memcg: charge swapcache to proper memcg
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  */
19
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/limits.h>
31 #include <linux/mutex.h>
32 #include <linux/slab.h>
33 #include <linux/swap.h>
34 #include <linux/spinlock.h>
35 #include <linux/fs.h>
36 #include <linux/seq_file.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mm_inline.h>
39 #include <linux/page_cgroup.h>
40 #include "internal.h"
41
42 #include <asm/uaccess.h>
43
44 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
45 #define MEM_CGROUP_RECLAIM_RETRIES      5
46
47 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
48 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 0 */
49 int do_swap_account __read_mostly;
50 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
51 #else
52 #define do_swap_account         (0)
53 #endif
54
55 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
56
57 /*
58  * Statistics for memory cgroup.
59  */
60 enum mem_cgroup_stat_index {
61         /*
62          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
63          */
64         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
65         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
66         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
67         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
68
69         MEM_CGROUP_STAT_NSTATS,
70 };
71
72 struct mem_cgroup_stat_cpu {
73         s64 count[MEM_CGROUP_STAT_NSTATS];
74 } ____cacheline_aligned_in_smp;
75
76 struct mem_cgroup_stat {
77         struct mem_cgroup_stat_cpu cpustat[0];
78 };
79
80 /*
81  * For accounting under irq disable, no need for increment preempt count.
82  */
83 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
84                 enum mem_cgroup_stat_index idx, int val)
85 {
86         stat->count[idx] += val;
87 }
88
89 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
90                 enum mem_cgroup_stat_index idx)
91 {
92         int cpu;
93         s64 ret = 0;
94         for_each_possible_cpu(cpu)
95                 ret += stat->cpustat[cpu].count[idx];
96         return ret;
97 }
98
99 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
100 {
101         s64 ret;
102
103         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
104         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
105         return ret;
106 }
107
108 /*
109  * per-zone information in memory controller.
110  */
111 struct mem_cgroup_per_zone {
112         /*
113          * spin_lock to protect the per cgroup LRU
114          */
115         struct list_head        lists[NR_LRU_LISTS];
116         unsigned long           count[NR_LRU_LISTS];
117
118         struct zone_reclaim_stat reclaim_stat;
119 };
120 /* Macro for accessing counter */
121 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
122
123 struct mem_cgroup_per_node {
124         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
125 };
126
127 struct mem_cgroup_lru_info {
128         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
129 };
130
131 /*
132  * The memory controller data structure. The memory controller controls both
133  * page cache and RSS per cgroup. We would eventually like to provide
134  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
135  * to help the administrator determine what knobs to tune.
136  *
137  * TODO: Add a water mark for the memory controller. Reclaim will begin when
138  * we hit the water mark. May be even add a low water mark, such that
139  * no reclaim occurs from a cgroup at it's low water mark, this is
140  * a feature that will be implemented much later in the future.
141  */
142 struct mem_cgroup {
143         struct cgroup_subsys_state css;
144         /*
145          * the counter to account for memory usage
146          */
147         struct res_counter res;
148         /*
149          * the counter to account for mem+swap usage.
150          */
151         struct res_counter memsw;
152         /*
153          * Per cgroup active and inactive list, similar to the
154          * per zone LRU lists.
155          */
156         struct mem_cgroup_lru_info info;
157
158         /*
159           protect against reclaim related member.
160         */
161         spinlock_t reclaim_param_lock;
162
163         int     prev_priority;  /* for recording reclaim priority */
164
165         /*
166          * While reclaiming in a hiearchy, we cache the last child we
167          * reclaimed from.
168          */
169         int last_scanned_child;
170         /*
171          * Should the accounting and control be hierarchical, per subtree?
172          */
173         bool use_hierarchy;
174         unsigned long   last_oom_jiffies;
175         atomic_t        refcnt;
176
177         unsigned int    swappiness;
178
179         /*
180          * statistics. This must be placed at the end of memcg.
