  * open coded manner. You see what this really does.
2350                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2351                  */
2352                 mutex_lock(&set_limit_mutex);
2353                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2354                 if (memswlimit < val) {
2355                         ret = -EINVAL;
2356                         mutex_unlock(&set_limit_mutex);
2357                         break;
2358                 }
2359                 ret = res_counter_set_limit(&memcg->res, val);
2360                 if (!ret) {
2361                         if (memswlimit == val)
2362                                 memcg->memsw_is_minimum = true;
2363                         else
2364                                 memcg->memsw_is_minimum = false;
2365                 }
2366                 mutex_unlock(&set_limit_mutex);
2367
2368                 if (!ret)
2369                         break;
2370
2371                 progress = mem_cgroup_hierarchical_reclaim(memcg, NULL,
2372                                                 GFP_KERNEL,
2373                                                 MEM_CGROUP_RECLAIM_SHRINK);
2374                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2375                 /* Usage is reduced ? */
2376                 if (curusage >= oldusage)
2377                         retry_count--;
2378                 else
2379                         oldusage = curusage;
2380         }
2381
2382         return ret;
2383 }
2384
2385 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2386                                         unsigned long long val)
2387 {
2388         int retry_count;
2389         u64 memlimit, oldusage, curusage;
2390         int children = mem_cgroup_count_children(memcg);
2391         int ret = -EBUSY;
2392
2393         /* see mem_cgroup_resize_res_limit */
2394         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2395         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2396         while (retry_count) {
2397                 if (signal_pending(current)) {
2398                         ret = -EINTR;
2399                         break;
2400                 }
2401                 /*
2402                  * Rather than hide all in some function, I do this in
2403                  * open coded manner. You see what this really does.
2404                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2405                  */
2406                 mutex_lock(&set_limit_mutex);
2407                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2408                 if (memlimit > val) {
2409                         ret = -EINVAL;
2410                         mutex_unlock(&set_limit_mutex);
2411                         break;
2412                 }
2413                 ret = res_counter_set_limit(&memcg->memsw, val);
2414                 if (!ret) {
2415                         if (memlimit == val)
2416                                 memcg->memsw_is_minimum = true;
2417                         else
2418                                 memcg->memsw_is_minimum = false;
2419                 }
2420                 mutex_unlock(&set_limit_mutex);
2421
2422                 if (!ret)
2423                         break;
2424
2425                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2426                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2427                                                 MEM_CGROUP_RECLAIM_SHRINK);
2428                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2429                 /* Usage is reduced ? */
2430                 if (curusage >= oldusage)
2431                         retry_count--;
2432                 else
2433                         oldusage = curusage;
2434         }
2435         return ret;
2436 }
2437
2438 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2439                                                 gfp_t gfp_mask, int nid,
2440                                                 int zid)
2441 {
2442         unsigned long nr_reclaimed = 0;
2443         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2444         unsigned long reclaimed;
2445         int loop = 0;
2446         struct mem_cgroup_tree_per_zone *mctz;
2447         unsigned long long excess;
2448
2449         if (order > 0)
2450                 return 0;
2451
2452         mctz = soft_limit_tree_node_zone(nid, zid);
2453         /*
2454          * This loop can run a while, specially if mem_cgroup's continuously
2455          * keep exceeding their soft limit and putting the system under
2456          * pressure
2457          */
2458         do {
2459                 if (next_mz)
2460                         mz = next_mz;
2461                 else
2462                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2463                 if (!mz)
2464                         break;
2465
2466                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2467                                                 gfp_mask,
2468                                                 MEM_CGROUP_RECLAIM_SOFT);
2469                 nr_reclaimed += reclaimed;
2470                 spin_lock(&mctz->lock);
2471
2472                 /*
2473                  * If we failed to reclaim anything from this memory cgroup
2474                  * it is time to move on to the next cgroup
2475                  */
2476                 next_mz = NULL;
2477                 if (!reclaimed) {
2478                         do {
2479                                 /*
2480                                  * Loop until we find yet another one.
2481                                  *
2482                                  * By the time we get the soft_limit lock
2483                                  * again, someone might have aded the
2484                                  * group back on the RB tree. Iterate to
2485                                  * make sure we get a different mem.
