memcg: fix section mismatch
[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 hierarchy_mutex
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         atomic_t        refcnt;
166
167         unsigned int    swappiness;
168
169         /*
170          * statistics. This must be placed at the end of memcg.
171          */
172         struct mem_cgroup_stat stat;
173 };
174
175 enum charge_type {
176         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
177         MEM_CGROUP_CHARGE_TYPE_MAPPED,
178         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
179         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
180         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
181         NR_CHARGE_TYPE,
182 };
183
184 /* only for here (for easy reading.) */
185 #define PCGF_CACHE      (1UL << PCG_CACHE)
186 #define PCGF_USED       (1UL << PCG_USED)
187 #define PCGF_LOCK       (1UL << PCG_LOCK)
188 static const unsigned long
189 pcg_default_flags[NR_CHARGE_TYPE] = {
190         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
191         PCGF_USED | PCGF_LOCK, /* Anon */
192         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
193         0, /* FORCE */
194 };
195
196 /* for encoding cft->private value on file */
197 #define _MEM                    (0)
198 #define _MEMSWAP                (1)
199 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
200 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
201 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
202
203 static void mem_cgroup_get(struct mem_cgroup *mem);
204 static void mem_cgroup_put(struct mem_cgroup *mem);
205
206 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
207                                          struct page_cgroup *pc,
208                                          bool charge)
209 {
210         int val = (charge)? 1 : -1;
211         struct mem_cgroup_stat *stat = &mem->stat;
212         struct mem_cgroup_stat_cpu *cpustat;
213         int cpu = get_cpu();
214
215         cpustat = &stat->cpustat[cpu];
216         if (PageCgroupCache(pc))
217                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
218         else
219                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
220
221         if (charge)
222                 __mem_cgroup_stat_add_safe(cpustat,
223                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
224         else
225                 __mem_cgroup_stat_add_safe(cpustat,
226                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
227         put_cpu();
228 }
229
230 static struct mem_cgroup_per_zone *
231 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
232 {
233         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
234 }
235
236 static struct mem_cgroup_per_zone *
237 page_cgroup_zoneinfo(struct page_cgroup *pc)
238 {
239         struct mem_cgroup *mem = pc->mem_cgroup;
240         int nid = page_cgroup_nid(pc);
241         int zid = page_cgroup_zid(pc);
242
243         if (!mem)
244                 return NULL;
245
246         return mem_cgroup_zoneinfo(mem, nid, zid);
247 }
248
249 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
250                                         enum lru_list idx)
251 {
252         int nid, zid;
253         struct mem_cgroup_per_zone *mz;
254         u64 total = 0;
255
256         for_each_online_node(nid)
257                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
258                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
259                         total += MEM_CGROUP_ZSTAT(mz, idx);
260                 }
261         return total;
262 }
263
264 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
265 {
266         return container_of(cgroup_subsys_state(cont,
267                                 mem_cgroup_subsys_id), struct mem_cgroup,
268                                 css);
269 }
270
271 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
272 {
273         /*
274          * mm_update_next_owner() may clear mm->owner to NULL
275          * if it races with swapoff, page migration, etc.
276          * So this can be called with p == NULL.
277          */
278         if (unlikely(!p))
279                 return NULL;
280
281         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
282                                 struct mem_cgroup, css);
283 }
284
285 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
286 {
287         struct mem_cgroup *mem = NULL;
288         /*
289          * Because we have no locks, mm->owner's may be being moved to other
290          * cgroup. We use css_tryget() here even if this looks
291          * pessimistic (rather than adding locks here).
292          */
293         rcu_read_lock();
294         do {
295                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
296                 if (unlikely(!mem))
297                         break;
298         } while (!css_tryget(&mem->css));
299         rcu_read_unlock();
300         return mem;
301 }
302
303 static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem)
304 {
305         if (!mem)
306                 return true;
307         return css_is_removed(&mem->css);
308 }
309
310 /*
311  * Following LRU functions are allowed to be used without PCG_LOCK.
312  * Operations are called by routine of global LRU independently from memcg.
313  * What we have to take care of here is validness of pc->mem_cgroup.
314  *
315  * Changes to pc->mem_cgroup happens when
316  * 1. charge
317  * 2. moving account
318  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
319  * It is added to LRU before charge.
320  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
321  * When moving account, the page is not on LRU. It's isolated.
322  */
323
324 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
325 {
326         struct page_cgroup *pc;
327         struct mem_cgroup *mem;
328         struct mem_cgroup_per_zone *mz;
329
330         if (mem_cgroup_disabled())
331                 return;
332         pc = lookup_page_cgroup(page);
333         /* can happen while we handle swapcache. */
334         if (list_empty(&pc->lru) || !pc->mem_cgroup)
335                 return;
336         /*
337          * We don't check PCG_USED bit. It's cleared when the "page" is finally
338          * removed from global LRU.
339          */
340         mz = page_cgroup_zoneinfo(pc);
341         mem = pc->mem_cgroup;
342         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
343         list_del_init(&pc->lru);
344         return;
345 }
346
347 void mem_cgroup_del_lru(struct page *page)
348 {
349         mem_cgroup_del_lru_list(page, page_lru(page));
350 }
351
352 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
353 {
354         struct mem_cgroup_per_zone *mz;
355         struct page_cgroup *pc;
356
357         if (mem_cgroup_disabled())
358                 return;
359
360         pc = lookup_page_cgroup(page);
361         /*
362          * Used bit is set without atomic ops but after smp_wmb().
363          * For making pc->mem_cgroup visible, insert smp_rmb() here.
364          */
365         smp_rmb();
366         /* unused page is not rotated. */
367         if (!PageCgroupUsed(pc))
368                 return;
369         mz = page_cgroup_zoneinfo(pc);
370         list_move(&pc->lru, &mz->lists[lru]);
371 }
372
373 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
374 {
375         struct page_cgroup *pc;
376         struct mem_cgroup_per_zone *mz;
377
378         if (mem_cgroup_disabled())
379                 return;
380         pc = lookup_page_cgroup(page);
381         /*
382          * Used bit is set without atomic ops but after smp_wmb().
383          * For making pc->mem_cgroup visible, insert smp_rmb() here.
384          */
385         smp_rmb();
386         if (!PageCgroupUsed(pc))
387                 return;
388
389         mz = page_cgroup_zoneinfo(pc);
390         MEM_CGROUP_ZSTAT(mz, lru) += 1;
391         list_add(&pc->lru, &mz->lists[lru]);
392 }
393
394 /*
395  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
396  * lru because the page may.be reused after it's fully uncharged (because of
397  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
398  * it again. This function is only used to charge SwapCache. It's done under
399  * lock_page and expected that zone->lru_lock is never held.
400  */
401 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
402 {
403         unsigned long flags;
404         struct zone *zone = page_zone(page);
405         struct page_cgroup *pc = lookup_page_cgroup(page);
406
407         spin_lock_irqsave(&zone->lru_lock, flags);
408         /*
409          * Forget old LRU when this page_cgroup is *not* used. This Used bit
410          * is guarded by lock_page() because the page is SwapCache.
411          */
412         if (!PageCgroupUsed(pc))
413                 mem_cgroup_del_lru_list(page, page_lru(page));
414         spin_unlock_irqrestore(&zone->lru_lock, flags);
415 }
416
417 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
418 {
419         unsigned long flags;
420         struct zone *zone = page_zone(page);
421         struct page_cgroup *pc = lookup_page_cgroup(page);
422
423         spin_lock_irqsave(&zone->lru_lock, flags);
424         /* link when the page is linked to LRU but page_cgroup isn't */
425         if (PageLRU(page) && list_empty(&pc->lru))
426                 mem_cgroup_add_lru_list(page, page_lru(page));
427         spin_unlock_irqrestore(&zone->lru_lock, flags);
428 }
429
430
431 void mem_cgroup_move_lists(struct page *page,
432                            enum lru_list from, enum lru_list to)
433 {
434         if (mem_cgroup_disabled())
435                 return;
436         mem_cgroup_del_lru_list(page, from);
437         mem_cgroup_add_lru_list(page, to);
438 }
439
440 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
441 {
442         int ret;
443
444         task_lock(task);
445         ret = task->mm && mm_match_cgroup(task->mm, mem);
446         task_unlock(task);
447         return ret;
448 }
449
450 /*
451  * Calculate mapped_ratio under memory controller. This will be used in
452  * vmscan.c for deteremining we have to reclaim mapped pages.
