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