memcg: fix swap accounting leak
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  */
19
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/mutex.h>
31 #include <linux/slab.h>
32 #include <linux/swap.h>
33 #include <linux/spinlock.h>
34 #include <linux/fs.h>
35 #include <linux/seq_file.h>
36 #include <linux/vmalloc.h>
37 #include <linux/mm_inline.h>
38 #include <linux/page_cgroup.h>
39 #include "internal.h"
40
41 #include <asm/uaccess.h>
42
43 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
44 #define MEM_CGROUP_RECLAIM_RETRIES      5
45
46 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
47 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 0 */
48 int do_swap_account __read_mostly;
49 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
50 #else
51 #define do_swap_account         (0)
52 #endif
53
54 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
55
56 /*
57  * Statistics for memory cgroup.
58  */
59 enum mem_cgroup_stat_index {
60         /*
61          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
62          */
63         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
64         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
65         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
66         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
67
68         MEM_CGROUP_STAT_NSTATS,
69 };
70
71 struct mem_cgroup_stat_cpu {
72         s64 count[MEM_CGROUP_STAT_NSTATS];
73 } ____cacheline_aligned_in_smp;
74
75 struct mem_cgroup_stat {
76         struct mem_cgroup_stat_cpu cpustat[0];
77 };
78
79 /*
80  * For accounting under irq disable, no need for increment preempt count.
81  */
82 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
83                 enum mem_cgroup_stat_index idx, int val)
84 {
85         stat->count[idx] += val;
86 }
87
88 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
89                 enum mem_cgroup_stat_index idx)
90 {
91         int cpu;
92         s64 ret = 0;
93         for_each_possible_cpu(cpu)
94                 ret += stat->cpustat[cpu].count[idx];
95         return ret;
96 }
97
98 /*
99  * per-zone information in memory controller.
100  */
101 struct mem_cgroup_per_zone {
102         /*
103          * spin_lock to protect the per cgroup LRU
104          */
105         struct list_head        lists[NR_LRU_LISTS];
106         unsigned long           count[NR_LRU_LISTS];
107
108         struct zone_reclaim_stat reclaim_stat;
109 };
110 /* Macro for accessing counter */
111 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
112
113 struct mem_cgroup_per_node {
114         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
115 };
116
117 struct mem_cgroup_lru_info {
118         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
119 };
120
121 /*
122  * The memory controller data structure. The memory controller controls both
123  * page cache and RSS per cgroup. We would eventually like to provide
124  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
125  * to help the administrator determine what knobs to tune.
126  *
127  * TODO: Add a water mark for the memory controller. Reclaim will begin when
128  * we hit the water mark. May be even add a low water mark, such that
129  * no reclaim occurs from a cgroup at it's low water mark, this is
130  * a feature that will be implemented much later in the future.
131  */
132 struct mem_cgroup {
133         struct cgroup_subsys_state css;
134         /*
135          * the counter to account for memory usage
136          */
137         struct res_counter res;
138         /*
139          * the counter to account for mem+swap usage.
140          */
141         struct res_counter memsw;
142         /*
143          * Per cgroup active and inactive list, similar to the
144          * per zone LRU lists.
145          */
146         struct mem_cgroup_lru_info info;
147
148         /*
149           protect against reclaim related member.
150         */
151         spinlock_t reclaim_param_lock;
152
153         int     prev_priority;  /* for recording reclaim priority */
154
155         /*
156          * While reclaiming in a hiearchy, we cache the last child we
157          * reclaimed from. Protected by cgroup_lock()
158          */
159         struct mem_cgroup *last_scanned_child;
160         /*
161          * Should the accounting and control be hierarchical, per subtree?
162          */
163         bool use_hierarchy;
164         unsigned long   last_oom_jiffies;
165         int             obsolete;
166         atomic_t        refcnt;
167
168         unsigned int    swappiness;
169
170         /*
171          * statistics. This must be placed at the end of memcg.
172          */
173         struct mem_cgroup_stat stat;
174 };
175
176 enum charge_type {
177         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
178         MEM_CGROUP_CHARGE_TYPE_MAPPED,
179         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
180         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
181         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
182         NR_CHARGE_TYPE,
183 };
184
185 /* only for here (for easy reading.) */
186 #define PCGF_CACHE      (1UL << PCG_CACHE)
187 #define PCGF_USED       (1UL << PCG_USED)
188 #define PCGF_LOCK       (1UL << PCG_LOCK)
189 static const unsigned long
190 pcg_default_flags[NR_CHARGE_TYPE] = {
191         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
192         PCGF_USED | PCGF_LOCK, /* Anon */
193         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
194         0, /* FORCE */
195 };
196
197 /* for encoding cft->private value on file */
198 #define _MEM                    (0)
199 #define _MEMSWAP                (1)
200 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
201 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
202 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
203
204 static void mem_cgroup_get(struct mem_cgroup *mem);
205 static void mem_cgroup_put(struct mem_cgroup *mem);
206
207 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
208                                          struct page_cgroup *pc,
209                                          bool charge)
210 {
211         int val = (charge)? 1 : -1;
212         struct mem_cgroup_stat *stat = &mem->stat;
213         struct mem_cgroup_stat_cpu *cpustat;
214         int cpu = get_cpu();
215
216         cpustat = &stat->cpustat[cpu];
217         if (PageCgroupCache(pc))
218                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
219         else
220                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
221
222         if (charge)
223                 __mem_cgroup_stat_add_safe(cpustat,
224                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
225         else
226                 __mem_cgroup_stat_add_safe(cpustat,
227                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
228         put_cpu();
229 }
230
231 static struct mem_cgroup_per_zone *
232 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
233 {
234         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
235 }
236
237 static struct mem_cgroup_per_zone *
238 page_cgroup_zoneinfo(struct page_cgroup *pc)
239 {
240         struct mem_cgroup *mem = pc->mem_cgroup;
241         int nid = page_cgroup_nid(pc);
242         int zid = page_cgroup_zid(pc);
243
244         if (!mem)
245                 return NULL;
246
247         return mem_cgroup_zoneinfo(mem, nid, zid);
248 }
249
250 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
251                                         enum lru_list idx)
252 {
253         int nid, zid;
254         struct mem_cgroup_per_zone *mz;
255         u64 total = 0;
256
257         for_each_online_node(nid)
258                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
259                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
260                         total += MEM_CGROUP_ZSTAT(mz, idx);
261                 }
262         return total;
263 }
264
265 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
266 {
267         return container_of(cgroup_subsys_state(cont,
268                                 mem_cgroup_subsys_id), struct mem_cgroup,
269                                 css);
270 }
271
272 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
273 {
274         /*
275          * mm_update_next_owner() may clear mm->owner to NULL
276          * if it races with swapoff, page migration, etc.
277          * So this can be called with p == NULL.
278          */
279         if (unlikely(!p))
280                 return NULL;
281
282         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
283                                 struct mem_cgroup, css);
284 }
285
286 /*
287  * Following LRU functions are allowed to be used without PCG_LOCK.
288  * Operations are called by routine of global LRU independently from memcg.
289  * What we have to take care of here is validness of pc->mem_cgroup.
290  *
291  * Changes to pc->mem_cgroup happens when
292  * 1. charge
293  * 2. moving account
294  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
295  * It is added to LRU before charge.
296  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
297  * When moving account, the page is not on LRU. It's isolated.