453  */
454 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
455 {
456         long total, rss;
457
458         /*
459          * usage is recorded in bytes. But, here, we assume the number of
460          * physical pages can be represented by "long" on any arch.
461          */
462         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
463         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
464         return (int)((rss * 100L) / total);
465 }
466
467 /*
468  * prev_priority control...this will be used in memory reclaim path.
469  */
470 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
471 {
472         int prev_priority;
473
474         spin_lock(&mem->reclaim_param_lock);
475         prev_priority = mem->prev_priority;
476         spin_unlock(&mem->reclaim_param_lock);
477
478         return prev_priority;
479 }
480
481 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
482 {
483         spin_lock(&mem->reclaim_param_lock);
484         if (priority < mem->prev_priority)
485                 mem->prev_priority = priority;
486         spin_unlock(&mem->reclaim_param_lock);
487 }
488
489 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
490 {
491         spin_lock(&mem->reclaim_param_lock);
492         mem->prev_priority = priority;
493         spin_unlock(&mem->reclaim_param_lock);
494 }
495
496 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
497 {
498         unsigned long active;
499         unsigned long inactive;
500         unsigned long gb;
501         unsigned long inactive_ratio;
502
503         inactive = mem_cgroup_get_all_zonestat(memcg, LRU_INACTIVE_ANON);
504         active = mem_cgroup_get_all_zonestat(memcg, LRU_ACTIVE_ANON);
505
506         gb = (inactive + active) >> (30 - PAGE_SHIFT);
507         if (gb)
508                 inactive_ratio = int_sqrt(10 * gb);
509         else
510                 inactive_ratio = 1;
511
512         if (present_pages) {
513                 present_pages[0] = inactive;
514                 present_pages[1] = active;
515         }
516
517         return inactive_ratio;
518 }
519
520 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
521 {
522         unsigned long active;
523         unsigned long inactive;
524         unsigned long present_pages[2];
525         unsigned long inactive_ratio;
526
527         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
528
529         inactive = present_pages[0];
530         active = present_pages[1];
531
532         if (inactive * inactive_ratio < active)
533                 return 1;
534
535         return 0;
536 }
537
538 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
539                                        struct zone *zone,
540                                        enum lru_list lru)
541 {
542         int nid = zone->zone_pgdat->node_id;
543         int zid = zone_idx(zone);
544         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
545
546         return MEM_CGROUP_ZSTAT(mz, lru);
547 }
548
549 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
550                                                       struct zone *zone)
551 {
552         int nid = zone->zone_pgdat->node_id;
553         int zid = zone_idx(zone);
554         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
555
556         return &mz->reclaim_stat;
557 }
558
559 struct zone_reclaim_stat *
560 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
561 {
562         struct page_cgroup *pc;
563         struct mem_cgroup_per_zone *mz;
564
565         if (mem_cgroup_disabled())
566                 return NULL;
567
568         pc = lookup_page_cgroup(page);
569         /*
570          * Used bit is set without atomic ops but after smp_wmb().
571          * For making pc->mem_cgroup visible, insert smp_rmb() here.
572          */
573         smp_rmb();
574         if (!PageCgroupUsed(pc))
575                 return NULL;
576
577         mz = page_cgroup_zoneinfo(pc);
578         if (!mz)
579                 return NULL;
580
581         return &mz->reclaim_stat;
582 }
583
584 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
585                                         struct list_head *dst,
586                                         unsigned long *scanned, int order,
587                                         int mode, struct zone *z,
588                                         struct mem_cgroup *mem_cont,
589                                         int active, int file)
590 {
591         unsigned long nr_taken = 0;
592         struct page *page;
593         unsigned long scan;
594         LIST_HEAD(pc_list);
595         struct list_head *src;
596         struct page_cgroup *pc, *tmp;
597         int nid = z->zone_pgdat->node_id;
598         int zid = zone_idx(z);
599         struct mem_cgroup_per_zone *mz;
600         int lru = LRU_FILE * !!file + !!active;
601
602         BUG_ON(!mem_cont);
603         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
604         src = &mz->lists[lru];
605
606         scan = 0;
607         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
608                 if (scan >= nr_to_scan)
609                         break;
610
611                 page = pc->page;
612                 if (unlikely(!PageCgroupUsed(pc)))
613                         continue;
614                 if (unlikely(!PageLRU(page)))
615                         continue;
616
617                 scan++;
618                 if (__isolate_lru_page(page, mode, file) == 0) {
619                         list_move(&page->lru, dst);
620                         nr_taken++;
621                 }
622         }
623
624         *scanned = scan;
625         return nr_taken;
626 }
627
628 #define mem_cgroup_from_res_counter(counter, member)    \
629         container_of(counter, struct mem_cgroup, member)
630
631 /*
632  * This routine finds the DFS walk successor. This routine should be
633  * called with hierarchy_mutex held
634  */
635 static struct mem_cgroup *
636 __mem_cgroup_get_next_node(struct mem_cgroup *curr, struct mem_cgroup *root_mem)
637 {
638         struct cgroup *cgroup, *curr_cgroup, *root_cgroup;
639
640         curr_cgroup = curr->css.cgroup;
641         root_cgroup = root_mem->css.cgroup;
642
643         if (!list_empty(&curr_cgroup->children)) {
644                 /*
645                  * Walk down to children
646                  */
647                 cgroup = list_entry(curr_cgroup->children.next,
648                                                 struct cgroup, sibling);
649                 curr = mem_cgroup_from_cont(cgroup);
650                 goto done;
651         }
652
653 visit_parent:
654         if (curr_cgroup == root_cgroup) {
655                 /* caller handles NULL case */
656                 curr = NULL;
657                 goto done;
658         }
659
660         /*
661          * Goto next sibling
662          */
663         if (curr_cgroup->sibling.next != &curr_cgroup->parent->children) {
664                 cgroup = list_entry(curr_cgroup->sibling.next, struct cgroup,
665                                                 sibling);
666                 curr = mem_cgroup_from_cont(cgroup);
667                 goto done;
668         }
669
670         /*
671          * Go up to next parent and next parent's sibling if need be
672          */
673         curr_cgroup = curr_cgroup->parent;
674         goto visit_parent;
675
676 done:
677         return curr;
678 }
679
680 /*
681  * Visit the first child (need not be the first child as per the ordering
682  * of the cgroup list, since we track last_scanned_child) of @mem and use
683  * that to reclaim free pages from.
684  */
685 static struct mem_cgroup *
686 mem_cgroup_get_next_node(struct mem_cgroup *root_mem)
687 {
688         struct cgroup *cgroup;
689         struct mem_cgroup *orig, *next;
690         bool obsolete;
691
692         /*
693          * Scan all children under the mem_cgroup mem
694          */
695         mutex_lock(&mem_cgroup_subsys.hierarchy_mutex);
696
697         orig = root_mem->last_scanned_child;
698         obsolete = mem_cgroup_is_obsolete(orig);
699
700         if (list_empty(&root_mem->css.cgroup->children)) {
701                 /*
702                  * root_mem might have children before and last_scanned_child
703                  * may point to one of them. We put it later.