2486                                  * mem_cgroup_largest_soft_limit_node returns
2487                                  * NULL if no other cgroup is present on
2488                                  * the tree
2489                                  */
2490                                 next_mz =
2491                                 __mem_cgroup_largest_soft_limit_node(mctz);
2492                                 if (next_mz == mz) {
2493                                         css_put(&next_mz->mem->css);
2494                                         next_mz = NULL;
2495                                 } else /* next_mz == NULL or other memcg */
2496                                         break;
2497                         } while (1);
2498                 }
2499                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2500                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2501                 /*
2502                  * One school of thought says that we should not add
2503                  * back the node to the tree if reclaim returns 0.
2504                  * But our reclaim could return 0, simply because due
2505                  * to priority we are exposing a smaller subset of
2506                  * memory to reclaim from. Consider this as a longer
2507                  * term TODO.
2508                  */
2509                 /* If excess == 0, no tree ops */
2510                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2511                 spin_unlock(&mctz->lock);
2512                 css_put(&mz->mem->css);
2513                 loop++;
2514                 /*
2515                  * Could not reclaim anything and there are no more
2516                  * mem cgroups to try or we seem to be looping without
2517                  * reclaiming anything.
2518                  */
2519                 if (!nr_reclaimed &&
2520                         (next_mz == NULL ||
2521                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2522                         break;
2523         } while (!nr_reclaimed);
2524         if (next_mz)
2525                 css_put(&next_mz->mem->css);
2526         return nr_reclaimed;
2527 }
2528
2529 /*
2530  * This routine traverse page_cgroup in given list and drop them all.
2531  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2532  */
2533 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2534                                 int node, int zid, enum lru_list lru)
2535 {
2536         struct zone *zone;
2537         struct mem_cgroup_per_zone *mz;
2538         struct page_cgroup *pc, *busy;
2539         unsigned long flags, loop;
2540         struct list_head *list;
2541         int ret = 0;
2542
2543         zone = &NODE_DATA(node)->node_zones[zid];
2544         mz = mem_cgroup_zoneinfo(mem, node, zid);
2545         list = &mz->lists[lru];
2546
2547         loop = MEM_CGROUP_ZSTAT(mz, lru);
2548         /* give some margin against EBUSY etc...*/
2549         loop += 256;
2550         busy = NULL;
2551         while (loop--) {
2552                 ret = 0;
2553                 spin_lock_irqsave(&zone->lru_lock, flags);
2554                 if (list_empty(list)) {
2555                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2556                         break;
2557                 }
2558                 pc = list_entry(list->prev, struct page_cgroup, lru);
2559                 if (busy == pc) {
2560                         list_move(&pc->lru, list);
2561                         busy = 0;
2562                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2563                         continue;
2564                 }
2565                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2566
2567                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2568                 if (ret == -ENOMEM)
2569                         break;
2570
2571                 if (ret == -EBUSY || ret == -EINVAL) {
2572                         /* found lock contention or "pc" is obsolete. */
2573                         busy = pc;
2574                         cond_resched();
2575                 } else
2576                         busy = NULL;
2577         }
2578
2579         if (!ret && !list_empty(list))
2580                 return -EBUSY;
2581         return ret;
2582 }
2583
2584 /*
2585  * make mem_cgroup's charge to be 0 if there is no task.
2586  * This enables deleting this mem_cgroup.