298  */
299
300 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
301 {
302         struct page_cgroup *pc;
303         struct mem_cgroup *mem;
304         struct mem_cgroup_per_zone *mz;
305
306         if (mem_cgroup_disabled())
307                 return;
308         pc = lookup_page_cgroup(page);
309         /* can happen while we handle swapcache. */
310         if (list_empty(&pc->lru))
311                 return;
312         mz = page_cgroup_zoneinfo(pc);
313         mem = pc->mem_cgroup;
314         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
315         list_del_init(&pc->lru);
316         return;
317 }
318
319 void mem_cgroup_del_lru(struct page *page)
320 {
321         mem_cgroup_del_lru_list(page, page_lru(page));
322 }
323
324 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
325 {
326         struct mem_cgroup_per_zone *mz;
327         struct page_cgroup *pc;
328
329         if (mem_cgroup_disabled())
330                 return;
331
332         pc = lookup_page_cgroup(page);
333         smp_rmb();
334         /* unused page is not rotated. */
335         if (!PageCgroupUsed(pc))
336                 return;
337         mz = page_cgroup_zoneinfo(pc);
338         list_move(&pc->lru, &mz->lists[lru]);
339 }
340
341 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
342 {
343         struct page_cgroup *pc;
344         struct mem_cgroup_per_zone *mz;
345
346         if (mem_cgroup_disabled())
347                 return;
348         pc = lookup_page_cgroup(page);
349         /* barrier to sync with "charge" */
350         smp_rmb();
351         if (!PageCgroupUsed(pc))
352                 return;
353
354         mz = page_cgroup_zoneinfo(pc);
355         MEM_CGROUP_ZSTAT(mz, lru) += 1;
356         list_add(&pc->lru, &mz->lists[lru]);
357 }
358 /*
359  * To add swapcache into LRU. Be careful to all this function.
360  * zone->lru_lock shouldn't be held and irq must not be disabled.
361  */
362 static void mem_cgroup_lru_fixup(struct page *page)
363 {
364         if (!isolate_lru_page(page))
365                 putback_lru_page(page);
366 }
367
368 void mem_cgroup_move_lists(struct page *page,
369                            enum lru_list from, enum lru_list to)
370 {
371         if (mem_cgroup_disabled())
372                 return;
373         mem_cgroup_del_lru_list(page, from);
374         mem_cgroup_add_lru_list(page, to);
375 }
376
377 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
378 {
379         int ret;
380
381         task_lock(task);
382         ret = task->mm && mm_match_cgroup(task->mm, mem);
383         task_unlock(task);
384         return ret;
385 }
386
387 /*
388  * Calculate mapped_ratio under memory controller. This will be used in
389  * vmscan.c for deteremining we have to reclaim mapped pages.
390  */
391 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
392 {
393         long total, rss;
394
395         /*
396          * usage is recorded in bytes. But, here, we assume the number of
397          * physical pages can be represented by "long" on any arch.
398          */
399         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
400         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
401         return (int)((rss * 100L) / total);
402 }
403
404 /*
405  * prev_priority control...this will be used in memory reclaim path.
406  */
407 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
408 {
409         int prev_priority;
410
411         spin_lock(&mem->reclaim_param_lock);
412         prev_priority = mem->prev_priority;
413         spin_unlock(&mem->reclaim_param_lock);
414
415         return prev_priority;
416 }
417
418 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
419 {
420         spin_lock(&mem->reclaim_param_lock);
421         if (priority < mem->prev_priority)
422                 mem->prev_priority = priority;
423         spin_unlock(&mem->reclaim_param_lock);
424 }
425
426 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
427 {
428         spin_lock(&mem->reclaim_param_lock);
429         mem->prev_priority = priority;
430         spin_unlock(&mem->reclaim_param_lock);
431 }
432
433 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
434 {
435         unsigned long active;
436         unsigned long inactive;
437         unsigned long gb;
438         unsigned long inactive_ratio;
439
440         inactive = mem_cgroup_get_all_zonestat(memcg, LRU_INACTIVE_ANON);
441         active = mem_cgroup_get_all_zonestat(memcg, LRU_ACTIVE_ANON);
442
443         gb = (inactive + active) >> (30 - PAGE_SHIFT);
444         if (gb)
445                 inactive_ratio = int_sqrt(10 * gb);
446         else
447                 inactive_ratio = 1;
448
449         if (present_pages) {
450                 present_pages[0] = inactive;
451                 present_pages[1] = active;
452         }
453
454         return inactive_ratio;
455 }
456
457 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
458 {
459         unsigned long active;
460         unsigned long inactive;
461         unsigned long present_pages[2];
462         unsigned long inactive_ratio;
463
464         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
465
466         inactive = present_pages[0];
467         active = present_pages[1];
468
469         if (inactive * inactive_ratio < active)
470                 return 1;
471
472         return 0;
473 }
474
475 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
476                                        struct zone *zone,
477                                        enum lru_list lru)
478 {
479         int nid = zone->zone_pgdat->node_id;
480         int zid = zone_idx(zone);
481         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
482
483         return MEM_CGROUP_ZSTAT(mz, lru);
484 }
485
486 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
487                                                       struct zone *zone)
488 {
489         int nid = zone->zone_pgdat->node_id;
490         int zid = zone_idx(zone);
491         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
492
493         return &mz->reclaim_stat;
494 }
495
496 struct zone_reclaim_stat *
497 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
498 {
499         struct page_cgroup *pc;
500         struct mem_cgroup_per_zone *mz;
501
502         if (mem_cgroup_disabled())
503                 return NULL;
504
505         pc = lookup_page_cgroup(page);
506         mz = page_cgroup_zoneinfo(pc);
507         if (!mz)
508                 return NULL;
509
510         return &mz->reclaim_stat;
511 }
512
513 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
514                                         struct list_head *dst,
515                                         unsigned long *scanned, int order,
516                                         int mode, struct zone *z,
517                                         struct mem_cgroup *mem_cont,
518                                         int active, int file)
519 {
520         unsigned long nr_taken = 0;
521         struct page *page;
522         unsigned long scan;
523         LIST_HEAD(pc_list);
524         struct list_head *src;
525         struct page_cgroup *pc, *tmp;
526         int nid = z->zone_pgdat->node_id;
527         int zid = zone_idx(z);
528         struct mem_cgroup_per_zone *mz;
529         int lru = LRU_FILE * !!file + !!active;
530
531         BUG_ON(!mem_cont);
532         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
533         src = &mz->lists[lru];
534
535         scan = 0;
536         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
537                 if (scan >= nr_to_scan)
538                         break;
539
540                 page = pc->page;
541                 if (unlikely(!PageCgroupUsed(pc)))
542                         continue;
543                 if (unlikely(!PageLRU(page)))
544                         continue;
545
546                 scan++;
547                 if (__isolate_lru_page(page, mode, file) == 0) {
548                         list_move(&page->lru, dst);
549                         nr_taken++;
550                 }
551         }
552
553         *scanned = scan;
554         return nr_taken;
555 }
556
557 #define mem_cgroup_from_res_counter(counter, member)    \
558         container_of(counter, struct mem_cgroup, member)
559
560 /*
561  * This routine finds the DFS walk successor. This routine should be
562  * called with cgroup_mutex held
563  */
564 static struct mem_cgroup *
565 mem_cgroup_get_next_node(struct mem_cgroup *curr, struct mem_cgroup *root_mem)
566 {
567         struct cgroup *cgroup, *curr_cgroup, *root_cgroup;
568
569         curr_cgroup = curr->css.cgroup;
570         root_cgroup = root_mem->css.cgroup;
571
572         if (!list_empty(&curr_cgroup->children)) {
573                 /*
574                  * Walk down to children
575                  */
576                 mem_cgroup_put(curr);
577                 cgroup = list_entry(curr_cgroup->children.next,
578                                                 struct cgroup, sibling);
579                 curr = mem_cgroup_from_cont(cgroup);
580                 mem_cgroup_get(curr);
581                 goto done;
582         }
583
584 visit_parent:
585         if (curr_cgroup == root_cgroup) {
586                 mem_cgroup_put(curr);
587                 curr = root_mem;
588                 mem_cgroup_get(curr);
589                 goto done;
590         }
591
592         /*
593          * Goto next sibling
594          */
595         if (curr_cgroup->sibling.next != &curr_cgroup->parent->children) {
596                 mem_cgroup_put(curr);
597                 cgroup = list_entry(curr_cgroup->sibling.next, struct cgroup,
598                                                 sibling);
599                 curr = mem_cgroup_from_cont(cgroup);
600                 mem_cgroup_get(curr);
601                 goto done;
602         }
603
604         /*
605          * Go up to next parent and next parent's sibling if need be
606          */
607         curr_cgroup = curr_cgroup->parent;
608         goto visit_parent;
609
610 done:
611         root_mem->last_scanned_child = curr;
612         return curr;
613 }
614
615 /*
616  * Visit the first child (need not be the first child as per the ordering
617  * of the cgroup list, since we track last_scanned_child) of @mem and use
618  * that to reclaim free pages from.