181          */
182         struct mem_cgroup_stat stat;
183 };
184
185 enum charge_type {
186         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
187         MEM_CGROUP_CHARGE_TYPE_MAPPED,
188         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
189         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
190         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
191         NR_CHARGE_TYPE,
192 };
193
194 /* only for here (for easy reading.) */
195 #define PCGF_CACHE      (1UL << PCG_CACHE)
196 #define PCGF_USED       (1UL << PCG_USED)
197 #define PCGF_LOCK       (1UL << PCG_LOCK)
198 static const unsigned long
199 pcg_default_flags[NR_CHARGE_TYPE] = {
200         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
201         PCGF_USED | PCGF_LOCK, /* Anon */
202         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
203         0, /* FORCE */
204 };
205
206 /* for encoding cft->private value on file */
207 #define _MEM                    (0)
208 #define _MEMSWAP                (1)
209 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
210 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
211 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
212
213 static void mem_cgroup_get(struct mem_cgroup *mem);
214 static void mem_cgroup_put(struct mem_cgroup *mem);
215 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
216
217 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
218                                          struct page_cgroup *pc,
219                                          bool charge)
220 {
221         int val = (charge)? 1 : -1;
222         struct mem_cgroup_stat *stat = &mem->stat;
223         struct mem_cgroup_stat_cpu *cpustat;
224         int cpu = get_cpu();
225
226         cpustat = &stat->cpustat[cpu];
227         if (PageCgroupCache(pc))
228                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
229         else
230                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
231
232         if (charge)
233                 __mem_cgroup_stat_add_safe(cpustat,
234                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
235         else
236                 __mem_cgroup_stat_add_safe(cpustat,
237                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
238         put_cpu();
239 }
240
241 static struct mem_cgroup_per_zone *
242 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
243 {
244         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
245 }
246
247 static struct mem_cgroup_per_zone *
248 page_cgroup_zoneinfo(struct page_cgroup *pc)
249 {
250         struct mem_cgroup *mem = pc->mem_cgroup;
251         int nid = page_cgroup_nid(pc);
252         int zid = page_cgroup_zid(pc);
253
254         if (!mem)
255                 return NULL;
256
257         return mem_cgroup_zoneinfo(mem, nid, zid);
258 }
259
260 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
261                                         enum lru_list idx)
262 {
263         int nid, zid;
264         struct mem_cgroup_per_zone *mz;
265         u64 total = 0;
266
267         for_each_online_node(nid)
268                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
269                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
270                         total += MEM_CGROUP_ZSTAT(mz, idx);
271                 }
272         return total;
273 }
274
275 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
276 {
277         return container_of(cgroup_subsys_state(cont,
278                                 mem_cgroup_subsys_id), struct mem_cgroup,
279                                 css);
280 }
281
282 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
283 {
284         /*
285          * mm_update_next_owner() may clear mm->owner to NULL
286          * if it races with swapoff, page migration, etc.
287          * So this can be called with p == NULL.
288          */
289         if (unlikely(!p))
290                 return NULL;
291
292         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
293                                 struct mem_cgroup, css);
294 }
295
296 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
297 {
298         struct mem_cgroup *mem = NULL;
299
300         if (!mm)
301                 return NULL;
302         /*
303          * Because we have no locks, mm->owner's may be being moved to other
304          * cgroup. We use css_tryget() here even if this looks
305          * pessimistic (rather than adding locks here).
306          */
307         rcu_read_lock();
308         do {
309                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
310                 if (unlikely(!mem))
311                         break;
312         } while (!css_tryget(&mem->css));
313         rcu_read_unlock();
314         return mem;
315 }
316
317 static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem)
318 {
319         if (!mem)
320                 return true;
321         return css_is_removed(&mem->css);
322 }
323
324
325 /*
326  * Call callback function against all cgroup under hierarchy tree.
327  */
328 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
329                           int (*func)(struct mem_cgroup *, void *))
330 {
331         int found, ret, nextid;
332         struct cgroup_subsys_state *css;
333         struct mem_cgroup *mem;
334
335         if (!root->use_hierarchy)
336                 return (*func)(root, data);
337
338         nextid = 1;
339         do {
340                 ret = 0;
341                 mem = NULL;
342
343                 rcu_read_lock();
344                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
345                                    &found);
346                 if (css && css_tryget(css))
347                         mem = container_of(css, struct mem_cgroup, css);
348                 rcu_read_unlock();
349
350                 if (mem) {
351                         ret = (*func)(mem, data);
352                         css_put(&mem->css);
353                 }
354                 nextid = found + 1;
355         } while (!ret && css);
356
357         return ret;
358 }
359
360 /*
361  * Following LRU functions are allowed to be used without PCG_LOCK.
362  * Operations are called by routine of global LRU independently from memcg.
363  * What we have to take care of here is validness of pc->mem_cgroup.
364  *
365  * Changes to pc->mem_cgroup happens when
366  * 1. charge
367  * 2. moving account
368  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
369  * It is added to LRU before charge.
370  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
371  * When moving account, the page is not on LRU. It's isolated.
372  */
373
374 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
375 {
376         struct page_cgroup *pc;
377         struct mem_cgroup *mem;
378         struct mem_cgroup_per_zone *mz;
379
380         if (mem_cgroup_disabled())
381                 return;
382         pc = lookup_page_cgroup(page);
383         /* can happen while we handle swapcache. */
384         if (list_empty(&pc->lru) || !pc->mem_cgroup)
385                 return;
386         /*
387          * We don't check PCG_USED bit. It's cleared when the "page" is finally
388          * removed from global LRU.
389          */
390         mz = page_cgroup_zoneinfo(pc);
391         mem = pc->mem_cgroup;
392         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
393         list_del_init(&pc->lru);
394         return;
395 }
396
397 void mem_cgroup_del_lru(struct page *page)
398 {
399         mem_cgroup_del_lru_list(page, page_lru(page));
400 }
401
402 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
403 {
404         struct mem_cgroup_per_zone *mz;
405         struct page_cgroup *pc;
406
407         if (mem_cgroup_disabled())
408                 return;
409
410         pc = lookup_page_cgroup(page);
411         /*
412          * Used bit is set without atomic ops but after smp_wmb().