2587  */
2588 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2589 {
2590         int ret;
2591         int node, zid, shrink;
2592         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2593         struct cgroup *cgrp = mem->css.cgroup;
2594
2595         css_get(&mem->css);
2596
2597         shrink = 0;
2598         /* should free all ? */
2599         if (free_all)
2600                 goto try_to_free;
2601 move_account:
2602         while (mem->res.usage > 0) {
2603                 ret = -EBUSY;
2604                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2605                         goto out;
2606                 ret = -EINTR;
2607                 if (signal_pending(current))
2608                         goto out;
2609                 /* This is for making all *used* pages to be on LRU. */
2610                 lru_add_drain_all();
2611                 drain_all_stock_sync();
2612                 ret = 0;
2613                 for_each_node_state(node, N_HIGH_MEMORY) {
2614                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2615                                 enum lru_list l;
2616                                 for_each_lru(l) {
2617                                         ret = mem_cgroup_force_empty_list(mem,
2618                                                         node, zid, l);
2619                                         if (ret)
2620                                                 break;
2621                                 }
2622                         }
2623                         if (ret)
2624                                 break;
2625                 }
2626                 /* it seems parent cgroup doesn't have enough mem */
2627                 if (ret == -ENOMEM)
2628                         goto try_to_free;
2629                 cond_resched();
2630         }
2631         ret = 0;
2632 out:
2633         css_put(&mem->css);
2634         return ret;
2635
2636 try_to_free:
2637         /* returns EBUSY if there is a task or if we come here twice. */
2638         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2639                 ret = -EBUSY;
2640                 goto out;
2641         }
2642         /* we call try-to-free pages for make this cgroup empty */
2643         lru_add_drain_all();
2644         /* try to free all pages in this cgroup */
2645         shrink = 1;
2646         while (nr_retries && mem->res.usage > 0) {
2647                 int progress;
2648
2649                 if (signal_pending(current)) {
2650                         ret = -EINTR;
2651                         goto out;
2652                 }
2653                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2654                                                 false, get_swappiness(mem));
2655                 if (!progress) {
2656                         nr_retries--;
2657                         /* maybe some writeback is necessary */
2658                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2659                 }
2660
2661         }
2662         lru_add_drain();
2663         /* try move_account...there may be some *locked* pages. */
2664         if (mem->res.usage)
2665                 goto move_account;
2666         ret = 0;
2667         goto out;
2668 }
2669
2670 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2671 {
2672         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2673 }
2674
2675
2676 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2677 {
2678         return mem_cgroup_from_cont(cont)->use_hierarchy;
2679 }
2680
2681 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2682                                         u64 val)
2683 {
2684         int retval = 0;
2685         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2686         struct cgroup *parent = cont->parent;
2687         struct mem_cgroup *parent_mem = NULL;
2688
2689         if (parent)
2690                 parent_mem = mem_cgroup_from_cont(parent);
2691
2692         cgroup_lock();
2693         /*
2694          * If parent's use_hierarchy is set, we can't make any modifications
2695          * in the child subtrees. If it is unset, then the change can
2696          * occur, provided the current cgroup has no children.
2697          *
2698          * For the root cgroup, parent_mem is NULL, we allow value to be
2699          * set if there are no children.
2700          */
2701         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2702                                 (val == 1 || val == 0)) {
2703                 if (list_empty(&cont->children))
2704                         mem->use_hierarchy = val;
2705                 else
2706                         retval = -EBUSY;
2707         } else
2708                 retval = -EINVAL;
2709         cgroup_unlock();
2710
2711         return retval;
2712 }
2713
2714 struct mem_cgroup_idx_data {
2715         s64 val;
2716         enum mem_cgroup_stat_index idx;
2717 };
2718
2719 static int
2720 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2721 {
2722         struct mem_cgroup_idx_data *d = data;
2723         d->val += mem_cgroup_read_stat(&mem->stat, d->idx);
2724         return 0;
2725 }
2726
2727 static void
2728 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2729                                 enum mem_cgroup_stat_index idx, s64 *val)
2730 {
2731         struct mem_cgroup_idx_data d;
2732         d.idx = idx;
2733         d.val = 0;
2734         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2735         *val = d.val;
2736 }
2737
2738 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2739 {
2740         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2741         u64 idx_val, val;
2742         int type, name;
2743
2744         type = MEMFILE_TYPE(cft->private);
2745         name = MEMFILE_ATTR(cft->private);
2746         switch (type) {
2747         case _MEM:
2748                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2749                         mem_cgroup_get_recursive_idx_stat(mem,
2750                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2751                         val = idx_val;
2752                         mem_cgroup_get_recursive_idx_stat(mem,
2753                                 MEM_CGROUP_STAT_RSS, &idx_val);
2754                         val += idx_val;
2755                         val <<= PAGE_SHIFT;
2756                 } else
2757                         val = res_counter_read_u64(&mem->res, name);
2758                 break;
2759         case _MEMSWAP:
2760                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2761                         mem_cgroup_get_recursive_idx_stat(mem,
2762                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2763                         val = idx_val;
2764                         mem_cgroup_get_recursive_idx_stat(mem,
2765                                 MEM_CGROUP_STAT_RSS, &idx_val);
2766                         val += idx_val;
2767                         mem_cgroup_get_recursive_idx_stat(mem,
2768                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2769                         val += idx_val;
2770                         val <<= PAGE_SHIFT;
2771                 } else
2772                         val = res_counter_read_u64(&mem->memsw, name);
2773                 break;
2774         default:
2775                 BUG();
2776                 break;
2777         }
2778         return val;
2779 }
2780 /*
2781  * The user of this function is...