704                  */
705                 if (orig)
706                         VM_BUG_ON(!obsolete);
707                 next = NULL;
708                 goto done;
709         }
710
711         if (!orig || obsolete) {
712                 cgroup = list_first_entry(&root_mem->css.cgroup->children,
713                                 struct cgroup, sibling);
714                 next = mem_cgroup_from_cont(cgroup);
715         } else
716                 next = __mem_cgroup_get_next_node(orig, root_mem);
717
718 done:
719         if (next)
720                 mem_cgroup_get(next);
721         root_mem->last_scanned_child = next;
722         if (orig)
723                 mem_cgroup_put(orig);
724         mutex_unlock(&mem_cgroup_subsys.hierarchy_mutex);
725         return (next) ? next : root_mem;
726 }
727
728 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
729 {
730         if (do_swap_account) {
731                 if (res_counter_check_under_limit(&mem->res) &&
732                         res_counter_check_under_limit(&mem->memsw))
733                         return true;
734         } else
735                 if (res_counter_check_under_limit(&mem->res))
736                         return true;
737         return false;
738 }
739
740 static unsigned int get_swappiness(struct mem_cgroup *memcg)
741 {
742         struct cgroup *cgrp = memcg->css.cgroup;
743         unsigned int swappiness;
744
745         /* root ? */
746         if (cgrp->parent == NULL)
747                 return vm_swappiness;
748
749         spin_lock(&memcg->reclaim_param_lock);
750         swappiness = memcg->swappiness;
751         spin_unlock(&memcg->reclaim_param_lock);
752
753         return swappiness;
754 }
755
756 /*
757  * Dance down the hierarchy if needed to reclaim memory. We remember the
758  * last child we reclaimed from, so that we don't end up penalizing
759  * one child extensively based on its position in the children list.
760  *
761  * root_mem is the original ancestor that we've been reclaim from.
762  */
763 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
764                                                 gfp_t gfp_mask, bool noswap)
765 {
766         struct mem_cgroup *next_mem;
767         int ret = 0;
768
769         /*
770          * Reclaim unconditionally and don't check for return value.
771          * We need to reclaim in the current group and down the tree.
772          * One might think about checking for children before reclaiming,
773          * but there might be left over accounting, even after children
774          * have left.
775          */
776         ret += try_to_free_mem_cgroup_pages(root_mem, gfp_mask, noswap,
777                                            get_swappiness(root_mem));
778         if (mem_cgroup_check_under_limit(root_mem))
779                 return 1;       /* indicate reclaim has succeeded */
780         if (!root_mem->use_hierarchy)
781                 return ret;
782
783         next_mem = mem_cgroup_get_next_node(root_mem);
784
785         while (next_mem != root_mem) {
786                 if (mem_cgroup_is_obsolete(next_mem)) {
787                         next_mem = mem_cgroup_get_next_node(root_mem);
788                         continue;
789                 }
790                 ret += try_to_free_mem_cgroup_pages(next_mem, gfp_mask, noswap,
791                                                    get_swappiness(next_mem));
792                 if (mem_cgroup_check_under_limit(root_mem))
793                         return 1;       /* indicate reclaim has succeeded */
794                 next_mem = mem_cgroup_get_next_node(root_mem);
795         }
796         return ret;
797 }
798
799 bool mem_cgroup_oom_called(struct task_struct *task)
800 {
801         bool ret = false;
802         struct mem_cgroup *mem;
803         struct mm_struct *mm;
804
805         rcu_read_lock();
806         mm = task->mm;
807         if (!mm)
808                 mm = &init_mm;
809         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
810         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
811                 ret = true;
812         rcu_read_unlock();
813         return ret;
814 }
815 /*
816  * Unlike exported interface, "oom" parameter is added. if oom==true,
817  * oom-killer can be invoked.
818  */
819 static int __mem_cgroup_try_charge(struct mm_struct *mm,
820                         gfp_t gfp_mask, struct mem_cgroup **memcg,
821                         bool oom)
822 {
823         struct mem_cgroup *mem, *mem_over_limit;
824         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
825         struct res_counter *fail_res;
826
827         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
828                 /* Don't account this! */
829                 *memcg = NULL;
830                 return 0;
831         }
832
833         /*
834          * We always charge the cgroup the mm_struct belongs to.
835          * The mm_struct's mem_cgroup changes on task migration if the
836          * thread group leader migrates. It's possible that mm is not
837          * set, if so charge the init_mm (happens for pagecache usage).
838          */
839         mem = *memcg;
840         if (likely(!mem)) {
841                 mem = try_get_mem_cgroup_from_mm(mm);
842                 *memcg = mem;
843         } else {
844                 css_get(&mem->css);
845         }
846         if (unlikely(!mem))
847                 return 0;
848
849         VM_BUG_ON(mem_cgroup_is_obsolete(mem));
850
851         while (1) {
852                 int ret;
853                 bool noswap = false;
854
855                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
856                 if (likely(!ret)) {
857                         if (!do_swap_account)
858                                 break;
859                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
860                                                         &fail_res);
861                         if (likely(!ret))
862                                 break;
863                         /* mem+swap counter fails */
864                         res_counter_uncharge(&mem->res, PAGE_SIZE);
865                         noswap = true;
866                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
867                                                                         memsw);
868                 } else
869                         /* mem counter fails */
870                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
871                                                                         res);
872
873                 if (!(gfp_mask & __GFP_WAIT))
874                         goto nomem;
875
876                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
877                                                         noswap);
878                 if (ret)
879                         continue;
880
881                 /*
882                  * try_to_free_mem_cgroup_pages() might not give us a full
883                  * picture of reclaim. Some pages are reclaimed and might be
884                  * moved to swap cache or just unmapped from the cgroup.
885                  * Check the limit again to see if the reclaim reduced the
886                  * current usage of the cgroup before giving up
887                  *
888                  */
889                 if (mem_cgroup_check_under_limit(mem_over_limit))
890                         continue;
891
892                 if (!nr_retries--) {
893                         if (oom) {
894                                 mutex_lock(&memcg_tasklist);
895                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
896                                 mutex_unlock(&memcg_tasklist);
897                                 mem_over_limit->last_oom_jiffies = jiffies;
898                         }
899                         goto nomem;
900                 }
901         }
902         return 0;
903 nomem:
904         css_put(&mem->css);
905         return -ENOMEM;
906 }
907
908 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
909 {
910         struct mem_cgroup *mem;
911         swp_entry_t ent;
912
913         if (!PageSwapCache(page))
914                 return NULL;
915
916         ent.val = page_private(page);
917         mem = lookup_swap_cgroup(ent);
918         if (!mem)
919                 return NULL;
920         if (!css_tryget(&mem->css))
921                 return NULL;
922         return mem;
923 }
924
925 /*
926  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
927  * USED state. If already USED, uncharge and return.
928  */
929
930 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
931                                      struct page_cgroup *pc,
932                                      enum charge_type ctype)
933 {
934         /* try_charge() can return NULL to *memcg, taking care of it. */
935         if (!mem)
936                 return;
937
938         lock_page_cgroup(pc);
939         if (unlikely(PageCgroupUsed(pc))) {
940                 unlock_page_cgroup(pc);
941                 res_counter_uncharge(&mem->res, PAGE_SIZE);
942                 if (do_swap_account)
943                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
944                 css_put(&mem->css);
945                 return;
946         }
947         pc->mem_cgroup = mem;
948         smp_wmb();
949         pc->flags = pcg_default_flags[ctype];
950
951         mem_cgroup_charge_statistics(mem, pc, true);
952
953         unlock_page_cgroup(pc);
954 }
955
956 /**
957  * mem_cgroup_move_account - move account of the page
958  * @pc: page_cgroup of the page.
959  * @from: mem_cgroup which the page is moved from.
960  * @to: mem_cgroup which the page is moved to. @from != @to.
961  *
962  * The caller must confirm following.
963  * - page is not on LRU (isolate_page() is useful.)
964  *
965  * returns 0 at success,
966  * returns -EBUSY when lock is busy or "pc" is unstable.
967  *
968  * This function does "uncharge" from old cgroup but doesn't do "charge" to
969  * new cgroup. It should be done by a caller.