413          * For making pc->mem_cgroup visible, insert smp_rmb() here.
414          */
415         smp_rmb();
416         /* unused page is not rotated. */
417         if (!PageCgroupUsed(pc))
418                 return;
419         mz = page_cgroup_zoneinfo(pc);
420         list_move(&pc->lru, &mz->lists[lru]);
421 }
422
423 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
424 {
425         struct page_cgroup *pc;
426         struct mem_cgroup_per_zone *mz;
427
428         if (mem_cgroup_disabled())
429                 return;
430         pc = lookup_page_cgroup(page);
431         /*
432          * Used bit is set without atomic ops but after smp_wmb().
433          * For making pc->mem_cgroup visible, insert smp_rmb() here.
434          */
435         smp_rmb();
436         if (!PageCgroupUsed(pc))
437                 return;
438
439         mz = page_cgroup_zoneinfo(pc);
440         MEM_CGROUP_ZSTAT(mz, lru) += 1;
441         list_add(&pc->lru, &mz->lists[lru]);
442 }
443
444 /*
445  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
446  * lru because the page may.be reused after it's fully uncharged (because of
447  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
448  * it again. This function is only used to charge SwapCache. It's done under
449  * lock_page and expected that zone->lru_lock is never held.
450  */
451 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
452 {
453         unsigned long flags;
454         struct zone *zone = page_zone(page);
455         struct page_cgroup *pc = lookup_page_cgroup(page);
456
457         spin_lock_irqsave(&zone->lru_lock, flags);
458         /*
459          * Forget old LRU when this page_cgroup is *not* used. This Used bit
460          * is guarded by lock_page() because the page is SwapCache.
461          */
462         if (!PageCgroupUsed(pc))
463                 mem_cgroup_del_lru_list(page, page_lru(page));
464         spin_unlock_irqrestore(&zone->lru_lock, flags);
465 }
466
467 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
468 {
469         unsigned long flags;
470         struct zone *zone = page_zone(page);
471         struct page_cgroup *pc = lookup_page_cgroup(page);
472
473         spin_lock_irqsave(&zone->lru_lock, flags);
474         /* link when the page is linked to LRU but page_cgroup isn't */
475         if (PageLRU(page) && list_empty(&pc->lru))
476                 mem_cgroup_add_lru_list(page, page_lru(page));
477         spin_unlock_irqrestore(&zone->lru_lock, flags);
478 }
479
480
481 void mem_cgroup_move_lists(struct page *page,
482                            enum lru_list from, enum lru_list to)
483 {
484         if (mem_cgroup_disabled())
485                 return;
486         mem_cgroup_del_lru_list(page, from);
487         mem_cgroup_add_lru_list(page, to);
488 }
489
490 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
491 {
492         int ret;
493         struct mem_cgroup *curr = NULL;
494
495         task_lock(task);
496         rcu_read_lock();
497         curr = try_get_mem_cgroup_from_mm(task->mm);
498         rcu_read_unlock();
499         task_unlock(task);
500         if (!curr)
501                 return 0;
502         if (curr->use_hierarchy)
503                 ret = css_is_ancestor(&curr->css, &mem->css);
504         else
505                 ret = (curr == mem);
506         css_put(&curr->css);
507         return ret;
508 }
509
510 /*
511  * prev_priority control...this will be used in memory reclaim path.
512  */
513 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
514 {
515         int prev_priority;
516
517         spin_lock(&mem->reclaim_param_lock);
518         prev_priority = mem->prev_priority;
519         spin_unlock(&mem->reclaim_param_lock);
520
521         return prev_priority;
522 }
523
524 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
525 {
526         spin_lock(&mem->reclaim_param_lock);
527         if (priority < mem->prev_priority)
528                 mem->prev_priority = priority;
529         spin_unlock(&mem->reclaim_param_lock);
530 }
531
532 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
533 {
534         spin_lock(&mem->reclaim_param_lock);
535         mem->prev_priority = priority;
536         spin_unlock(&mem->reclaim_param_lock);
537 }
538
539 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
540 {
541         unsigned long active;
542         unsigned long inactive;
543         unsigned long gb;
544         unsigned long inactive_ratio;
545
546         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
547         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
548
549         gb = (inactive + active) >> (30 - PAGE_SHIFT);
550         if (gb)
551                 inactive_ratio = int_sqrt(10 * gb);
552         else
553                 inactive_ratio = 1;
554
555         if (present_pages) {
556                 present_pages[0] = inactive;
557                 present_pages[1] = active;
558         }
559
560         return inactive_ratio;
561 }
562
563 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
564 {
565         unsigned long active;
566         unsigned long inactive;
567         unsigned long present_pages[2];
568         unsigned long inactive_ratio;
569
570         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
571
572         inactive = present_pages[0];
573         active = present_pages[1];
574
575         if (inactive * inactive_ratio < active)
576                 return 1;
577
578         return 0;
579 }
580
581 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
582                                        struct zone *zone,
583                                        enum lru_list lru)
584 {
585         int nid = zone->zone_pgdat->node_id;
586         int zid = zone_idx(zone);
587         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
588
589         return MEM_CGROUP_ZSTAT(mz, lru);
590 }
591
592 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
593                                                       struct zone *zone)
594 {
595         int nid = zone->zone_pgdat->node_id;
596         int zid = zone_idx(zone);
597         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
598
599         return &mz->reclaim_stat;
600 }
601
602 struct zone_reclaim_stat *
603 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
604 {
605         struct page_cgroup *pc;
606         struct mem_cgroup_per_zone *mz;
607
608         if (mem_cgroup_disabled())
609                 return NULL;
610
611         pc = lookup_page_cgroup(page);
612         /*
613          * Used bit is set without atomic ops but after smp_wmb().