2782  * RES_LIMIT.
2783  */
2784 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2785                             const char *buffer)
2786 {
2787         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2788         int type, name;
2789         unsigned long long val;
2790         int ret;
2791
2792         type = MEMFILE_TYPE(cft->private);
2793         name = MEMFILE_ATTR(cft->private);
2794         switch (name) {
2795         case RES_LIMIT:
2796                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2797                         ret = -EINVAL;
2798                         break;
2799                 }
2800                 /* This function does all necessary parse...reuse it */
2801                 ret = res_counter_memparse_write_strategy(buffer, &val);
2802                 if (ret)
2803                         break;
2804                 if (type == _MEM)
2805                         ret = mem_cgroup_resize_limit(memcg, val);
2806                 else
2807                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2808                 break;
2809         case RES_SOFT_LIMIT:
2810                 ret = res_counter_memparse_write_strategy(buffer, &val);
2811                 if (ret)
2812                         break;
2813                 /*
2814                  * For memsw, soft limits are hard to implement in terms
2815                  * of semantics, for now, we support soft limits for
2816                  * control without swap
2817                  */
2818                 if (type == _MEM)
2819                         ret = res_counter_set_soft_limit(&memcg->res, val);
2820                 else
2821                         ret = -EINVAL;
2822                 break;
2823         default:
2824                 ret = -EINVAL; /* should be BUG() ? */
2825                 break;
2826         }
2827         return ret;
2828 }
2829
2830 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2831                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2832 {
2833         struct cgroup *cgroup;
2834         unsigned long long min_limit, min_memsw_limit, tmp;
2835
2836         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2837         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2838         cgroup = memcg->css.cgroup;
2839         if (!memcg->use_hierarchy)
2840                 goto out;
2841
2842         while (cgroup->parent) {
2843                 cgroup = cgroup->parent;
2844                 memcg = mem_cgroup_from_cont(cgroup);
2845                 if (!memcg->use_hierarchy)
2846                         break;
2847                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2848                 min_limit = min(min_limit, tmp);
2849                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2850                 min_memsw_limit = min(min_memsw_limit, tmp);
2851         }
2852 out:
2853         *mem_limit = min_limit;
2854         *memsw_limit = min_memsw_limit;
2855         return;
2856 }
2857
2858 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2859 {
2860         struct mem_cgroup *mem;
2861         int type, name;
2862
2863         mem = mem_cgroup_from_cont(cont);
2864         type = MEMFILE_TYPE(event);
2865         name = MEMFILE_ATTR(event);
2866         switch (name) {
2867         case RES_MAX_USAGE:
2868                 if (type == _MEM)
2869                         res_counter_reset_max(&mem->res);
2870                 else
2871                         res_counter_reset_max(&mem->memsw);
2872                 break;
2873         case RES_FAILCNT:
2874                 if (type == _MEM)
2875                         res_counter_reset_failcnt(&mem->res);
2876                 else
2877                         res_counter_reset_failcnt(&mem->memsw);
2878                 break;
2879         }
2880
2881         return 0;
2882 }
2883
2884
2885 /* For read statistics */
2886 enum {
2887         MCS_CACHE,
2888         MCS_RSS,
2889         MCS_FILE_MAPPED,
2890         MCS_PGPGIN,
2891         MCS_PGPGOUT,
2892         MCS_SWAP,
2893         MCS_INACTIVE_ANON,
2894         MCS_ACTIVE_ANON,
2895         MCS_INACTIVE_FILE,
2896         MCS_ACTIVE_FILE,
2897         MCS_UNEVICTABLE,
2898         NR_MCS_STAT,
2899 };
2900
2901 struct mcs_total_stat {
2902         s64 stat[NR_MCS_STAT];
2903 };
2904
2905 struct {
2906         char *local_name;
2907         char *total_name;
2908 } memcg_stat_strings[NR_MCS_STAT] = {
2909         {"cache", "total_cache"},
2910         {"rss", "total_rss"},
2911         {"mapped_file", "total_mapped_file"},
2912         {"pgpgin", "total_pgpgin"},
2913         {"pgpgout", "total_pgpgout"},
2914         {"swap", "total_swap"},
2915         {"inactive_anon", "total_inactive_anon"},
2916         {"active_anon", "total_active_anon"},
2917         {"inactive_file", "total_inactive_file"},
2918         {"active_file", "total_active_file"},
2919         {"unevictable", "total_unevictable"}
2920 };
2921
2922
2923 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2924 {
2925         struct mcs_total_stat *s = data;
2926         s64 val;
2927
2928         /* per cpu stat */
2929         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2930         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2931         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2932         s->stat[MCS_RSS] += val * PAGE_SIZE;
2933         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_FILE_MAPPED);