970  */
971
972 static int mem_cgroup_move_account(struct page_cgroup *pc,
973         struct mem_cgroup *from, struct mem_cgroup *to)
974 {
975         struct mem_cgroup_per_zone *from_mz, *to_mz;
976         int nid, zid;
977         int ret = -EBUSY;
978
979         VM_BUG_ON(from == to);
980         VM_BUG_ON(PageLRU(pc->page));
981
982         nid = page_cgroup_nid(pc);
983         zid = page_cgroup_zid(pc);
984         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
985         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
986
987         if (!trylock_page_cgroup(pc))
988                 return ret;
989
990         if (!PageCgroupUsed(pc))
991                 goto out;
992
993         if (pc->mem_cgroup != from)
994                 goto out;
995
996         res_counter_uncharge(&from->res, PAGE_SIZE);
997         mem_cgroup_charge_statistics(from, pc, false);
998         if (do_swap_account)
999                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1000         css_put(&from->css);
1001
1002         css_get(&to->css);
1003         pc->mem_cgroup = to;
1004         mem_cgroup_charge_statistics(to, pc, true);
1005         ret = 0;
1006 out:
1007         unlock_page_cgroup(pc);
1008         return ret;
1009 }
1010
1011 /*
1012  * move charges to its parent.
1013  */
1014
1015 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1016                                   struct mem_cgroup *child,
1017                                   gfp_t gfp_mask)
1018 {
1019         struct page *page = pc->page;
1020         struct cgroup *cg = child->css.cgroup;
1021         struct cgroup *pcg = cg->parent;
1022         struct mem_cgroup *parent;
1023         int ret;
1024
1025         /* Is ROOT ? */
1026         if (!pcg)
1027                 return -EINVAL;
1028
1029
1030         parent = mem_cgroup_from_cont(pcg);
1031
1032
1033         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1034         if (ret || !parent)
1035                 return ret;
1036
1037         if (!get_page_unless_zero(page)) {
1038                 ret = -EBUSY;
1039                 goto uncharge;
1040         }
1041
1042         ret = isolate_lru_page(page);
1043
1044         if (ret)
1045                 goto cancel;
1046
1047         ret = mem_cgroup_move_account(pc, child, parent);
1048
1049         putback_lru_page(page);
1050         if (!ret) {
1051                 put_page(page);
1052                 /* drop extra refcnt by try_charge() */
1053                 css_put(&parent->css);
1054                 return 0;
1055         }
1056
1057 cancel:
1058         put_page(page);
1059 uncharge:
1060         /* drop extra refcnt by try_charge() */
1061         css_put(&parent->css);
1062         /* uncharge if move fails */
1063         res_counter_uncharge(&parent->res, PAGE_SIZE);
1064         if (do_swap_account)
1065                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1066         return ret;
1067 }
1068
1069 /*
1070  * Charge the memory controller for page usage.
1071  * Return
1072  * 0 if the charge was successful
1073  * < 0 if the cgroup is over its limit
1074  */
1075 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1076                                 gfp_t gfp_mask, enum charge_type ctype,
1077                                 struct mem_cgroup *memcg)
1078 {
1079         struct mem_cgroup *mem;
1080         struct page_cgroup *pc;
1081         int ret;
1082
1083         pc = lookup_page_cgroup(page);
1084         /* can happen at boot */
1085         if (unlikely(!pc))
1086                 return 0;
1087         prefetchw(pc);
1088
1089         mem = memcg;
1090         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1091         if (ret || !mem)
1092                 return ret;
1093
1094         __mem_cgroup_commit_charge(mem, pc, ctype);
1095         return 0;
1096 }
1097
1098 int mem_cgroup_newpage_charge(struct page *page,
1099                               struct mm_struct *mm, gfp_t gfp_mask)
1100 {
1101         if (mem_cgroup_disabled())
1102                 return 0;
1103         if (PageCompound(page))
1104                 return 0;
1105         /*
1106          * If already mapped, we don't have to account.
1107          * If page cache, page->mapping has address_space.
1108          * But page->mapping may have out-of-use anon_vma pointer,
1109          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1110          * is NULL.
1111          */
1112         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1113                 return 0;
1114         if (unlikely(!mm))
1115                 mm = &init_mm;
1116         return mem_cgroup_charge_common(page, mm, gfp_mask,
1117                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1118 }
1119
1120 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1121                                 gfp_t gfp_mask)
1122 {
1123         struct mem_cgroup *mem = NULL;
1124         int ret;
1125
1126         if (mem_cgroup_disabled())
1127                 return 0;
1128         if (PageCompound(page))
1129                 return 0;
1130         /*
1131          * Corner case handling. This is called from add_to_page_cache()
1132          * in usual. But some FS (shmem) precharges this page before calling it
1133          * and call add_to_page_cache() with GFP_NOWAIT.
1134          *
1135          * For GFP_NOWAIT case, the page may be pre-charged before calling
1136          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1137          * charge twice. (It works but has to pay a bit larger cost.)
1138          * And when the page is SwapCache, it should take swap information
1139          * into account. This is under lock_page() now.
1140          */
1141         if (!(gfp_mask & __GFP_WAIT)) {
1142                 struct page_cgroup *pc;
1143
1144
1145                 pc = lookup_page_cgroup(page);
1146                 if (!pc)
1147                         return 0;
1148                 lock_page_cgroup(pc);
1149                 if (PageCgroupUsed(pc)) {
1150                         unlock_page_cgroup(pc);
1151                         return 0;
1152                 }
1153                 unlock_page_cgroup(pc);
1154         }
1155
1156         if (do_swap_account && PageSwapCache(page)) {
1157                 mem = try_get_mem_cgroup_from_swapcache(page);
1158                 if (mem)
1159                         mm = NULL;
1160                   else
1161                         mem = NULL;
1162                 /* SwapCache may be still linked to LRU now. */
1163                 mem_cgroup_lru_del_before_commit_swapcache(page);
1164         }
1165
1166         if (unlikely(!mm && !mem))
1167                 mm = &init_mm;
1168
1169         if (page_is_file_cache(page))
1170                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1171                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1172
1173         ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1174                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1175         if (mem)
1176                 css_put(&mem->css);
1177         if (PageSwapCache(page))
1178                 mem_cgroup_lru_add_after_commit_swapcache(page);
1179
1180         if (do_swap_account && !ret && PageSwapCache(page)) {
1181                 swp_entry_t ent = {.val = page_private(page)};
1182                 /* avoid double counting */
1183                 mem = swap_cgroup_record(ent, NULL);
1184                 if (mem) {
1185                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1186                         mem_cgroup_put(mem);
1187                 }
1188         }
1189         return ret;
1190 }
1191
1192 /*
1193  * While swap-in, try_charge -> commit or cancel, the page is locked.
1194  * And when try_charge() successfully returns, one refcnt to memcg without
1195  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1196  * "commit()" or removed by "cancel()"
1197  */
1198 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1199                                  struct page *page,
1200                                  gfp_t mask, struct mem_cgroup **ptr)
1201 {
1202         struct mem_cgroup *mem;
1203         int ret;
1204
1205         if (mem_cgroup_disabled())
1206                 return 0;
1207
1208         if (!do_swap_account)
1209                 goto charge_cur_mm;
1210         /*
1211          * A racing thread's fault, or swapoff, may have already updated
1212          * the pte, and even removed page from swap cache: return success
1213          * to go on to do_swap_page()'s pte_same() test, which should fail.