614          * For making pc->mem_cgroup visible, insert smp_rmb() here.
615          */
616         smp_rmb();
617         if (!PageCgroupUsed(pc))
618                 return NULL;
619
620         mz = page_cgroup_zoneinfo(pc);
621         if (!mz)
622                 return NULL;
623
624         return &mz->reclaim_stat;
625 }
626
627 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
628                                         struct list_head *dst,
629                                         unsigned long *scanned, int order,
630                                         int mode, struct zone *z,
631                                         struct mem_cgroup *mem_cont,
632                                         int active, int file)
633 {
634         unsigned long nr_taken = 0;
635         struct page *page;
636         unsigned long scan;
637         LIST_HEAD(pc_list);
638         struct list_head *src;
639         struct page_cgroup *pc, *tmp;
640         int nid = z->zone_pgdat->node_id;
641         int zid = zone_idx(z);
642         struct mem_cgroup_per_zone *mz;
643         int lru = LRU_FILE * !!file + !!active;
644
645         BUG_ON(!mem_cont);
646         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
647         src = &mz->lists[lru];
648
649         scan = 0;
650         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
651                 if (scan >= nr_to_scan)
652                         break;
653
654                 page = pc->page;
655                 if (unlikely(!PageCgroupUsed(pc)))
656                         continue;
657                 if (unlikely(!PageLRU(page)))
658                         continue;
659
660                 scan++;
661                 if (__isolate_lru_page(page, mode, file) == 0) {
662                         list_move(&page->lru, dst);
663                         nr_taken++;
664                 }
665         }
666
667         *scanned = scan;
668         return nr_taken;
669 }
670
671 #define mem_cgroup_from_res_counter(counter, member)    \
672         container_of(counter, struct mem_cgroup, member)
673
674 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
675 {
676         if (do_swap_account) {
677                 if (res_counter_check_under_limit(&mem->res) &&
678                         res_counter_check_under_limit(&mem->memsw))
679                         return true;
680         } else
681                 if (res_counter_check_under_limit(&mem->res))
682                         return true;
683         return false;
684 }
685
686 static unsigned int get_swappiness(struct mem_cgroup *memcg)
687 {
688         struct cgroup *cgrp = memcg->css.cgroup;
689         unsigned int swappiness;
690
691         /* root ? */
692         if (cgrp->parent == NULL)
693                 return vm_swappiness;
694
695         spin_lock(&memcg->reclaim_param_lock);
696         swappiness = memcg->swappiness;
697         spin_unlock(&memcg->reclaim_param_lock);
698
699         return swappiness;
700 }
701
702 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
703 {
704         int *val = data;
705         (*val)++;
706         return 0;
707 }
708
709 /**
710  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
711  * @memcg: The memory cgroup that went over limit
712  * @p: Task that is going to be killed
713  *
714  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
715  * enabled
716  */
717 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
718 {
719         struct cgroup *task_cgrp;
720         struct cgroup *mem_cgrp;
721         /*
722          * Need a buffer in BSS, can't rely on allocations. The code relies
723          * on the assumption that OOM is serialized for memory controller.
724          * If this assumption is broken, revisit this code.
725          */
726         static char memcg_name[PATH_MAX];
727         int ret;
728
729         if (!memcg)
730                 return;
731
732
733         rcu_read_lock();
734
735         mem_cgrp = memcg->css.cgroup;
736         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
737
738         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
739         if (ret < 0) {
740                 /*
741                  * Unfortunately, we are unable to convert to a useful name
742                  * But we'll still print out the usage information
743                  */
744                 rcu_read_unlock();
745                 goto done;
746         }
747         rcu_read_unlock();
748
749         printk(KERN_INFO "Task in %s killed", memcg_name);
750
751         rcu_read_lock();
752         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
753         if (ret < 0) {
754                 rcu_read_unlock();
755                 goto done;
756         }
757         rcu_read_unlock();
758
759         /*
760          * Continues from above, so we don't need an KERN_ level
761          */
762         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
763 done:
764
765         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
766                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
767                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
768                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
769         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
770                 "failcnt %llu\n",
771                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
772                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
773                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
774 }
775
776 /*
777  * This function returns the number of memcg under hierarchy tree. Returns
778  * 1(self count) if no children.