2934         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
2935         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2936         s->stat[MCS_PGPGIN] += val;
2937         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2938         s->stat[MCS_PGPGOUT] += val;
2939         if (do_swap_account) {
2940                 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT);
2941                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
2942         }
2943
2944         /* per zone stat */
2945         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2946         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2947         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2948         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2949         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2950         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2951         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2952         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2953         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2954         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2955         return 0;
2956 }
2957
2958 static void
2959 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2960 {
2961         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2962 }
2963
2964 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2965                                  struct cgroup_map_cb *cb)
2966 {
2967         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2968         struct mcs_total_stat mystat;
2969         int i;
2970
2971         memset(&mystat, 0, sizeof(mystat));
2972         mem_cgroup_get_local_stat(mem_cont, &mystat);
2973
2974         for (i = 0; i < NR_MCS_STAT; i++) {
2975                 if (i == MCS_SWAP && !do_swap_account)
2976                         continue;
2977                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2978         }
2979
2980         /* Hierarchical information */
2981         {
2982                 unsigned long long limit, memsw_limit;
2983                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2984                 cb->fill(cb, "hierarchical_memory_limit", limit);
2985                 if (do_swap_account)
2986                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2987         }
2988
2989         memset(&mystat, 0, sizeof(mystat));
2990         mem_cgroup_get_total_stat(mem_cont, &mystat);
2991         for (i = 0; i < NR_MCS_STAT; i++) {
2992                 if (i == MCS_SWAP && !do_swap_account)
2993                         continue;
2994                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2995         }
2996
2997 #ifdef CONFIG_DEBUG_VM
2998         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2999
3000         {
3001                 int nid, zid;
3002                 struct mem_cgroup_per_zone *mz;
3003                 unsigned long recent_rotated[2] = {0, 0};
3004                 unsigned long recent_scanned[2] = {0, 0};
3005
3006                 for_each_online_node(nid)
3007                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3008                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3009
3010                                 recent_rotated[0] +=
3011                                         mz->reclaim_stat.recent_rotated[0];
3012                                 recent_rotated[1] +=
3013                                         mz->reclaim_stat.recent_rotated[1];
3014                                 recent_scanned[0] +=
3015                                         mz->reclaim_stat.recent_scanned[0];
3016                                 recent_scanned[1] +=
3017                                         mz->reclaim_stat.recent_scanned[1];
3018                         }
3019                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3020                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3021                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3022                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3023         }
3024 #endif
3025
3026         return 0;
3027 }
3028
3029 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3030 {
3031         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3032
3033         return get_swappiness(memcg);
3034 }
3035
3036 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3037                                        u64 val)
3038 {
3039         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3040         struct mem_cgroup *parent;
3041
3042         if (val > 100)
3043                 return -EINVAL;
3044
3045         if (cgrp->parent == NULL)
3046                 return -EINVAL;
3047
3048         parent = mem_cgroup_from_cont(cgrp->parent);
3049
3050         cgroup_lock();
3051
3052         /* If under hierarchy, only empty-root can set this value */
3053         if ((parent->use_hierarchy) ||
3054             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3055                 cgroup_unlock();
3056                 return -EINVAL;
3057         }
3058
3059         spin_lock(&memcg->reclaim_param_lock);
3060         memcg->swappiness = val;
3061         spin_unlock(&memcg->reclaim_param_lock);
3062
3063         cgroup_unlock();
3064
3065         return 0;
3066 }
3067
3068
3069 static struct cftype mem_cgroup_files[] = {