619  */
620 static struct mem_cgroup *
621 mem_cgroup_get_first_node(struct mem_cgroup *root_mem)
622 {
623         struct cgroup *cgroup;
624         struct mem_cgroup *ret;
625         bool obsolete = (root_mem->last_scanned_child &&
626                                 root_mem->last_scanned_child->obsolete);
627
628         /*
629          * Scan all children under the mem_cgroup mem
630          */
631         cgroup_lock();
632         if (list_empty(&root_mem->css.cgroup->children)) {
633                 ret = root_mem;
634                 goto done;
635         }
636
637         if (!root_mem->last_scanned_child || obsolete) {
638
639                 if (obsolete)
640                         mem_cgroup_put(root_mem->last_scanned_child);
641
642                 cgroup = list_first_entry(&root_mem->css.cgroup->children,
643                                 struct cgroup, sibling);
644                 ret = mem_cgroup_from_cont(cgroup);
645                 mem_cgroup_get(ret);
646         } else
647                 ret = mem_cgroup_get_next_node(root_mem->last_scanned_child,
648                                                 root_mem);
649
650 done:
651         root_mem->last_scanned_child = ret;
652         cgroup_unlock();
653         return ret;
654 }
655
656 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
657 {
658         if (do_swap_account) {
659                 if (res_counter_check_under_limit(&mem->res) &&
660                         res_counter_check_under_limit(&mem->memsw))
661                         return true;
662         } else
663                 if (res_counter_check_under_limit(&mem->res))
664                         return true;
665         return false;
666 }
667
668 static unsigned int get_swappiness(struct mem_cgroup *memcg)
669 {
670         struct cgroup *cgrp = memcg->css.cgroup;
671         unsigned int swappiness;
672
673         /* root ? */
674         if (cgrp->parent == NULL)
675                 return vm_swappiness;
676
677         spin_lock(&memcg->reclaim_param_lock);
678         swappiness = memcg->swappiness;
679         spin_unlock(&memcg->reclaim_param_lock);
680
681         return swappiness;
682 }
683
684 /*
685  * Dance down the hierarchy if needed to reclaim memory. We remember the
686  * last child we reclaimed from, so that we don't end up penalizing
687  * one child extensively based on its position in the children list.
688  *
689  * root_mem is the original ancestor that we've been reclaim from.
690  */
691 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
692                                                 gfp_t gfp_mask, bool noswap)
693 {
694         struct mem_cgroup *next_mem;
695         int ret = 0;
696
697         /*
698          * Reclaim unconditionally and don't check for return value.
699          * We need to reclaim in the current group and down the tree.
700          * One might think about checking for children before reclaiming,
701          * but there might be left over accounting, even after children
702          * have left.
703          */
704         ret = try_to_free_mem_cgroup_pages(root_mem, gfp_mask, noswap,
705                                            get_swappiness(root_mem));
706         if (mem_cgroup_check_under_limit(root_mem))
707                 return 0;
708         if (!root_mem->use_hierarchy)
709                 return ret;
710
711         next_mem = mem_cgroup_get_first_node(root_mem);
712
713         while (next_mem != root_mem) {
714                 if (next_mem->obsolete) {
715                         mem_cgroup_put(next_mem);
716                         cgroup_lock();
717                         next_mem = mem_cgroup_get_first_node(root_mem);
718                         cgroup_unlock();
719                         continue;
720                 }
721                 ret = try_to_free_mem_cgroup_pages(next_mem, gfp_mask, noswap,
722                                                    get_swappiness(next_mem));
723                 if (mem_cgroup_check_under_limit(root_mem))
724                         return 0;
725                 cgroup_lock();
726                 next_mem = mem_cgroup_get_next_node(next_mem, root_mem);
727                 cgroup_unlock();
728         }
729         return ret;
730 }
731
732 bool mem_cgroup_oom_called(struct task_struct *task)
733 {
734         bool ret = false;
735         struct mem_cgroup *mem;
736         struct mm_struct *mm;
737
738         rcu_read_lock();
739         mm = task->mm;
740         if (!mm)
741                 mm = &init_mm;
742         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
743         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
744                 ret = true;
745         rcu_read_unlock();
746         return ret;
747 }
748 /*
749  * Unlike exported interface, "oom" parameter is added. if oom==true,
750  * oom-killer can be invoked.
751  */
752 static int __mem_cgroup_try_charge(struct mm_struct *mm,
753                         gfp_t gfp_mask, struct mem_cgroup **memcg,
754                         bool oom)
755 {
756         struct mem_cgroup *mem, *mem_over_limit;
757         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
758         struct res_counter *fail_res;
759
760         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
761                 /* Don't account this! */
762                 *memcg = NULL;
763                 return 0;
764         }
765
766         /*
767          * We always charge the cgroup the mm_struct belongs to.
768          * The mm_struct's mem_cgroup changes on task migration if the
769          * thread group leader migrates. It's possible that mm is not
770          * set, if so charge the init_mm (happens for pagecache usage).
771          */
772         if (likely(!*memcg)) {
773                 rcu_read_lock();
774                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
775                 if (unlikely(!mem)) {
776                         rcu_read_unlock();
777                         return 0;
778                 }
779                 /*
780                  * For every charge from the cgroup, increment reference count
781                  */
782                 css_get(&mem->css);
783                 *memcg = mem;
784                 rcu_read_unlock();
785         } else {
786                 mem = *memcg;
787                 css_get(&mem->css);
788         }
789
790         while (1) {
791                 int ret;
792                 bool noswap = false;
793
794                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
795                 if (likely(!ret)) {
796                         if (!do_swap_account)
797                                 break;
798                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
799                                                         &fail_res);
800                         if (likely(!ret))
801                                 break;
802                         /* mem+swap counter fails */
803                         res_counter_uncharge(&mem->res, PAGE_SIZE);
804                         noswap = true;
805                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
806                                                                         memsw);
807                 } else
808                         /* mem counter fails */
809                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
810                                                                         res);
811
812                 if (!(gfp_mask & __GFP_WAIT))
813                         goto nomem;
814
815                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
816                                                         noswap);
817
818                 /*
819                  * try_to_free_mem_cgroup_pages() might not give us a full
820                  * picture of reclaim. Some pages are reclaimed and might be
821                  * moved to swap cache or just unmapped from the cgroup.
822                  * Check the limit again to see if the reclaim reduced the
823                  * current usage of the cgroup before giving up
824                  *
825                  */
826                 if (mem_cgroup_check_under_limit(mem_over_limit))
827                         continue;
828
829                 if (!nr_retries--) {
830                         if (oom) {
831                                 mutex_lock(&memcg_tasklist);
832                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
833                                 mutex_unlock(&memcg_tasklist);
834                                 mem_over_limit->last_oom_jiffies = jiffies;
835                         }
836                         goto nomem;
837                 }
838         }
839         return 0;
840 nomem:
841         css_put(&mem->css);
842         return -ENOMEM;
843 }
844
845 /*
846  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
847  * USED state. If already USED, uncharge and return.
848  */
849
850 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
851                                      struct page_cgroup *pc,
852                                      enum charge_type ctype)
853 {
854         /* try_charge() can return NULL to *memcg, taking care of it. */
855         if (!mem)
856                 return;
857
858         lock_page_cgroup(pc);
859         if (unlikely(PageCgroupUsed(pc))) {
860                 unlock_page_cgroup(pc);
861                 res_counter_uncharge(&mem->res, PAGE_SIZE);
862                 if (do_swap_account)
863                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
864                 css_put(&mem->css);
865                 return;
866         }
867         pc->mem_cgroup = mem;
868         smp_wmb();
869         pc->flags = pcg_default_flags[ctype];
870
871         mem_cgroup_charge_statistics(mem, pc, true);
872
873         unlock_page_cgroup(pc);
874 }
875
876 /**
877  * mem_cgroup_move_account - move account of the page
878  * @pc: page_cgroup of the page.