779  */
780 static int mem_cgroup_count_children(struct mem_cgroup *mem)
781 {
782         int num = 0;
783         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
784         return num;
785 }
786
787 /*
788  * Visit the first child (need not be the first child as per the ordering
789  * of the cgroup list, since we track last_scanned_child) of @mem and use
790  * that to reclaim free pages from.
791  */
792 static struct mem_cgroup *
793 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
794 {
795         struct mem_cgroup *ret = NULL;
796         struct cgroup_subsys_state *css;
797         int nextid, found;
798
799         if (!root_mem->use_hierarchy) {
800                 css_get(&root_mem->css);
801                 ret = root_mem;
802         }
803
804         while (!ret) {
805                 rcu_read_lock();
806                 nextid = root_mem->last_scanned_child + 1;
807                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
808                                    &found);
809                 if (css && css_tryget(css))
810                         ret = container_of(css, struct mem_cgroup, css);
811
812                 rcu_read_unlock();
813                 /* Updates scanning parameter */
814                 spin_lock(&root_mem->reclaim_param_lock);
815                 if (!css) {
816                         /* this means start scan from ID:1 */
817                         root_mem->last_scanned_child = 0;
818                 } else
819                         root_mem->last_scanned_child = found;
820                 spin_unlock(&root_mem->reclaim_param_lock);
821         }
822
823         return ret;
824 }
825
826 /*
827  * Scan the hierarchy if needed to reclaim memory. We remember the last child
828  * we reclaimed from, so that we don't end up penalizing one child extensively
829  * based on its position in the children list.
830  *
831  * root_mem is the original ancestor that we've been reclaim from.
832  *
833  * We give up and return to the caller when we visit root_mem twice.
834  * (other groups can be removed while we're walking....)
835  *
836  * If shrink==true, for avoiding to free too much, this returns immedieately.
837  */
838 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
839                                    gfp_t gfp_mask, bool noswap, bool shrink)
840 {
841         struct mem_cgroup *victim;
842         int ret, total = 0;
843         int loop = 0;
844
845         while (loop < 2) {
846                 victim = mem_cgroup_select_victim(root_mem);
847                 if (victim == root_mem)
848                         loop++;
849                 if (!mem_cgroup_local_usage(&victim->stat)) {
850                         /* this cgroup's local usage == 0 */
851                         css_put(&victim->css);
852                         continue;
853                 }
854                 /* we use swappiness of local cgroup */
855                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
856                                                    get_swappiness(victim));
857                 css_put(&victim->css);
858                 /*
859                  * At shrinking usage, we can't check we should stop here or
860                  * reclaim more. It's depends on callers. last_scanned_child
861                  * will work enough for keeping fairness under tree.
862                  */
863                 if (shrink)
864                         return ret;
865                 total += ret;
866                 if (mem_cgroup_check_under_limit(root_mem))
867                         return 1 + total;
868         }
869         return total;
870 }
871
872 bool mem_cgroup_oom_called(struct task_struct *task)
873 {
874         bool ret = false;
875         struct mem_cgroup *mem;
876         struct mm_struct *mm;
877
878         rcu_read_lock();
879         mm = task->mm;
880         if (!mm)
881                 mm = &init_mm;
882         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
883         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
884                 ret = true;
885         rcu_read_unlock();
886         return ret;
887 }
888
889 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
890 {
891         mem->last_oom_jiffies = jiffies;
892         return 0;
893 }
894
895 static void record_last_oom(struct mem_cgroup *mem)
896 {
897         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
898 }
899
900
901 /*
902  * Unlike exported interface, "oom" parameter is added. if oom==true,
903  * oom-killer can be invoked.
904  */
905 static int __mem_cgroup_try_charge(struct mm_struct *mm,
906                         gfp_t gfp_mask, struct mem_cgroup **memcg,
907                         bool oom)
908 {
909         struct mem_cgroup *mem, *mem_over_limit;
910         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
911         struct res_counter *fail_res;
912
913         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
914                 /* Don't account this! */
915                 *memcg = NULL;
916                 return 0;
917         }
918
919         /*
920          * We always charge the cgroup the mm_struct belongs to.
921          * The mm_struct's mem_cgroup changes on task migration if the
922          * thread group leader migrates. It's possible that mm is not
923          * set, if so charge the init_mm (happens for pagecache usage).