3070         {
3071                 .name = "usage_in_bytes",
3072                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3073                 .read_u64 = mem_cgroup_read,
3074         },
3075         {
3076                 .name = "max_usage_in_bytes",
3077                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3078                 .trigger = mem_cgroup_reset,
3079                 .read_u64 = mem_cgroup_read,
3080         },
3081         {
3082                 .name = "limit_in_bytes",
3083                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3084                 .write_string = mem_cgroup_write,
3085                 .read_u64 = mem_cgroup_read,
3086         },
3087         {
3088                 .name = "soft_limit_in_bytes",
3089                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3090                 .write_string = mem_cgroup_write,
3091                 .read_u64 = mem_cgroup_read,
3092         },
3093         {
3094                 .name = "failcnt",
3095                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3096                 .trigger = mem_cgroup_reset,
3097                 .read_u64 = mem_cgroup_read,
3098         },
3099         {
3100                 .name = "stat",
3101                 .read_map = mem_control_stat_show,
3102         },
3103         {
3104                 .name = "force_empty",
3105                 .trigger = mem_cgroup_force_empty_write,
3106         },
3107         {
3108                 .name = "use_hierarchy",
3109                 .write_u64 = mem_cgroup_hierarchy_write,
3110                 .read_u64 = mem_cgroup_hierarchy_read,
3111         },
3112         {
3113                 .name = "swappiness",
3114                 .read_u64 = mem_cgroup_swappiness_read,
3115                 .write_u64 = mem_cgroup_swappiness_write,
3116         },
3117 };
3118
3119 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3120 static struct cftype memsw_cgroup_files[] = {
3121         {
3122                 .name = "memsw.usage_in_bytes",
3123                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3124                 .read_u64 = mem_cgroup_read,
3125         },
3126         {
3127                 .name = "memsw.max_usage_in_bytes",
3128                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3129                 .trigger = mem_cgroup_reset,
3130                 .read_u64 = mem_cgroup_read,
3131         },
3132         {
3133                 .name = "memsw.limit_in_bytes",
3134                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3135                 .write_string = mem_cgroup_write,
3136                 .read_u64 = mem_cgroup_read,
3137         },
3138         {
3139                 .name = "memsw.failcnt",
3140                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3141                 .trigger = mem_cgroup_reset,
3142                 .read_u64 = mem_cgroup_read,
3143         },
3144 };
3145
3146 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3147 {
3148         if (!do_swap_account)
3149                 return 0;
3150         return cgroup_add_files(cont, ss, memsw_cgroup_files,
3151                                 ARRAY_SIZE(memsw_cgroup_files));
3152 };
3153 #else
3154 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3155 {
3156         return 0;
3157 }
3158 #endif
3159
3160 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3161 {
3162         struct mem_cgroup_per_node *pn;
3163         struct mem_cgroup_per_zone *mz;
3164         enum lru_list l;
3165         int zone, tmp = node;
3166         /*
3167          * This routine is called against possible nodes.
3168          * But it's BUG to call kmalloc() against offline node.
3169          *
3170          * TODO: this routine can waste much memory for nodes which will
3171          *       never be onlined. It's better to use memory hotplug callback
3172          *       function.
3173          */
3174         if (!node_state(node, N_NORMAL_MEMORY))
3175                 tmp = -1;
3176         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3177         if (!pn)
3178                 return 1;
3179
3180         mem->info.nodeinfo[node] = pn;
3181         memset(pn, 0, sizeof(*pn));
3182
3183         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3184                 mz = &pn->zoneinfo[zone];
3185                 for_each_lru(l)
3186                         INIT_LIST_HEAD(&mz->lists[l]);
3187                 mz->usage_in_excess = 0;
3188                 mz->on_tree = false;
3189                 mz->mem = mem;
3190         }
3191         return 0;
3192 }
3193
3194 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3195 {
3196         kfree(mem->info.nodeinfo[node]);
3197 }
3198
3199 static int mem_cgroup_size(void)
3200 {
3201         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
3202         return sizeof(struct mem_cgroup) + cpustat_size;
3203 }
3204
3205 static struct mem_cgroup *mem_cgroup_alloc(void)
3206 {
3207         struct mem_cgroup *mem;
3208         int size = mem_cgroup_size();
3209
3210         if (size < PAGE_SIZE)
3211                 mem = kmalloc(size, GFP_KERNEL);
3212         else
3213                 mem = vmalloc(size);
3214
3215         if (mem)
3216                 memset(mem, 0, size);
3217         return mem;
3218 }
3219
3220 /*
3221  * At destroying mem_cgroup, references from swap_cgroup can remain.