879  * @from: mem_cgroup which the page is moved from.
880  * @to: mem_cgroup which the page is moved to. @from != @to.
881  *
882  * The caller must confirm following.
883  * - page is not on LRU (isolate_page() is useful.)
884  *
885  * returns 0 at success,
886  * returns -EBUSY when lock is busy or "pc" is unstable.
887  *
888  * This function does "uncharge" from old cgroup but doesn't do "charge" to
889  * new cgroup. It should be done by a caller.
890  */
891
892 static int mem_cgroup_move_account(struct page_cgroup *pc,
893         struct mem_cgroup *from, struct mem_cgroup *to)
894 {
895         struct mem_cgroup_per_zone *from_mz, *to_mz;
896         int nid, zid;
897         int ret = -EBUSY;
898
899         VM_BUG_ON(from == to);
900         VM_BUG_ON(PageLRU(pc->page));
901
902         nid = page_cgroup_nid(pc);
903         zid = page_cgroup_zid(pc);
904         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
905         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
906
907         if (!trylock_page_cgroup(pc))
908                 return ret;
909
910         if (!PageCgroupUsed(pc))
911                 goto out;
912
913         if (pc->mem_cgroup != from)
914                 goto out;
915
916         css_put(&from->css);
917         res_counter_uncharge(&from->res, PAGE_SIZE);
918         mem_cgroup_charge_statistics(from, pc, false);
919         if (do_swap_account)
920                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
921         pc->mem_cgroup = to;
922         mem_cgroup_charge_statistics(to, pc, true);
923         css_get(&to->css);
924         ret = 0;
925 out:
926         unlock_page_cgroup(pc);
927         return ret;
928 }
929
930 /*
931  * move charges to its parent.
932  */
933
934 static int mem_cgroup_move_parent(struct page_cgroup *pc,
935                                   struct mem_cgroup *child,
936                                   gfp_t gfp_mask)
937 {
938         struct page *page = pc->page;
939         struct cgroup *cg = child->css.cgroup;
940         struct cgroup *pcg = cg->parent;
941         struct mem_cgroup *parent;
942         int ret;
943
944         /* Is ROOT ? */
945         if (!pcg)
946                 return -EINVAL;
947
948
949         parent = mem_cgroup_from_cont(pcg);
950
951
952         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
953         if (ret || !parent)
954                 return ret;
955
956         if (!get_page_unless_zero(page))
957                 return -EBUSY;
958
959         ret = isolate_lru_page(page);
960
961         if (ret)
962                 goto cancel;
963
964         ret = mem_cgroup_move_account(pc, child, parent);
965
966         /* drop extra refcnt by try_charge() (move_account increment one) */
967         css_put(&parent->css);
968         putback_lru_page(page);
969         if (!ret) {
970                 put_page(page);
971                 return 0;
972         }
973         /* uncharge if move fails */
974 cancel:
975         res_counter_uncharge(&parent->res, PAGE_SIZE);
976         if (do_swap_account)
977                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
978         put_page(page);
979         return ret;
980 }
981
982 /*
983  * Charge the memory controller for page usage.
984  * Return
985  * 0 if the charge was successful
986  * < 0 if the cgroup is over its limit
987  */
988 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
989                                 gfp_t gfp_mask, enum charge_type ctype,
990                                 struct mem_cgroup *memcg)
991 {
992         struct mem_cgroup *mem;
993         struct page_cgroup *pc;
994         int ret;
995
996         pc = lookup_page_cgroup(page);
997         /* can happen at boot */
998         if (unlikely(!pc))
999                 return 0;
1000         prefetchw(pc);
1001
1002         mem = memcg;
1003         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1004         if (ret || !mem)
1005                 return ret;
1006
1007         __mem_cgroup_commit_charge(mem, pc, ctype);
1008         return 0;
1009 }
1010
1011 int mem_cgroup_newpage_charge(struct page *page,
1012                               struct mm_struct *mm, gfp_t gfp_mask)
1013 {
1014         if (mem_cgroup_disabled())
1015                 return 0;
1016         if (PageCompound(page))
1017                 return 0;
1018         /*
1019          * If already mapped, we don't have to account.
1020          * If page cache, page->mapping has address_space.
1021          * But page->mapping may have out-of-use anon_vma pointer,
1022          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1023          * is NULL.
1024          */
1025         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1026                 return 0;
1027         if (unlikely(!mm))
1028                 mm = &init_mm;
1029         return mem_cgroup_charge_common(page, mm, gfp_mask,
1030                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1031 }
1032
1033 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1034                                 gfp_t gfp_mask)
1035 {
1036         if (mem_cgroup_disabled())
1037                 return 0;
1038         if (PageCompound(page))
1039                 return 0;
1040         /*
1041          * Corner case handling. This is called from add_to_page_cache()
1042          * in usual. But some FS (shmem) precharges this page before calling it
1043          * and call add_to_page_cache() with GFP_NOWAIT.
1044          *
1045          * For GFP_NOWAIT case, the page may be pre-charged before calling
1046          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1047          * charge twice. (It works but has to pay a bit larger cost.)
1048          */
1049         if (!(gfp_mask & __GFP_WAIT)) {
1050                 struct page_cgroup *pc;
1051
1052
1053                 pc = lookup_page_cgroup(page);
1054                 if (!pc)
1055                         return 0;
1056                 lock_page_cgroup(pc);
1057                 if (PageCgroupUsed(pc)) {
1058                         unlock_page_cgroup(pc);
1059                         return 0;
1060                 }
1061                 unlock_page_cgroup(pc);
1062         }
1063
1064         if (unlikely(!mm))
1065                 mm = &init_mm;
1066
1067         if (page_is_file_cache(page))
1068                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1069                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1070         else
1071                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1072                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, NULL);
1073 }
1074
1075 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1076                                  struct page *page,
1077                                  gfp_t mask, struct mem_cgroup **ptr)
1078 {
1079         struct mem_cgroup *mem;
1080         swp_entry_t     ent;
1081
1082         if (mem_cgroup_disabled())
1083                 return 0;
1084
1085         if (!do_swap_account)
1086                 goto charge_cur_mm;
1087
1088         /*
1089          * A racing thread's fault, or swapoff, may have already updated
1090          * the pte, and even removed page from swap cache: return success
1091          * to go on to do_swap_page()'s pte_same() test, which should fail.
1092          */
1093         if (!PageSwapCache(page))
1094                 return 0;
1095
1096         ent.val = page_private(page);
1097
1098         mem = lookup_swap_cgroup(ent);
1099         if (!mem || mem->obsolete)
1100                 goto charge_cur_mm;
1101         *ptr = mem;
1102         return __mem_cgroup_try_charge(NULL, mask, ptr, true);
1103 charge_cur_mm:
1104         if (unlikely(!mm))
1105                 mm = &init_mm;
1106         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1107 }
1108
1109 #ifdef CONFIG_SWAP
1110
1111 int mem_cgroup_cache_charge_swapin(struct page *page,
1112                         struct mm_struct *mm, gfp_t mask, bool locked)
1113 {
1114         int ret = 0;
1115
1116         if (mem_cgroup_disabled())
1117                 return 0;
1118         if (unlikely(!mm))
1119                 mm = &init_mm;
1120         if (!locked)
1121                 lock_page(page);
1122         /*
1123          * If not locked, the page can be dropped from SwapCache until
1124          * we reach here.