1214          */
1215         if (!PageSwapCache(page))
1216                 return 0;
1217         mem = try_get_mem_cgroup_from_swapcache(page);
1218         if (!mem)
1219                 goto charge_cur_mm;
1220         *ptr = mem;
1221         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1222         /* drop extra refcnt from tryget */
1223         css_put(&mem->css);
1224         return ret;
1225 charge_cur_mm:
1226         if (unlikely(!mm))
1227                 mm = &init_mm;
1228         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1229 }
1230
1231 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1232 {
1233         struct page_cgroup *pc;
1234
1235         if (mem_cgroup_disabled())
1236                 return;
1237         if (!ptr)
1238                 return;
1239         pc = lookup_page_cgroup(page);
1240         mem_cgroup_lru_del_before_commit_swapcache(page);
1241         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1242         mem_cgroup_lru_add_after_commit_swapcache(page);
1243         /*
1244          * Now swap is on-memory. This means this page may be
1245          * counted both as mem and swap....double count.
1246          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1247          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1248          * may call delete_from_swap_cache() before reach here.
1249          */
1250         if (do_swap_account && PageSwapCache(page)) {
1251                 swp_entry_t ent = {.val = page_private(page)};
1252                 struct mem_cgroup *memcg;
1253                 memcg = swap_cgroup_record(ent, NULL);
1254                 if (memcg) {
1255                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1256                         mem_cgroup_put(memcg);
1257                 }
1258
1259         }
1260         /* add this page(page_cgroup) to the LRU we want. */
1261
1262 }
1263
1264 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1265 {
1266         if (mem_cgroup_disabled())
1267                 return;
1268         if (!mem)
1269                 return;
1270         res_counter_uncharge(&mem->res, PAGE_SIZE);
1271         if (do_swap_account)
1272                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1273         css_put(&mem->css);
1274 }
1275
1276
1277 /*
1278  * uncharge if !page_mapped(page)
1279  */
1280 static struct mem_cgroup *
1281 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1282 {
1283         struct page_cgroup *pc;
1284         struct mem_cgroup *mem = NULL;
1285         struct mem_cgroup_per_zone *mz;
1286
1287         if (mem_cgroup_disabled())
1288                 return NULL;
1289
1290         if (PageSwapCache(page))
1291                 return NULL;
1292
1293         /*
1294          * Check if our page_cgroup is valid
1295          */
1296         pc = lookup_page_cgroup(page);
1297         if (unlikely(!pc || !PageCgroupUsed(pc)))
1298                 return NULL;
1299
1300         lock_page_cgroup(pc);
1301
1302         mem = pc->mem_cgroup;
1303
1304         if (!PageCgroupUsed(pc))
1305                 goto unlock_out;
1306
1307         switch (ctype) {
1308         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1309                 if (page_mapped(page))
1310                         goto unlock_out;
1311                 break;
1312         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1313                 if (!PageAnon(page)) {  /* Shared memory */
1314                         if (page->mapping && !page_is_file_cache(page))
1315                                 goto unlock_out;
1316                 } else if (page_mapped(page)) /* Anon */
1317                                 goto unlock_out;
1318                 break;
1319         default:
1320                 break;
1321         }
1322
1323         res_counter_uncharge(&mem->res, PAGE_SIZE);
1324         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1325                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1326
1327         mem_cgroup_charge_statistics(mem, pc, false);
1328         ClearPageCgroupUsed(pc);
1329         /*
1330          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1331          * freed from LRU. This is safe because uncharged page is expected not
1332          * to be reused (freed soon). Exception is SwapCache, it's handled by
1333          * special functions.
1334          */
1335
1336         mz = page_cgroup_zoneinfo(pc);
1337         unlock_page_cgroup(pc);
1338
1339         /* at swapout, this memcg will be accessed to record to swap */
1340         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1341                 css_put(&mem->css);
1342
1343         return mem;
1344
1345 unlock_out:
1346         unlock_page_cgroup(pc);
1347         return NULL;
1348 }
1349
1350 void mem_cgroup_uncharge_page(struct page *page)
1351 {
1352         /* early check. */
1353         if (page_mapped(page))
1354                 return;
1355         if (page->mapping && !PageAnon(page))
1356                 return;
1357         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1358 }
1359
1360 void mem_cgroup_uncharge_cache_page(struct page *page)
1361 {
1362         VM_BUG_ON(page_mapped(page));
1363         VM_BUG_ON(page->mapping);
1364         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1365 }
1366
1367 /*
1368  * called from __delete_from_swap_cache() and drop "page" account.
1369  * memcg information is recorded to swap_cgroup of "ent"
1370  */
1371 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1372 {
1373         struct mem_cgroup *memcg;
1374
1375         memcg = __mem_cgroup_uncharge_common(page,
1376                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1377         /* record memcg information */
1378         if (do_swap_account && memcg) {
1379                 swap_cgroup_record(ent, memcg);
1380                 mem_cgroup_get(memcg);
1381         }
1382         if (memcg)
1383                 css_put(&memcg->css);
1384 }
1385
1386 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1387 /*
1388  * called from swap_entry_free(). remove record in swap_cgroup and
1389  * uncharge "memsw" account.
1390  */
1391 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1392 {
1393         struct mem_cgroup *memcg;
1394
1395         if (!do_swap_account)
1396                 return;
1397
1398         memcg = swap_cgroup_record(ent, NULL);
1399         if (memcg) {
1400                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1401                 mem_cgroup_put(memcg);
1402         }
1403 }
1404 #endif
1405
1406 /*
1407  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1408  * page belongs to.
1409  */
1410 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1411 {
1412         struct page_cgroup *pc;
1413         struct mem_cgroup *mem = NULL;
1414         int ret = 0;
1415
1416         if (mem_cgroup_disabled())
1417                 return 0;
1418
1419         pc = lookup_page_cgroup(page);
1420         lock_page_cgroup(pc);
1421         if (PageCgroupUsed(pc)) {
1422                 mem = pc->mem_cgroup;
1423                 css_get(&mem->css);
1424         }
1425         unlock_page_cgroup(pc);
1426
1427         if (mem) {
1428                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1429                 css_put(&mem->css);
1430         }
1431         *ptr = mem;
1432         return ret;
1433 }
1434
1435 /* remove redundant charge if migration failed*/
1436 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1437                 struct page *oldpage, struct page *newpage)
1438 {
1439         struct page *target, *unused;
1440         struct page_cgroup *pc;
1441         enum charge_type ctype;
1442
1443         if (!mem)
1444                 return;
1445
1446         /* at migration success, oldpage->mapping is NULL. */
1447         if (oldpage->mapping) {
1448                 target = oldpage;
1449                 unused = NULL;
1450         } else {
1451                 target = newpage;
1452                 unused = oldpage;
1453         }
1454
1455         if (PageAnon(target))
1456                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1457         else if (page_is_file_cache(target))
1458                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1459         else
1460                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1461
1462         /* unused page is not on radix-tree now. */
1463         if (unused)
1464                 __mem_cgroup_uncharge_common(unused, ctype);
1465
1466         pc = lookup_page_cgroup(target);
1467         /*
1468          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1469          * So, double-counting is effectively avoided.
1470          */
1471         __mem_cgroup_commit_charge(mem, pc, ctype);
1472
1473         /*
1474          * Both of oldpage and newpage are still under lock_page().
1475          * Then, we don't have to care about race in radix-tree.
1476          * But we have to be careful that this page is unmapped or not.
1477          *
1478          * There is a case for !page_mapped(). At the start of
1479          * migration, oldpage was mapped. But now, it's zapped.
1480          * But we know *target* page is not freed/reused under us.
1481          * mem_cgroup_uncharge_page() does all necessary checks.
1482          */
1483         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1484                 mem_cgroup_uncharge_page(target);
1485 }
1486
1487 /*
1488  * A call to try to shrink memory usage under specified resource controller.
1489  * This is typically used for page reclaiming for shmem for reducing side
1490  * effect of page allocation from shmem, which is used by some mem_cgroup.