3222  * (scanning all at force_empty is too costly...)
3223  *
3224  * Instead of clearing all references at force_empty, we remember
3225  * the number of reference from swap_cgroup and free mem_cgroup when
3226  * it goes down to 0.
3227  *
3228  * Removal of cgroup itself succeeds regardless of refs from swap.
3229  */
3230
3231 static void __mem_cgroup_free(struct mem_cgroup *mem)
3232 {
3233         int node;
3234
3235         mem_cgroup_remove_from_trees(mem);
3236         free_css_id(&mem_cgroup_subsys, &mem->css);
3237
3238         for_each_node_state(node, N_POSSIBLE)
3239                 free_mem_cgroup_per_zone_info(mem, node);
3240
3241         if (mem_cgroup_size() < PAGE_SIZE)
3242                 kfree(mem);
3243         else
3244                 vfree(mem);
3245 }
3246
3247 static void mem_cgroup_get(struct mem_cgroup *mem)
3248 {
3249         atomic_inc(&mem->refcnt);
3250 }
3251
3252 static void mem_cgroup_put(struct mem_cgroup *mem)
3253 {
3254         if (atomic_dec_and_test(&mem->refcnt)) {
3255                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3256                 __mem_cgroup_free(mem);
3257                 if (parent)
3258                         mem_cgroup_put(parent);
3259         }
3260 }
3261
3262 /*
3263  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3264  */
3265 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3266 {
3267         if (!mem->res.parent)
3268                 return NULL;
3269         return mem_cgroup_from_res_counter(mem->res.parent, res);
3270 }
3271
3272 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3273 static void __init enable_swap_cgroup(void)
3274 {
3275         if (!mem_cgroup_disabled() && really_do_swap_account)
3276                 do_swap_account = 1;
3277 }
3278 #else
3279 static void __init enable_swap_cgroup(void)
3280 {
3281 }
3282 #endif
3283
3284 static int mem_cgroup_soft_limit_tree_init(void)
3285 {
3286         struct mem_cgroup_tree_per_node *rtpn;
3287         struct mem_cgroup_tree_per_zone *rtpz;
3288         int tmp, node, zone;
3289
3290         for_each_node_state(node, N_POSSIBLE) {
3291                 tmp = node;
3292                 if (!node_state(node, N_NORMAL_MEMORY))
3293                         tmp = -1;
3294                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3295                 if (!rtpn)
3296                         return 1;
3297
3298                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3299
3300                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3301                         rtpz = &rtpn->rb_tree_per_zone[zone];
3302                         rtpz->rb_root = RB_ROOT;
3303                         spin_lock_init(&rtpz->lock);
3304                 }
3305         }
3306         return 0;
3307 }
3308
3309 static struct cgroup_subsys_state * __ref
3310 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3311 {
3312         struct mem_cgroup *mem, *parent;
3313         long error = -ENOMEM;
3314         int node;
3315
3316         mem = mem_cgroup_alloc();
3317         if (!mem)
3318                 return ERR_PTR(error);
3319
3320         for_each_node_state(node, N_POSSIBLE)
3321                 if (alloc_mem_cgroup_per_zone_info(mem, node))
3322                         goto free_out;
3323
3324         /* root ? */
3325         if (cont->parent == NULL) {
3326                 int cpu;
3327                 enable_swap_cgroup();
3328                 parent = NULL;
3329                 root_mem_cgroup = mem;
3330                 if (mem_cgroup_soft_limit_tree_init())
3331                         goto free_out;
3332                 for_each_possible_cpu(cpu) {
3333                         struct memcg_stock_pcp *stock =
3334                                                 &per_cpu(memcg_stock, cpu);
3335                         INIT_WORK(&stock->work, drain_local_stock);
3336                 }
3337                 hotcpu_notifier(memcg_stock_cpu_callback, 0);
3338
3339         } else {
3340                 parent = mem_cgroup_from_cont(cont->parent);
3341                 mem->use_hierarchy = parent->use_hierarchy;
3342         }
3343
3344         if (parent && parent->use_hierarchy) {
3345                 res_counter_init(&mem->res, &parent->res);
3346                 res_counter_init(&mem->memsw, &parent->memsw);
3347                 /*
3348                  * We increment refcnt of the parent to ensure that we can
3349                  * safely access it on res_counter_charge/uncharge.
3350                  * This refcnt will be decremented when freeing this