1125          */
1126         if (PageSwapCache(page)) {
1127                 struct mem_cgroup *mem = NULL;
1128                 swp_entry_t ent;
1129
1130                 ent.val = page_private(page);
1131                 if (do_swap_account) {
1132                         mem = lookup_swap_cgroup(ent);
1133                         if (mem && mem->obsolete)
1134                                 mem = NULL;
1135                         if (mem)
1136                                 mm = NULL;
1137                 }
1138                 ret = mem_cgroup_charge_common(page, mm, mask,
1139                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1140
1141                 if (!ret && do_swap_account) {
1142                         /* avoid double counting */
1143                         mem = swap_cgroup_record(ent, NULL);
1144                         if (mem) {
1145                                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1146                                 mem_cgroup_put(mem);
1147                         }
1148                 }
1149         }
1150         if (!locked)
1151                 unlock_page(page);
1152         /* add this page(page_cgroup) to the LRU we want. */
1153         mem_cgroup_lru_fixup(page);
1154
1155         return ret;
1156 }
1157 #endif
1158
1159 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1160 {
1161         struct page_cgroup *pc;
1162
1163         if (mem_cgroup_disabled())
1164                 return;
1165         if (!ptr)
1166                 return;
1167         pc = lookup_page_cgroup(page);
1168         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1169         /*
1170          * Now swap is on-memory. This means this page may be
1171          * counted both as mem and swap....double count.
1172          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1173          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1174          * may call delete_from_swap_cache() before reach here.
1175          */
1176         if (do_swap_account && PageSwapCache(page)) {
1177                 swp_entry_t ent = {.val = page_private(page)};
1178                 struct mem_cgroup *memcg;
1179                 memcg = swap_cgroup_record(ent, NULL);
1180                 if (memcg) {
1181                         /* If memcg is obsolete, memcg can be != ptr */
1182                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1183                         mem_cgroup_put(memcg);
1184                 }
1185
1186         }
1187         /* add this page(page_cgroup) to the LRU we want. */
1188         mem_cgroup_lru_fixup(page);
1189 }
1190
1191 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1192 {
1193         if (mem_cgroup_disabled())
1194                 return;
1195         if (!mem)
1196                 return;
1197         res_counter_uncharge(&mem->res, PAGE_SIZE);
1198         if (do_swap_account)
1199                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1200         css_put(&mem->css);
1201 }
1202
1203
1204 /*
1205  * uncharge if !page_mapped(page)
1206  */
1207 static struct mem_cgroup *
1208 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1209 {
1210         struct page_cgroup *pc;
1211         struct mem_cgroup *mem = NULL;
1212         struct mem_cgroup_per_zone *mz;
1213
1214         if (mem_cgroup_disabled())
1215                 return NULL;
1216
1217         if (PageSwapCache(page))
1218                 return NULL;
1219
1220         /*
1221          * Check if our page_cgroup is valid
1222          */
1223         pc = lookup_page_cgroup(page);
1224         if (unlikely(!pc || !PageCgroupUsed(pc)))
1225                 return NULL;
1226
1227         lock_page_cgroup(pc);
1228
1229         mem = pc->mem_cgroup;
1230
1231         if (!PageCgroupUsed(pc))
1232                 goto unlock_out;
1233
1234         switch (ctype) {
1235         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1236                 if (page_mapped(page))
1237                         goto unlock_out;
1238                 break;
1239         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1240                 if (!PageAnon(page)) {  /* Shared memory */
1241                         if (page->mapping && !page_is_file_cache(page))
1242                                 goto unlock_out;
1243                 } else if (page_mapped(page)) /* Anon */
1244                                 goto unlock_out;
1245                 break;
1246         default:
1247                 break;
1248         }
1249
1250         res_counter_uncharge(&mem->res, PAGE_SIZE);
1251         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1252                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1253
1254         mem_cgroup_charge_statistics(mem, pc, false);
1255         ClearPageCgroupUsed(pc);
1256
1257         mz = page_cgroup_zoneinfo(pc);
1258         unlock_page_cgroup(pc);
1259
1260         /* at swapout, this memcg will be accessed to record to swap */
1261         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1262                 css_put(&mem->css);
1263
1264         return mem;
1265
1266 unlock_out:
1267         unlock_page_cgroup(pc);
1268         return NULL;
1269 }
1270
1271 void mem_cgroup_uncharge_page(struct page *page)
1272 {
1273         /* early check. */
1274         if (page_mapped(page))
1275                 return;
1276         if (page->mapping && !PageAnon(page))
1277                 return;
1278         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1279 }
1280
1281 void mem_cgroup_uncharge_cache_page(struct page *page)
1282 {
1283         VM_BUG_ON(page_mapped(page));
1284         VM_BUG_ON(page->mapping);
1285         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1286 }
1287
1288 /*
1289  * called from __delete_from_swap_cache() and drop "page" account.
1290  * memcg information is recorded to swap_cgroup of "ent"
1291  */
1292 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1293 {
1294         struct mem_cgroup *memcg;
1295
1296         memcg = __mem_cgroup_uncharge_common(page,
1297                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1298         /* record memcg information */
1299         if (do_swap_account && memcg) {
1300                 swap_cgroup_record(ent, memcg);
1301                 mem_cgroup_get(memcg);
1302         }
1303         if (memcg)
1304                 css_put(&memcg->css);
1305 }
1306
1307 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1308 /*
1309  * called from swap_entry_free(). remove record in swap_cgroup and
1310  * uncharge "memsw" account.
1311  */
1312 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1313 {
1314         struct mem_cgroup *memcg;
1315
1316         if (!do_swap_account)
1317                 return;
1318
1319         memcg = swap_cgroup_record(ent, NULL);
1320         if (memcg) {
1321                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1322                 mem_cgroup_put(memcg);
1323         }
1324 }
1325 #endif
1326
1327 /*
1328  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1329  * page belongs to.
1330  */
1331 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1332 {
1333         struct page_cgroup *pc;
1334         struct mem_cgroup *mem = NULL;
1335         int ret = 0;
1336
1337         if (mem_cgroup_disabled())
1338                 return 0;
1339
1340         pc = lookup_page_cgroup(page);
1341         lock_page_cgroup(pc);
1342         if (PageCgroupUsed(pc)) {
1343                 mem = pc->mem_cgroup;
1344                 css_get(&mem->css);
1345         }
1346         unlock_page_cgroup(pc);
1347
1348         if (mem) {
1349                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1350                 css_put(&mem->css);
1351         }
1352         *ptr = mem;
1353         return ret;
1354 }
1355
1356 /* remove redundant charge if migration failed*/
1357 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1358                 struct page *oldpage, struct page *newpage)
1359 {
1360         struct page *target, *unused;
1361         struct page_cgroup *pc;
1362         enum charge_type ctype;
1363
1364         if (!mem)
1365                 return;
1366
1367         /* at migration success, oldpage->mapping is NULL. */
1368         if (oldpage->mapping) {
1369                 target = oldpage;
1370                 unused = NULL;
1371         } else {
1372                 target = newpage;
1373                 unused = oldpage;
1374         }
1375
1376         if (PageAnon(target))
1377                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1378         else if (page_is_file_cache(target))
1379                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1380         else
1381                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1382
1383         /* unused page is not on radix-tree now. */
1384         if (unused)
1385                 __mem_cgroup_uncharge_common(unused, ctype);
1386
1387         pc = lookup_page_cgroup(target);
1388         /*
1389          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1390          * So, double-counting is effectively avoided.
1391          */
1392         __mem_cgroup_commit_charge(mem, pc, ctype);
1393
1394         /*
1395          * Both of oldpage and newpage are still under lock_page().
1396          * Then, we don't have to care about race in radix-tree.
1397          * But we have to be careful that this page is unmapped or not.
1398          *
1399          * There is a case for !page_mapped(). At the start of
1400          * migration, oldpage was mapped. But now, it's zapped.
1401          * But we know *target* page is not freed/reused under us.
1402          * mem_cgroup_uncharge_page() does all necessary checks.
1403          */
1404         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1405                 mem_cgroup_uncharge_page(target);
1406 }
1407
1408 /*
1409  * A call to try to shrink memory usage under specified resource controller.
1410  * This is typically used for page reclaiming for shmem for reducing side
1411  * effect of page allocation from shmem, which is used by some mem_cgroup.