1491  */
1492 int mem_cgroup_shrink_usage(struct page *page,
1493                             struct mm_struct *mm,
1494                             gfp_t gfp_mask)
1495 {
1496         struct mem_cgroup *mem = NULL;
1497         int progress = 0;
1498         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1499
1500         if (mem_cgroup_disabled())
1501                 return 0;
1502         if (page)
1503                 mem = try_get_mem_cgroup_from_swapcache(page);
1504         if (!mem && mm)
1505                 mem = try_get_mem_cgroup_from_mm(mm);
1506         if (unlikely(!mem))
1507                 return 0;
1508
1509         do {
1510                 progress = mem_cgroup_hierarchical_reclaim(mem, gfp_mask, true);
1511                 progress += mem_cgroup_check_under_limit(mem);
1512         } while (!progress && --retry);
1513
1514         css_put(&mem->css);
1515         if (!retry)
1516                 return -ENOMEM;
1517         return 0;
1518 }
1519
1520 static DEFINE_MUTEX(set_limit_mutex);
1521
1522 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1523                                 unsigned long long val)
1524 {
1525
1526         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1527         int progress;
1528         u64 memswlimit;
1529         int ret = 0;
1530
1531         while (retry_count) {
1532                 if (signal_pending(current)) {
1533                         ret = -EINTR;
1534                         break;
1535                 }
1536                 /*
1537                  * Rather than hide all in some function, I do this in
1538                  * open coded manner. You see what this really does.
1539                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1540                  */
1541                 mutex_lock(&set_limit_mutex);
1542                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1543                 if (memswlimit < val) {
1544                         ret = -EINVAL;
1545                         mutex_unlock(&set_limit_mutex);
1546                         break;
1547                 }
1548                 ret = res_counter_set_limit(&memcg->res, val);
1549                 mutex_unlock(&set_limit_mutex);
1550
1551                 if (!ret)
1552                         break;
1553
1554                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1555                                                            false);
1556                 if (!progress)                  retry_count--;
1557         }
1558
1559         return ret;
1560 }
1561
1562 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1563                                 unsigned long long val)
1564 {
1565         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1566         u64 memlimit, oldusage, curusage;
1567         int ret;
1568
1569         if (!do_swap_account)
1570                 return -EINVAL;
1571
1572         while (retry_count) {
1573                 if (signal_pending(current)) {
1574                         ret = -EINTR;
1575                         break;
1576                 }
1577                 /*
1578                  * Rather than hide all in some function, I do this in
1579                  * open coded manner. You see what this really does.
1580                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1581                  */
1582                 mutex_lock(&set_limit_mutex);
1583                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1584                 if (memlimit > val) {
1585                         ret = -EINVAL;
1586                         mutex_unlock(&set_limit_mutex);
1587                         break;
1588                 }
1589                 ret = res_counter_set_limit(&memcg->memsw, val);
1590                 mutex_unlock(&set_limit_mutex);
1591
1592                 if (!ret)
1593                         break;
1594
1595                 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1596                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true);
1597                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1598                 if (curusage >= oldusage)
1599                         retry_count--;
1600         }
1601         return ret;
1602 }
1603
1604 /*
1605  * This routine traverse page_cgroup in given list and drop them all.
1606  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1607  */
1608 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1609                                 int node, int zid, enum lru_list lru)
1610 {
1611         struct zone *zone;
1612         struct mem_cgroup_per_zone *mz;
1613         struct page_cgroup *pc, *busy;
1614         unsigned long flags, loop;
1615         struct list_head *list;
1616         int ret = 0;
1617
1618         zone = &NODE_DATA(node)->node_zones[zid];
1619         mz = mem_cgroup_zoneinfo(mem, node, zid);
1620         list = &mz->lists[lru];
1621
1622         loop = MEM_CGROUP_ZSTAT(mz, lru);
1623         /* give some margin against EBUSY etc...*/
1624         loop += 256;
1625         busy = NULL;
1626         while (loop--) {
1627                 ret = 0;
1628                 spin_lock_irqsave(&zone->lru_lock, flags);
1629                 if (list_empty(list)) {
1630                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1631                         break;
1632                 }
1633                 pc = list_entry(list->prev, struct page_cgroup, lru);
1634                 if (busy == pc) {
1635                         list_move(&pc->lru, list);
1636                         busy = 0;
1637                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1638                         continue;
1639                 }
1640                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1641
1642                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1643                 if (ret == -ENOMEM)
1644                         break;
1645
1646                 if (ret == -EBUSY || ret == -EINVAL) {
1647                         /* found lock contention or "pc" is obsolete. */
1648                         busy = pc;
1649                         cond_resched();
1650                 } else
1651                         busy = NULL;
1652         }
1653
1654         if (!ret && !list_empty(list))
1655                 return -EBUSY;
1656         return ret;
1657 }
1658
1659 /*
1660  * make mem_cgroup's charge to be 0 if there is no task.
1661  * This enables deleting this mem_cgroup.
1662  */
1663 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1664 {
1665         int ret;
1666         int node, zid, shrink;
1667         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1668         struct cgroup *cgrp = mem->css.cgroup;
1669
1670         css_get(&mem->css);
1671
1672         shrink = 0;
1673         /* should free all ? */
1674         if (free_all)
1675                 goto try_to_free;
1676 move_account:
1677         while (mem->res.usage > 0) {
1678                 ret = -EBUSY;
1679                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1680                         goto out;
1681                 ret = -EINTR;
1682                 if (signal_pending(current))
1683                         goto out;
1684                 /* This is for making all *used* pages to be on LRU. */
1685                 lru_add_drain_all();
1686                 ret = 0;
1687                 for_each_node_state(node, N_POSSIBLE) {
1688                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1689                                 enum lru_list l;
1690                                 for_each_lru(l) {
1691                                         ret = mem_cgroup_force_empty_list(mem,
1692                                                         node, zid, l);
1693                                         if (ret)
1694                                                 break;
1695                                 }
1696                         }
1697                         if (ret)
1698                                 break;
1699                 }
1700                 /* it seems parent cgroup doesn't have enough mem */
1701                 if (ret == -ENOMEM)
1702                         goto try_to_free;
1703                 cond_resched();
1704         }
1705         ret = 0;
1706 out:
1707         css_put(&mem->css);
1708         return ret;
1709
1710 try_to_free:
1711         /* returns EBUSY if there is a task or if we come here twice. */
1712         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1713                 ret = -EBUSY;
1714                 goto out;
1715         }
1716         /* we call try-to-free pages for make this cgroup empty */
1717         lru_add_drain_all();
1718         /* try to free all pages in this cgroup */
1719         shrink = 1;
1720         while (nr_retries && mem->res.usage > 0) {
1721                 int progress;
1722
1723                 if (signal_pending(current)) {
1724                         ret = -EINTR;
1725                         goto out;
1726                 }
1727                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1728                                                 false, get_swappiness(mem));
1729                 if (!progress) {
1730                         nr_retries--;
1731                         /* maybe some writeback is necessary */
1732                         congestion_wait(WRITE, HZ/10);
1733                 }
1734
1735         }
1736         lru_add_drain();
1737         /* try move_account...there may be some *locked* pages. */
1738         if (mem->res.usage)
1739                 goto move_account;
1740         ret = 0;
1741         goto out;
1742 }
1743
1744 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1745 {
1746         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1747 }
1748
1749
1750 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1751 {
1752         return mem_cgroup_from_cont(cont)->use_hierarchy;
1753 }
1754
1755 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1756                                         u64 val)
1757 {
1758         int retval = 0;
1759         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1760         struct cgroup *parent = cont->parent;
1761         struct mem_cgroup *parent_mem = NULL;
1762
1763         if (parent)
1764                 parent_mem = mem_cgroup_from_cont(parent);
1765
1766         cgroup_lock();
1767         /*
1768          * If parent's use_hiearchy is set, we can't make any modifications
1769          * in the child subtrees. If it is unset, then the change can
1770          * occur, provided the current cgroup has no children.
1771          *
1772          * For the root cgroup, parent_mem is NULL, we allow value to be
1773          * set if there are no children.