924          */
925         mem = *memcg;
926         if (likely(!mem)) {
927                 mem = try_get_mem_cgroup_from_mm(mm);
928                 *memcg = mem;
929         } else {
930                 css_get(&mem->css);
931         }
932         if (unlikely(!mem))
933                 return 0;
934
935         VM_BUG_ON(mem_cgroup_is_obsolete(mem));
936
937         while (1) {
938                 int ret;
939                 bool noswap = false;
940
941                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
942                 if (likely(!ret)) {
943                         if (!do_swap_account)
944                                 break;
945                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
946                                                         &fail_res);
947                         if (likely(!ret))
948                                 break;
949                         /* mem+swap counter fails */
950                         res_counter_uncharge(&mem->res, PAGE_SIZE);
951                         noswap = true;
952                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
953                                                                         memsw);
954                 } else
955                         /* mem counter fails */
956                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
957                                                                         res);
958
959                 if (!(gfp_mask & __GFP_WAIT))
960                         goto nomem;
961
962                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
963                                                         noswap, false);
964                 if (ret)
965                         continue;
966
967                 /*
968                  * try_to_free_mem_cgroup_pages() might not give us a full
969                  * picture of reclaim. Some pages are reclaimed and might be
970                  * moved to swap cache or just unmapped from the cgroup.
971                  * Check the limit again to see if the reclaim reduced the
972                  * current usage of the cgroup before giving up
973                  *
974                  */
975                 if (mem_cgroup_check_under_limit(mem_over_limit))
976                         continue;
977
978                 if (!nr_retries--) {
979                         if (oom) {
980                                 mutex_lock(&memcg_tasklist);
981                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
982                                 mutex_unlock(&memcg_tasklist);
983                                 record_last_oom(mem_over_limit);
984                         }
985                         goto nomem;
986                 }
987         }
988         return 0;
989 nomem:
990         css_put(&mem->css);
991         return -ENOMEM;
992 }
993
994 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
995 {
996         struct mem_cgroup *mem;
997         struct page_cgroup *pc;
998         swp_entry_t ent;
999
1000         VM_BUG_ON(!PageLocked(page));
1001
1002         if (!PageSwapCache(page))
1003                 return NULL;
1004
1005         pc = lookup_page_cgroup(page);
1006         /*
1007          * Used bit of swapcache is solid under page lock.
1008          */
1009         if (PageCgroupUsed(pc))
1010                 mem = pc->mem_cgroup;
1011         else {
1012                 ent.val = page_private(page);
1013                 mem = lookup_swap_cgroup(ent);
1014         }
1015         if (!mem)
1016                 return NULL;
1017         if (!css_tryget(&mem->css))
1018                 return NULL;
1019         return mem;
1020 }
1021
1022 /*
1023  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1024  * USED state. If already USED, uncharge and return.
1025  */
1026
1027 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1028                                      struct page_cgroup *pc,
1029                                      enum charge_type ctype)
1030 {
1031         /* try_charge() can return NULL to *memcg, taking care of it. */
1032         if (!mem)
1033                 return;
1034
1035         lock_page_cgroup(pc);
1036         if (unlikely(PageCgroupUsed(pc))) {
1037                 unlock_page_cgroup(pc);
1038                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1039                 if (do_swap_account)
1040                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1041                 css_put(&mem->css);
1042                 return;
1043         }
1044         pc->mem_cgroup = mem;
1045         smp_wmb();
1046         pc->flags = pcg_default_flags[ctype];
1047
1048         mem_cgroup_charge_statistics(mem, pc, true);
1049
1050         unlock_page_cgroup(pc);
1051 }
1052
1053 /**
1054  * mem_cgroup_move_account - move account of the page
1055  * @pc: page_cgroup of the page.
1056  * @from: mem_cgroup which the page is moved from.
1057  * @to: mem_cgroup which the page is moved to. @from != @to.
1058  *
1059  * The caller must confirm following.
1060  * - page is not on LRU (isolate_page() is useful.)
1061  *
1062  * returns 0 at success,
1063  * returns -EBUSY when lock is busy or "pc" is unstable.
1064  *
1065  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1066  * new cgroup. It should be done by a caller.
1067  */
1068
1069 static int mem_cgroup_move_account(struct page_cgroup *pc,
1070         struct mem_cgroup *from, struct mem_cgroup *to)
1071 {
1072         struct mem_cgroup_per_zone *from_mz, *to_mz;
1073         int nid, zid;
1074         int ret = -EBUSY;
1075
1076         VM_BUG_ON(from == to);
1077         VM_BUG_ON(PageLRU(pc->page));
1078
1079         nid = page_cgroup_nid(pc);
1080         zid = page_cgroup_zid(pc);
1081         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1082         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1083
1084         if (!trylock_page_cgroup(pc))
1085                 return ret;
1086
1087         if (!PageCgroupUsed(pc))
1088                 goto out;
1089
1090         if (pc->mem_cgroup != from)
1091                 goto out;
1092
1093         res_counter_uncharge(&from->res, PAGE_SIZE);
1094         mem_cgroup_charge_statistics(from, pc, false);
1095         if (do_swap_account)
1096                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1097         css_put(&from->css);
1098
1099         css_get(&to->css);
1100         pc->mem_cgroup = to;
1101         mem_cgroup_charge_statistics(to, pc, true);
1102         ret = 0;
1103 out:
1104         unlock_page_cgroup(pc);
1105         return ret;
1106 }
1107
1108 /*
1109  * move charges to its parent.