1412  */
1413 int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
1414 {
1415         struct mem_cgroup *mem;
1416         int progress = 0;
1417         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1418
1419         if (mem_cgroup_disabled())
1420                 return 0;
1421         if (!mm)
1422                 return 0;
1423
1424         rcu_read_lock();
1425         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1426         if (unlikely(!mem)) {
1427                 rcu_read_unlock();
1428                 return 0;
1429         }
1430         css_get(&mem->css);
1431         rcu_read_unlock();
1432
1433         do {
1434                 progress = mem_cgroup_hierarchical_reclaim(mem, gfp_mask, true);
1435                 progress += mem_cgroup_check_under_limit(mem);
1436         } while (!progress && --retry);
1437
1438         css_put(&mem->css);
1439         if (!retry)
1440                 return -ENOMEM;
1441         return 0;
1442 }
1443
1444 static DEFINE_MUTEX(set_limit_mutex);
1445
1446 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1447                                 unsigned long long val)
1448 {
1449
1450         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1451         int progress;
1452         u64 memswlimit;
1453         int ret = 0;
1454
1455         while (retry_count) {
1456                 if (signal_pending(current)) {
1457                         ret = -EINTR;
1458                         break;
1459                 }
1460                 /*
1461                  * Rather than hide all in some function, I do this in
1462                  * open coded manner. You see what this really does.
1463                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1464                  */
1465                 mutex_lock(&set_limit_mutex);
1466                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1467                 if (memswlimit < val) {
1468                         ret = -EINVAL;
1469                         mutex_unlock(&set_limit_mutex);
1470                         break;
1471                 }
1472                 ret = res_counter_set_limit(&memcg->res, val);
1473                 mutex_unlock(&set_limit_mutex);
1474
1475                 if (!ret)
1476                         break;
1477
1478                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1479                                                            false);
1480                 if (!progress)                  retry_count--;
1481         }
1482
1483         return ret;
1484 }
1485
1486 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1487                                 unsigned long long val)
1488 {
1489         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1490         u64 memlimit, oldusage, curusage;
1491         int ret;
1492
1493         if (!do_swap_account)
1494                 return -EINVAL;
1495
1496         while (retry_count) {
1497                 if (signal_pending(current)) {
1498                         ret = -EINTR;
1499                         break;
1500                 }
1501                 /*
1502                  * Rather than hide all in some function, I do this in
1503                  * open coded manner. You see what this really does.
1504                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1505                  */
1506                 mutex_lock(&set_limit_mutex);
1507                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1508                 if (memlimit > val) {
1509                         ret = -EINVAL;
1510                         mutex_unlock(&set_limit_mutex);
1511                         break;
1512                 }
1513                 ret = res_counter_set_limit(&memcg->memsw, val);
1514                 mutex_unlock(&set_limit_mutex);
1515
1516                 if (!ret)
1517                         break;
1518
1519                 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1520                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true);
1521                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1522                 if (curusage >= oldusage)
1523                         retry_count--;
1524         }
1525         return ret;
1526 }
1527
1528 /*
1529  * This routine traverse page_cgroup in given list and drop them all.
1530  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1531  */
1532 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1533                                 int node, int zid, enum lru_list lru)
1534 {
1535         struct zone *zone;
1536         struct mem_cgroup_per_zone *mz;
1537         struct page_cgroup *pc, *busy;
1538         unsigned long flags, loop;
1539         struct list_head *list;
1540         int ret = 0;
1541
1542         zone = &NODE_DATA(node)->node_zones[zid];
1543         mz = mem_cgroup_zoneinfo(mem, node, zid);
1544         list = &mz->lists[lru];
1545
1546         loop = MEM_CGROUP_ZSTAT(mz, lru);
1547         /* give some margin against EBUSY etc...*/
1548         loop += 256;
1549         busy = NULL;
1550         while (loop--) {
1551                 ret = 0;
1552                 spin_lock_irqsave(&zone->lru_lock, flags);
1553                 if (list_empty(list)) {
1554                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1555                         break;
1556                 }
1557                 pc = list_entry(list->prev, struct page_cgroup, lru);
1558                 if (busy == pc) {
1559                         list_move(&pc->lru, list);
1560                         busy = 0;
1561                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1562                         continue;
1563                 }
1564                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1565
1566                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1567                 if (ret == -ENOMEM)
1568                         break;
1569
1570                 if (ret == -EBUSY || ret == -EINVAL) {
1571                         /* found lock contention or "pc" is obsolete. */
1572                         busy = pc;
1573                         cond_resched();
1574                 } else
1575                         busy = NULL;
1576         }
1577
1578         if (!ret && !list_empty(list))
1579                 return -EBUSY;
1580         return ret;
1581 }
1582
1583 /*
1584  * make mem_cgroup's charge to be 0 if there is no task.
1585  * This enables deleting this mem_cgroup.
1586  */
1587 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1588 {
1589         int ret;
1590         int node, zid, shrink;
1591         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1592         struct cgroup *cgrp = mem->css.cgroup;
1593
1594         css_get(&mem->css);
1595
1596         shrink = 0;
1597         /* should free all ? */
1598         if (free_all)
1599                 goto try_to_free;
1600 move_account:
1601         while (mem->res.usage > 0) {
1602                 ret = -EBUSY;
1603                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1604                         goto out;
1605                 ret = -EINTR;
1606                 if (signal_pending(current))
1607                         goto out;
1608                 /* This is for making all *used* pages to be on LRU. */
1609                 lru_add_drain_all();
1610                 ret = 0;
1611                 for_each_node_state(node, N_POSSIBLE) {
1612                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1613                                 enum lru_list l;
1614                                 for_each_lru(l) {
1615                                         ret = mem_cgroup_force_empty_list(mem,
1616                                                         node, zid, l);
1617                                         if (ret)
1618                                                 break;
1619                                 }
1620                         }
1621                         if (ret)
1622                                 break;
1623                 }
1624                 /* it seems parent cgroup doesn't have enough mem */
1625                 if (ret == -ENOMEM)
1626                         goto try_to_free;
1627                 cond_resched();
1628         }
1629         ret = 0;
1630 out:
1631         css_put(&mem->css);
1632         return ret;
1633
1634 try_to_free:
1635         /* returns EBUSY if there is a task or if we come here twice. */
1636         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1637                 ret = -EBUSY;
1638                 goto out;
1639         }
1640         /* we call try-to-free pages for make this cgroup empty */
1641         lru_add_drain_all();
1642         /* try to free all pages in this cgroup */
1643         shrink = 1;
1644         while (nr_retries && mem->res.usage > 0) {
1645                 int progress;
1646
1647                 if (signal_pending(current)) {
1648                         ret = -EINTR;
1649                         goto out;
1650                 }
1651                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1652                                                 false, get_swappiness(mem));
1653                 if (!progress) {
1654                         nr_retries--;
1655                         /* maybe some writeback is necessary */
1656                         congestion_wait(WRITE, HZ/10);
1657                 }
1658
1659         }
1660         lru_add_drain();
1661         /* try move_account...there may be some *locked* pages. */
1662         if (mem->res.usage)
1663                 goto move_account;
1664         ret = 0;
1665         goto out;
1666 }
1667
1668 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1669 {
1670         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1671 }
1672
1673
1674 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1675 {
1676         return mem_cgroup_from_cont(cont)->use_hierarchy;
1677 }
1678
1679 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1680                                         u64 val)
1681 {
1682         int retval = 0;
1683         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1684         struct cgroup *parent = cont->parent;
1685         struct mem_cgroup *parent_mem = NULL;
1686
1687         if (parent)
1688                 parent_mem = mem_cgroup_from_cont(parent);
1689
1690         cgroup_lock();
1691         /*
1692          * If parent's use_hiearchy is set, we can't make any modifications
1693          * in the child subtrees. If it is unset, then the change can
1694          * occur, provided the current cgroup has no children.
1695          *
1696          * For the root cgroup, parent_mem is NULL, we allow value to be
1697          * set if there are no children.