1774          */
1775         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1776                                 (val == 1 || val == 0)) {
1777                 if (list_empty(&cont->children))
1778                         mem->use_hierarchy = val;
1779                 else
1780                         retval = -EBUSY;
1781         } else
1782                 retval = -EINVAL;
1783         cgroup_unlock();
1784
1785         return retval;
1786 }
1787
1788 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1789 {
1790         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1791         u64 val = 0;
1792         int type, name;
1793
1794         type = MEMFILE_TYPE(cft->private);
1795         name = MEMFILE_ATTR(cft->private);
1796         switch (type) {
1797         case _MEM:
1798                 val = res_counter_read_u64(&mem->res, name);
1799                 break;
1800         case _MEMSWAP:
1801                 if (do_swap_account)
1802                         val = res_counter_read_u64(&mem->memsw, name);
1803                 break;
1804         default:
1805                 BUG();
1806                 break;
1807         }
1808         return val;
1809 }
1810 /*
1811  * The user of this function is...
1812  * RES_LIMIT.
1813  */
1814 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1815                             const char *buffer)
1816 {
1817         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1818         int type, name;
1819         unsigned long long val;
1820         int ret;
1821
1822         type = MEMFILE_TYPE(cft->private);
1823         name = MEMFILE_ATTR(cft->private);
1824         switch (name) {
1825         case RES_LIMIT:
1826                 /* This function does all necessary parse...reuse it */
1827                 ret = res_counter_memparse_write_strategy(buffer, &val);
1828                 if (ret)
1829                         break;
1830                 if (type == _MEM)
1831                         ret = mem_cgroup_resize_limit(memcg, val);
1832                 else
1833                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1834                 break;
1835         default:
1836                 ret = -EINVAL; /* should be BUG() ? */
1837                 break;
1838         }
1839         return ret;
1840 }
1841
1842 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1843                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1844 {
1845         struct cgroup *cgroup;
1846         unsigned long long min_limit, min_memsw_limit, tmp;
1847
1848         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1849         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1850         cgroup = memcg->css.cgroup;
1851         if (!memcg->use_hierarchy)
1852                 goto out;
1853
1854         while (cgroup->parent) {
1855                 cgroup = cgroup->parent;
1856                 memcg = mem_cgroup_from_cont(cgroup);
1857                 if (!memcg->use_hierarchy)
1858                         break;
1859                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
1860                 min_limit = min(min_limit, tmp);
1861                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1862                 min_memsw_limit = min(min_memsw_limit, tmp);
1863         }
1864 out:
1865         *mem_limit = min_limit;
1866         *memsw_limit = min_memsw_limit;
1867         return;
1868 }
1869
1870 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1871 {
1872         struct mem_cgroup *mem;
1873         int type, name;
1874
1875         mem = mem_cgroup_from_cont(cont);
1876         type = MEMFILE_TYPE(event);
1877         name = MEMFILE_ATTR(event);
1878         switch (name) {
1879         case RES_MAX_USAGE:
1880                 if (type == _MEM)
1881                         res_counter_reset_max(&mem->res);
1882                 else
1883                         res_counter_reset_max(&mem->memsw);
1884                 break;
1885         case RES_FAILCNT:
1886                 if (type == _MEM)
1887                         res_counter_reset_failcnt(&mem->res);
1888                 else
1889                         res_counter_reset_failcnt(&mem->memsw);
1890                 break;
1891         }
1892         return 0;
1893 }
1894
1895 static const struct mem_cgroup_stat_desc {
1896         const char *msg;
1897         u64 unit;
1898 } mem_cgroup_stat_desc[] = {
1899         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
1900         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
1901         [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
1902         [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
1903 };
1904
1905 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
1906                                  struct cgroup_map_cb *cb)
1907 {
1908         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
1909         struct mem_cgroup_stat *stat = &mem_cont->stat;
1910         int i;
1911
1912         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
1913                 s64 val;
1914
1915                 val = mem_cgroup_read_stat(stat, i);
1916                 val *= mem_cgroup_stat_desc[i].unit;
1917                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
1918         }
1919         /* showing # of active pages */
1920         {
1921                 unsigned long active_anon, inactive_anon;
1922                 unsigned long active_file, inactive_file;
1923                 unsigned long unevictable;
1924
1925                 inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
1926                                                 LRU_INACTIVE_ANON);
1927                 active_anon = mem_cgroup_get_all_zonestat(mem_cont,
1928                                                 LRU_ACTIVE_ANON);
1929                 inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
1930                                                 LRU_INACTIVE_FILE);
1931                 active_file = mem_cgroup_get_all_zonestat(mem_cont,
1932                                                 LRU_ACTIVE_FILE);
1933                 unevictable = mem_cgroup_get_all_zonestat(mem_cont,
1934                                                         LRU_UNEVICTABLE);
1935
1936                 cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
1937                 cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
1938                 cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
1939                 cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
1940                 cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
1941
1942         }
1943         {
1944                 unsigned long long limit, memsw_limit;
1945                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
1946                 cb->fill(cb, "hierarchical_memory_limit", limit);
1947                 if (do_swap_account)
1948                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
1949         }
1950
1951 #ifdef CONFIG_DEBUG_VM
1952         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
1953
1954         {
1955                 int nid, zid;
1956                 struct mem_cgroup_per_zone *mz;
1957                 unsigned long recent_rotated[2] = {0, 0};
1958                 unsigned long recent_scanned[2] = {0, 0};
1959
1960                 for_each_online_node(nid)
1961                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1962                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1963
1964                                 recent_rotated[0] +=
1965                                         mz->reclaim_stat.recent_rotated[0];
1966                                 recent_rotated[1] +=
1967                                         mz->reclaim_stat.recent_rotated[1];
1968                                 recent_scanned[0] +=
1969                                         mz->reclaim_stat.recent_scanned[0];
1970                                 recent_scanned[1] +=
1971                                         mz->reclaim_stat.recent_scanned[1];
1972                         }
1973                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
1974                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
1975                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
1976                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
1977         }
1978 #endif
1979
1980         return 0;
1981 }
1982
1983 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
1984 {
1985         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1986
1987         return get_swappiness(memcg);
1988 }
1989
1990 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
1991                                        u64 val)
1992 {
1993         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1994         struct mem_cgroup *parent;
1995         if (val > 100)
1996                 return -EINVAL;
1997
1998         if (cgrp->parent == NULL)
1999                 return -EINVAL;
2000
2001         parent = mem_cgroup_from_cont(cgrp->parent);
2002         /* If under hierarchy, only empty-root can set this value */
2003         if ((parent->use_hierarchy) ||
2004             (memcg->use_hierarchy && !list_empty(&cgrp->children)))
2005                 return -EINVAL;
2006
2007         spin_lock(&memcg->reclaim_param_lock);
2008         memcg->swappiness = val;
2009         spin_unlock(&memcg->reclaim_param_lock);
2010
2011         return 0;
2012 }
2013
2014
2015 static struct cftype mem_cgroup_files[] = {
2016         {
2017                 .name = "usage_in_bytes",
2018                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2019                 .read_u64 = mem_cgroup_read,
2020         },
2021         {
2022                 .name = "max_usage_in_bytes",
2023                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2024                 .trigger = mem_cgroup_reset,
2025                 .read_u64 = mem_cgroup_read,
2026         },
2027         {
2028                 .name = "limit_in_bytes",
2029                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2030                 .write_string = mem_cgroup_write,
2031                 .read_u64 = mem_cgroup_read,
2032         },
2033         {
2034                 .name = "failcnt",
2035                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2036                 .trigger = mem_cgroup_reset,
2037                 .read_u64 = mem_cgroup_read,
2038         },
2039         {
2040                 .name = "stat",
2041                 .read_map = mem_control_stat_show,
2042         },
2043         {
2044                 .name = "force_empty",
2045                 .trigger = mem_cgroup_force_empty_write,
2046         },
2047         {
2048                 .name = "use_hierarchy",
2049                 .write_u64 = mem_cgroup_hierarchy_write,
2050                 .read_u64 = mem_cgroup_hierarchy_read,
2051         },
2052         {
2053                 .name = "swappiness",
2054                 .read_u64 = mem_cgroup_swappiness_read,
2055                 .write_u64 = mem_cgroup_swappiness_write,
2056         },
2057 };
2058
2059 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2060 static struct cftype memsw_cgroup_files[] = {
2061         {
2062                 .name = "memsw.usage_in_bytes",
2063                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2064                 .read_u64 = mem_cgroup_read,
2065         },
2066         {
2067                 .name = "memsw.max_usage_in_bytes",
2068                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2069                 .trigger = mem_cgroup_reset,
2070                 .read_u64 = mem_cgroup_read,
2071         },
2072         {
2073                 .name = "memsw.limit_in_bytes",
2074                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2075                 .write_string = mem_cgroup_write,
2076                 .read_u64 = mem_cgroup_read,
2077         },
2078         {
2079                 .name = "memsw.failcnt",
2080                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2081                 .trigger = mem_cgroup_reset,
2082                 .read_u64 = mem_cgroup_read,
2083         },
2084 };
2085
2086 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2087 {
2088         if (!do_swap_account)
2089                 return 0;
2090         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2091                                 ARRAY_SIZE(memsw_cgroup_files));
2092 };
2093 #else
2094 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2095 {
2096         return 0;
2097 }
2098 #endif
2099
2100 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2101 {
2102         struct mem_cgroup_per_node *pn;
2103         struct mem_cgroup_per_zone *mz;
2104         enum lru_list l;
2105         int zone, tmp = node;
2106         /*
2107          * This routine is called against possible nodes.