1110  */
1111
1112 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1113                                   struct mem_cgroup *child,
1114                                   gfp_t gfp_mask)
1115 {
1116         struct page *page = pc->page;
1117         struct cgroup *cg = child->css.cgroup;
1118         struct cgroup *pcg = cg->parent;
1119         struct mem_cgroup *parent;
1120         int ret;
1121
1122         /* Is ROOT ? */
1123         if (!pcg)
1124                 return -EINVAL;
1125
1126
1127         parent = mem_cgroup_from_cont(pcg);
1128
1129
1130         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1131         if (ret || !parent)
1132                 return ret;
1133
1134         if (!get_page_unless_zero(page)) {
1135                 ret = -EBUSY;
1136                 goto uncharge;
1137         }
1138
1139         ret = isolate_lru_page(page);
1140
1141         if (ret)
1142                 goto cancel;
1143
1144         ret = mem_cgroup_move_account(pc, child, parent);
1145
1146         putback_lru_page(page);
1147         if (!ret) {
1148                 put_page(page);
1149                 /* drop extra refcnt by try_charge() */
1150                 css_put(&parent->css);
1151                 return 0;
1152         }
1153
1154 cancel:
1155         put_page(page);
1156 uncharge:
1157         /* drop extra refcnt by try_charge() */
1158         css_put(&parent->css);
1159         /* uncharge if move fails */
1160         res_counter_uncharge(&parent->res, PAGE_SIZE);
1161         if (do_swap_account)
1162                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1163         return ret;
1164 }
1165
1166 /*
1167  * Charge the memory controller for page usage.
1168  * Return
1169  * 0 if the charge was successful
1170  * < 0 if the cgroup is over its limit
1171  */
1172 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1173                                 gfp_t gfp_mask, enum charge_type ctype,
1174                                 struct mem_cgroup *memcg)
1175 {
1176         struct mem_cgroup *mem;
1177         struct page_cgroup *pc;
1178         int ret;
1179
1180         pc = lookup_page_cgroup(page);
1181         /* can happen at boot */
1182         if (unlikely(!pc))
1183                 return 0;
1184         prefetchw(pc);
1185
1186         mem = memcg;
1187         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1188         if (ret || !mem)
1189                 return ret;
1190
1191         __mem_cgroup_commit_charge(mem, pc, ctype);
1192         return 0;
1193 }
1194
1195 int mem_cgroup_newpage_charge(struct page *page,
1196                               struct mm_struct *mm, gfp_t gfp_mask)
1197 {
1198         if (mem_cgroup_disabled())
1199                 return 0;
1200         if (PageCompound(page))
1201                 return 0;
1202         /*
1203          * If already mapped, we don't have to account.
1204          * If page cache, page->mapping has address_space.
1205          * But page->mapping may have out-of-use anon_vma pointer,
1206          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1207          * is NULL.
1208          */
1209         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1210                 return 0;
1211         if (unlikely(!mm))
1212                 mm = &init_mm;
1213         return mem_cgroup_charge_common(page, mm, gfp_mask,
1214                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1215 }
1216
1217 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1218                                 gfp_t gfp_mask)
1219 {
1220         struct mem_cgroup *mem = NULL;
1221         int ret;
1222
1223         if (mem_cgroup_disabled())
1224                 return 0;
1225         if (PageCompound(page))
1226                 return 0;
1227         /*
1228          * Corner case handling. This is called from add_to_page_cache()
1229          * in usual. But some FS (shmem) precharges this page before calling it
1230          * and call add_to_page_cache() with GFP_NOWAIT.
1231          *
1232          * For GFP_NOWAIT case, the page may be pre-charged before calling
1233          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1234          * charge twice. (It works but has to pay a bit larger cost.)
1235          * And when the page is SwapCache, it should take swap information
1236          * into account. This is under lock_page() now.