1698          */
1699         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1700                                 (val == 1 || val == 0)) {
1701                 if (list_empty(&cont->children))
1702                         mem->use_hierarchy = val;
1703                 else
1704                         retval = -EBUSY;
1705         } else
1706                 retval = -EINVAL;
1707         cgroup_unlock();
1708
1709         return retval;
1710 }
1711
1712 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1713 {
1714         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1715         u64 val = 0;
1716         int type, name;
1717
1718         type = MEMFILE_TYPE(cft->private);
1719         name = MEMFILE_ATTR(cft->private);
1720         switch (type) {
1721         case _MEM:
1722                 val = res_counter_read_u64(&mem->res, name);
1723                 break;
1724         case _MEMSWAP:
1725                 if (do_swap_account)
1726                         val = res_counter_read_u64(&mem->memsw, name);
1727                 break;
1728         default:
1729                 BUG();
1730                 break;
1731         }
1732         return val;
1733 }
1734 /*
1735  * The user of this function is...
1736  * RES_LIMIT.
1737  */
1738 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1739                             const char *buffer)
1740 {
1741         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1742         int type, name;
1743         unsigned long long val;
1744         int ret;
1745
1746         type = MEMFILE_TYPE(cft->private);
1747         name = MEMFILE_ATTR(cft->private);
1748         switch (name) {
1749         case RES_LIMIT:
1750                 /* This function does all necessary parse...reuse it */
1751                 ret = res_counter_memparse_write_strategy(buffer, &val);
1752                 if (ret)
1753                         break;
1754                 if (type == _MEM)
1755                         ret = mem_cgroup_resize_limit(memcg, val);
1756                 else
1757                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1758                 break;
1759         default:
1760                 ret = -EINVAL; /* should be BUG() ? */
1761                 break;
1762         }
1763         return ret;
1764 }
1765
1766 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1767                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1768 {
1769         struct cgroup *cgroup;
1770         unsigned long long min_limit, min_memsw_limit, tmp;
1771
1772         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1773         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1774         cgroup = memcg->css.cgroup;
1775         if (!memcg->use_hierarchy)
1776                 goto out;
1777
1778         while (cgroup->parent) {
1779                 cgroup = cgroup->parent;
1780                 memcg = mem_cgroup_from_cont(cgroup);
1781                 if (!memcg->use_hierarchy)
1782                         break;
1783                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
1784                 min_limit = min(min_limit, tmp);
1785                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1786                 min_memsw_limit = min(min_memsw_limit, tmp);
1787         }
1788 out:
1789         *mem_limit = min_limit;
1790         *memsw_limit = min_memsw_limit;
1791         return;
1792 }
1793
1794 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1795 {
1796         struct mem_cgroup *mem;
1797         int type, name;
1798
1799         mem = mem_cgroup_from_cont(cont);
1800         type = MEMFILE_TYPE(event);
1801         name = MEMFILE_ATTR(event);
1802         switch (name) {
1803         case RES_MAX_USAGE:
1804                 if (type == _MEM)
1805                         res_counter_reset_max(&mem->res);
1806                 else
1807                         res_counter_reset_max(&mem->memsw);
1808                 break;
1809         case RES_FAILCNT:
1810                 if (type == _MEM)
1811                         res_counter_reset_failcnt(&mem->res);
1812                 else
1813                         res_counter_reset_failcnt(&mem->memsw);
1814                 break;
1815         }
1816         return 0;
1817 }
1818
1819 static const struct mem_cgroup_stat_desc {
1820         const char *msg;
1821         u64 unit;
1822 } mem_cgroup_stat_desc[] = {
1823         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
1824         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
1825         [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
1826         [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
1827 };
1828
1829 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
1830                                  struct cgroup_map_cb *cb)
1831 {
1832         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
1833         struct mem_cgroup_stat *stat = &mem_cont->stat;
1834         int i;
1835
1836         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
1837                 s64 val;
1838
1839                 val = mem_cgroup_read_stat(stat, i);
1840                 val *= mem_cgroup_stat_desc[i].unit;
1841                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
1842         }
1843         /* showing # of active pages */
1844         {
1845                 unsigned long active_anon, inactive_anon;
1846                 unsigned long active_file, inactive_file;
1847                 unsigned long unevictable;
1848
1849                 inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
1850                                                 LRU_INACTIVE_ANON);
1851                 active_anon = mem_cgroup_get_all_zonestat(mem_cont,
1852                                                 LRU_ACTIVE_ANON);
1853                 inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
1854                                                 LRU_INACTIVE_FILE);
1855                 active_file = mem_cgroup_get_all_zonestat(mem_cont,
1856                                                 LRU_ACTIVE_FILE);
1857                 unevictable = mem_cgroup_get_all_zonestat(mem_cont,
1858                                                         LRU_UNEVICTABLE);
1859
1860                 cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
1861                 cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
1862                 cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
1863                 cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
1864                 cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
1865
1866         }
1867         {
1868                 unsigned long long limit, memsw_limit;
1869                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
1870                 cb->fill(cb, "hierarchical_memory_limit", limit);
1871                 if (do_swap_account)
1872                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
1873         }
1874
1875 #ifdef CONFIG_DEBUG_VM
1876         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
1877
1878         {
1879                 int nid, zid;
1880                 struct mem_cgroup_per_zone *mz;
1881                 unsigned long recent_rotated[2] = {0, 0};
1882                 unsigned long recent_scanned[2] = {0, 0};
1883
1884                 for_each_online_node(nid)
1885                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1886                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1887
1888                                 recent_rotated[0] +=
1889                                         mz->reclaim_stat.recent_rotated[0];
1890                                 recent_rotated[1] +=
1891                                         mz->reclaim_stat.recent_rotated[1];
1892                                 recent_scanned[0] +=
1893                                         mz->reclaim_stat.recent_scanned[0];
1894                                 recent_scanned[1] +=
1895                                         mz->reclaim_stat.recent_scanned[1];
1896                         }
1897                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
1898                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
1899                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
1900                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
1901         }
1902 #endif
1903
1904         return 0;
1905 }
1906
1907 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
1908 {
1909         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1910
1911         return get_swappiness(memcg);
1912 }
1913
1914 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
1915                                        u64 val)
1916 {
1917         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1918         struct mem_cgroup *parent;
1919         if (val > 100)
1920                 return -EINVAL;
1921
1922         if (cgrp->parent == NULL)
1923                 return -EINVAL;
1924
1925         parent = mem_cgroup_from_cont(cgrp->parent);
1926         /* If under hierarchy, only empty-root can set this value */
1927         if ((parent->use_hierarchy) ||
1928             (memcg->use_hierarchy && !list_empty(&cgrp->children)))
1929                 return -EINVAL;
1930
1931         spin_lock(&memcg->reclaim_param_lock);
1932         memcg->swappiness = val;
1933         spin_unlock(&memcg->reclaim_param_lock);
1934
1935         return 0;
1936 }
1937
1938
1939 static struct cftype mem_cgroup_files[] = {
1940         {
1941                 .name = "usage_in_bytes",
1942                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
1943                 .read_u64 = mem_cgroup_read,
1944         },
1945         {
1946                 .name = "max_usage_in_bytes",
1947                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
1948                 .trigger = mem_cgroup_reset,
1949                 .read_u64 = mem_cgroup_read,
1950         },
1951         {
1952                 .name = "limit_in_bytes",
1953                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
1954                 .write_string = mem_cgroup_write,
1955                 .read_u64 = mem_cgroup_read,
1956         },
1957         {
1958                 .name = "failcnt",
1959                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
1960                 .trigger = mem_cgroup_reset,
1961                 .read_u64 = mem_cgroup_read,
1962         },
1963         {
1964                 .name = "stat",
1965                 .read_map = mem_control_stat_show,
1966         },
1967         {
1968                 .name = "force_empty",
1969                 .trigger = mem_cgroup_force_empty_write,
1970         },
1971         {
1972                 .name = "use_hierarchy",
1973                 .write_u64 = mem_cgroup_hierarchy_write,
1974                 .read_u64 = mem_cgroup_hierarchy_read,
1975         },
1976         {
1977                 .name = "swappiness",
1978                 .read_u64 = mem_cgroup_swappiness_read,
1979                 .write_u64 = mem_cgroup_swappiness_write,
1980         },
1981 };
1982
1983 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1984 static struct cftype memsw_cgroup_files[] = {
1985         {
1986                 .name = "memsw.usage_in_bytes",
1987                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
1988                 .read_u64 = mem_cgroup_read,
1989         },
1990         {
1991                 .name = "memsw.max_usage_in_bytes",
1992                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
1993                 .trigger = mem_cgroup_reset,
1994                 .read_u64 = mem_cgroup_read,
1995         },
1996         {
1997                 .name = "memsw.limit_in_bytes",
1998                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
1999                 .write_string = mem_cgroup_write,
2000                 .read_u64 = mem_cgroup_read,
2001         },
2002         {
2003                 .name = "memsw.failcnt",
2004                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2005                 .trigger = mem_cgroup_reset,
2006                 .read_u64 = mem_cgroup_read,
2007         },
2008 };
2009
2010 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2011 {
2012         if (!do_swap_account)
2013                 return 0;
2014         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2015                                 ARRAY_SIZE(memsw_cgroup_files));
2016 };
2017 #else
2018 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2019 {
2020         return 0;
2021 }
2022 #endif
2023
2024 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2025 {
2026         struct mem_cgroup_per_node *pn;
2027         struct mem_cgroup_per_zone *mz;
2028         enum lru_list l;
2029         int zone, tmp = node;
2030         /*
2031          * This routine is called against possible nodes.