2108          * But it's BUG to call kmalloc() against offline node.
2109          *
2110          * TODO: this routine can waste much memory for nodes which will
2111          *       never be onlined. It's better to use memory hotplug callback
2112          *       function.
2113          */
2114         if (!node_state(node, N_NORMAL_MEMORY))
2115                 tmp = -1;
2116         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2117         if (!pn)
2118                 return 1;
2119
2120         mem->info.nodeinfo[node] = pn;
2121         memset(pn, 0, sizeof(*pn));
2122
2123         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2124                 mz = &pn->zoneinfo[zone];
2125                 for_each_lru(l)
2126                         INIT_LIST_HEAD(&mz->lists[l]);
2127         }
2128         return 0;
2129 }
2130
2131 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2132 {
2133         kfree(mem->info.nodeinfo[node]);
2134 }
2135
2136 static int mem_cgroup_size(void)
2137 {
2138         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2139         return sizeof(struct mem_cgroup) + cpustat_size;
2140 }
2141
2142 static struct mem_cgroup *mem_cgroup_alloc(void)
2143 {
2144         struct mem_cgroup *mem;
2145         int size = mem_cgroup_size();
2146
2147         if (size < PAGE_SIZE)
2148                 mem = kmalloc(size, GFP_KERNEL);
2149         else
2150                 mem = vmalloc(size);
2151
2152         if (mem)
2153                 memset(mem, 0, size);
2154         return mem;
2155 }
2156
2157 /*
2158  * At destroying mem_cgroup, references from swap_cgroup can remain.
2159  * (scanning all at force_empty is too costly...)
2160  *
2161  * Instead of clearing all references at force_empty, we remember
2162  * the number of reference from swap_cgroup and free mem_cgroup when
2163  * it goes down to 0.
2164  *
2165  * Removal of cgroup itself succeeds regardless of refs from swap.
2166  */
2167
2168 static void __mem_cgroup_free(struct mem_cgroup *mem)
2169 {
2170         int node;
2171
2172         for_each_node_state(node, N_POSSIBLE)
2173                 free_mem_cgroup_per_zone_info(mem, node);
2174
2175         if (mem_cgroup_size() < PAGE_SIZE)
2176                 kfree(mem);
2177         else
2178                 vfree(mem);
2179 }
2180
2181 static void mem_cgroup_get(struct mem_cgroup *mem)
2182 {
2183         atomic_inc(&mem->refcnt);
2184 }
2185
2186 static void mem_cgroup_put(struct mem_cgroup *mem)
2187 {
2188         if (atomic_dec_and_test(&mem->refcnt))
2189                 __mem_cgroup_free(mem);
2190 }
2191
2192
2193 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2194 static void __init enable_swap_cgroup(void)
2195 {
2196         if (!mem_cgroup_disabled() && really_do_swap_account)
2197                 do_swap_account = 1;
2198 }
2199 #else
2200 static void __init enable_swap_cgroup(void)
2201 {
2202 }
2203 #endif
2204
2205 static struct cgroup_subsys_state * __ref
2206 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2207 {
2208         struct mem_cgroup *mem, *parent;
2209         int node;
2210
2211         mem = mem_cgroup_alloc();
2212         if (!mem)
2213                 return ERR_PTR(-ENOMEM);
2214
2215         for_each_node_state(node, N_POSSIBLE)
2216                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2217                         goto free_out;
2218         /* root ? */
2219         if (cont->parent == NULL) {
2220                 enable_swap_cgroup();
2221                 parent = NULL;
2222         } else {
2223                 parent = mem_cgroup_from_cont(cont->parent);
2224                 mem->use_hierarchy = parent->use_hierarchy;
2225         }
2226
2227         if (parent && parent->use_hierarchy) {
2228                 res_counter_init(&mem->res, &parent->res);
2229                 res_counter_init(&mem->memsw, &parent->memsw);
2230         } else {
2231                 res_counter_init(&mem->res, NULL);
2232                 res_counter_init(&mem->memsw, NULL);
2233         }
2234         mem->last_scanned_child = NULL;
2235         spin_lock_init(&mem->reclaim_param_lock);
2236
2237         if (parent)
2238                 mem->swappiness = get_swappiness(parent);
2239         atomic_set(&mem->refcnt, 1);
2240         return &mem->css;
2241 free_out:
2242         __mem_cgroup_free(mem);
2243         return ERR_PTR(-ENOMEM);
2244 }
2245
2246 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2247                                         struct cgroup *cont)
2248 {
2249         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2250         mem_cgroup_force_empty(mem, false);
2251 }
2252
2253 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2254                                 struct cgroup *cont)
2255 {
2256         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2257         struct mem_cgroup *last_scanned_child = mem->last_scanned_child;
2258
2259         if (last_scanned_child) {
2260                 VM_BUG_ON(!mem_cgroup_is_obsolete(last_scanned_child));
2261                 mem_cgroup_put(last_scanned_child);
2262         }
2263         mem_cgroup_put(mem);
2264 }
2265
2266 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2267                                 struct cgroup *cont)
2268 {
2269         int ret;
2270
2271         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2272                                 ARRAY_SIZE(mem_cgroup_files));
2273
2274         if (!ret)
2275                 ret = register_memsw_files(cont, ss);
2276         return ret;
2277 }
2278
2279 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2280                                 struct cgroup *cont,
2281                                 struct cgroup *old_cont,
2282                                 struct task_struct *p)
2283 {
2284         mutex_lock(&memcg_tasklist);
2285         /*
2286          * FIXME: It's better to move charges of this process from old
2287          * memcg to new memcg. But it's just on TODO-List now.
2288          */
2289         mutex_unlock(&memcg_tasklist);
2290 }
2291
2292 struct cgroup_subsys mem_cgroup_subsys = {
2293         .name = "memory",
2294         .subsys_id = mem_cgroup_subsys_id,
2295         .create = mem_cgroup_create,
2296         .pre_destroy = mem_cgroup_pre_destroy,
2297         .destroy = mem_cgroup_destroy,
2298         .populate = mem_cgroup_populate,
2299         .attach = mem_cgroup_move_task,
2300         .early_init = 0,
2301 };
2302
2303 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2304
2305 static int __init disable_swap_account(char *s)
2306 {
2307         really_do_swap_account = 0;
2308         return 1;
2309 }
2310 __setup("noswapaccount", disable_swap_account);
2311 #endif