2032          * But it's BUG to call kmalloc() against offline node.
2033          *
2034          * TODO: this routine can waste much memory for nodes which will
2035          *       never be onlined. It's better to use memory hotplug callback
2036          *       function.
2037          */
2038         if (!node_state(node, N_NORMAL_MEMORY))
2039                 tmp = -1;
2040         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2041         if (!pn)
2042                 return 1;
2043
2044         mem->info.nodeinfo[node] = pn;
2045         memset(pn, 0, sizeof(*pn));
2046
2047         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2048                 mz = &pn->zoneinfo[zone];
2049                 for_each_lru(l)
2050                         INIT_LIST_HEAD(&mz->lists[l]);
2051         }
2052         return 0;
2053 }
2054
2055 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2056 {
2057         kfree(mem->info.nodeinfo[node]);
2058 }
2059
2060 static int mem_cgroup_size(void)
2061 {
2062         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2063         return sizeof(struct mem_cgroup) + cpustat_size;
2064 }
2065
2066 static struct mem_cgroup *mem_cgroup_alloc(void)
2067 {
2068         struct mem_cgroup *mem;
2069         int size = mem_cgroup_size();
2070
2071         if (size < PAGE_SIZE)
2072                 mem = kmalloc(size, GFP_KERNEL);
2073         else
2074                 mem = vmalloc(size);
2075
2076         if (mem)
2077                 memset(mem, 0, size);
2078         return mem;
2079 }
2080
2081 /*
2082  * At destroying mem_cgroup, references from swap_cgroup can remain.
2083  * (scanning all at force_empty is too costly...)
2084  *
2085  * Instead of clearing all references at force_empty, we remember
2086  * the number of reference from swap_cgroup and free mem_cgroup when
2087  * it goes down to 0.
2088  *
2089  * When mem_cgroup is destroyed, mem->obsolete will be set to 0 and
2090  * entry which points to this memcg will be ignore at swapin.
2091  *
2092  * Removal of cgroup itself succeeds regardless of refs from swap.
2093  */
2094
2095 static void mem_cgroup_free(struct mem_cgroup *mem)
2096 {
2097         int node;
2098
2099         if (atomic_read(&mem->refcnt) > 0)
2100                 return;
2101
2102
2103         for_each_node_state(node, N_POSSIBLE)
2104                 free_mem_cgroup_per_zone_info(mem, node);
2105
2106         if (mem_cgroup_size() < PAGE_SIZE)
2107                 kfree(mem);
2108         else
2109                 vfree(mem);
2110 }
2111
2112 static void mem_cgroup_get(struct mem_cgroup *mem)
2113 {
2114         atomic_inc(&mem->refcnt);
2115 }
2116
2117 static void mem_cgroup_put(struct mem_cgroup *mem)
2118 {
2119         if (atomic_dec_and_test(&mem->refcnt)) {
2120                 if (!mem->obsolete)
2121                         return;
2122                 mem_cgroup_free(mem);
2123         }
2124 }
2125
2126
2127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2128 static void __init enable_swap_cgroup(void)
2129 {
2130         if (!mem_cgroup_disabled() && really_do_swap_account)
2131                 do_swap_account = 1;
2132 }
2133 #else
2134 static void __init enable_swap_cgroup(void)
2135 {
2136 }
2137 #endif
2138
2139 static struct cgroup_subsys_state *
2140 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2141 {
2142         struct mem_cgroup *mem, *parent;
2143         int node;
2144
2145         mem = mem_cgroup_alloc();
2146         if (!mem)
2147                 return ERR_PTR(-ENOMEM);
2148
2149         for_each_node_state(node, N_POSSIBLE)
2150                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2151                         goto free_out;
2152         /* root ? */
2153         if (cont->parent == NULL) {
2154                 enable_swap_cgroup();
2155                 parent = NULL;
2156         } else {
2157                 parent = mem_cgroup_from_cont(cont->parent);
2158                 mem->use_hierarchy = parent->use_hierarchy;
2159         }
2160
2161         if (parent && parent->use_hierarchy) {
2162                 res_counter_init(&mem->res, &parent->res);
2163                 res_counter_init(&mem->memsw, &parent->memsw);
2164         } else {
2165                 res_counter_init(&mem->res, NULL);
2166                 res_counter_init(&mem->memsw, NULL);
2167         }
2168         mem->last_scanned_child = NULL;
2169         spin_lock_init(&mem->reclaim_param_lock);
2170
2171         if (parent)
2172                 mem->swappiness = get_swappiness(parent);
2173
2174         return &mem->css;
2175 free_out:
2176         for_each_node_state(node, N_POSSIBLE)
2177                 free_mem_cgroup_per_zone_info(mem, node);
2178         mem_cgroup_free(mem);
2179         return ERR_PTR(-ENOMEM);
2180 }
2181
2182 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2183                                         struct cgroup *cont)
2184 {
2185         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2186         mem->obsolete = 1;
2187         mem_cgroup_force_empty(mem, false);
2188 }
2189
2190 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2191                                 struct cgroup *cont)
2192 {
2193         mem_cgroup_free(mem_cgroup_from_cont(cont));
2194 }
2195
2196 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2197                                 struct cgroup *cont)
2198 {
2199         int ret;
2200
2201         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2202                                 ARRAY_SIZE(mem_cgroup_files));
2203
2204         if (!ret)
2205                 ret = register_memsw_files(cont, ss);
2206         return ret;
2207 }
2208
2209 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2210                                 struct cgroup *cont,
2211                                 struct cgroup *old_cont,
2212                                 struct task_struct *p)
2213 {
2214         mutex_lock(&memcg_tasklist);
2215         /*
2216          * FIXME: It's better to move charges of this process from old
2217          * memcg to new memcg. But it's just on TODO-List now.
2218          */
2219         mutex_unlock(&memcg_tasklist);
2220 }
2221
2222 struct cgroup_subsys mem_cgroup_subsys = {
2223         .name = "memory",
2224         .subsys_id = mem_cgroup_subsys_id,
2225         .create = mem_cgroup_create,
2226         .pre_destroy = mem_cgroup_pre_destroy,
2227         .destroy = mem_cgroup_destroy,
2228         .populate = mem_cgroup_populate,
2229         .attach = mem_cgroup_move_task,
2230         .early_init = 0,
2231 };
2232
2233 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2234
2235 static int __init disable_swap_account(char *s)
2236 {
2237         really_do_swap_account = 0;
2238         return 1;
2239 }
2240 __setup("noswapaccount", disable_swap_account);
2241 #endif