1237          */
1238         if (!(gfp_mask & __GFP_WAIT)) {
1239                 struct page_cgroup *pc;
1240
1241
1242                 pc = lookup_page_cgroup(page);
1243                 if (!pc)
1244                         return 0;
1245                 lock_page_cgroup(pc);
1246                 if (PageCgroupUsed(pc)) {
1247                         unlock_page_cgroup(pc);
1248                         return 0;
1249                 }
1250                 unlock_page_cgroup(pc);
1251         }
1252
1253         if (do_swap_account && PageSwapCache(page)) {
1254                 mem = try_get_mem_cgroup_from_swapcache(page);
1255                 if (mem)
1256                         mm = NULL;
1257                   else
1258                         mem = NULL;
1259                 /* SwapCache may be still linked to LRU now. */
1260                 mem_cgroup_lru_del_before_commit_swapcache(page);
1261         }
1262
1263         if (unlikely(!mm && !mem))
1264                 mm = &init_mm;
1265
1266         if (page_is_file_cache(page))
1267                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1268                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1269
1270         ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1271                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1272         if (mem)
1273                 css_put(&mem->css);
1274         if (PageSwapCache(page))
1275                 mem_cgroup_lru_add_after_commit_swapcache(page);
1276
1277         if (do_swap_account && !ret && PageSwapCache(page)) {
1278                 swp_entry_t ent = {.val = page_private(page)};
1279                 /* avoid double counting */
1280                 mem = swap_cgroup_record(ent, NULL);
1281                 if (mem) {
1282                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1283                         mem_cgroup_put(mem);
1284                 }
1285         }
1286         return ret;
1287 }
1288
1289 /*
1290  * While swap-in, try_charge -> commit or cancel, the page is locked.
1291  * And when try_charge() successfully returns, one refcnt to memcg without
1292  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1293  * "commit()" or removed by "cancel()"
1294  */
1295 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1296                                  struct page *page,
1297                                  gfp_t mask, struct mem_cgroup **ptr)
1298 {
1299         struct mem_cgroup *mem;
1300         int ret;
1301
1302         if (mem_cgroup_disabled())
1303                 return 0;
1304
1305         if (!do_swap_account)
1306                 goto charge_cur_mm;
1307         /*
1308          * A racing thread's fault, or swapoff, may have already updated
1309          * the pte, and even removed page from swap cache: return success
1310          * to go on to do_swap_page()'s pte_same() test, which should fail.
1311          */
1312         if (!PageSwapCache(page))
1313                 return 0;
1314         mem = try_get_mem_cgroup_from_swapcache(page);
1315         if (!mem)
1316                 goto charge_cur_mm;
1317         *ptr = mem;
1318         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1319         /* drop extra refcnt from tryget */
1320         css_put(&mem->css);
1321         return ret;
1322 charge_cur_mm:
1323         if (unlikely(!mm))
1324                 mm = &init_mm;
1325         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1326 }
1327
1328 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1329 {
1330         struct page_cgroup *pc;
1331
1332         if (mem_cgroup_disabled())
1333                 return;
1334         if (!ptr)
1335                 return;
1336         pc = lookup_page_cgroup(page);
1337         mem_cgroup_lru_del_before_commit_swapcache(page);
1338         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1339         mem_cgroup_lru_add_after_commit_swapcache(page);
1340         /*
1341          * Now swap is on-memory. This means this page may be
1342          * counted both as mem and swap....double count.
1343          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1344          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1345          * may call delete_from_swap_cache() before reach here.
1346          */
1347         if (do_swap_account && PageSwapCache(page)) {
1348                 swp_entry_t ent = {.val = page_private(page)};
1349                 struct mem_cgroup *memcg;
1350                 memcg = swap_cgroup_record(ent, NULL);
1351                 if (memcg) {
1352                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1353                         mem_cgroup_put(memcg);
1354                 }
1355
1356         }
1357         /* add this page(page_cgroup) to the LRU we want. */
1358
1359 }
1360
1361 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1362 {
1363         if (mem_cgroup_disabled())
1364                 return;
1365         if (!mem)
1366                 return;
1367         res_counter_uncharge(&mem->res, PAGE_SIZE);
1368         if (do_swap_account)
1369                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1370         css_put(&mem->css);
1371 }
1372
1373
1374 /*
1375  * uncharge if !page_mapped(page)
1376  */
1377 static struct mem_cgroup *
1378 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1379 {
1380         struct page_cgroup *pc;
1381         struct mem_cgroup *mem = NULL;
1382         struct mem_cgroup_per_zone *mz;
1383
1384         if (mem_cgroup_disabled())
1385                 return NULL;
1386
1387         if (PageSwapCache(page))
1388                 return NULL;
1389
1390         /*
1391          * Check if our page_cgroup is valid
1392          */
1393         pc = lookup_page_cgroup(page);
1394         if (unlikely(!pc || !PageCgroupUsed(pc)))
1395                 return NULL;
1396
1397         lock_page_cgroup(pc);
1398
1399         mem = pc->mem_cgroup;
1400
1401         if (!PageCgroupUsed(pc))
1402                 goto unlock_out;
1403
1404         switch (ctype) {
1405         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1406                 if (page_mapped(page))
1407                         goto unlock_out;
1408                 break;
1409         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1410                 if (!PageAnon(page)) {  /* Shared memory */
1411                         if (page->mapping && !page_is_file_cache(page))
1412                                 goto unlock_out;
1413                 } else if (page_mapped(page)) /* Anon */
1414                                 goto unlock_out;
1415                 break;
1416         default:
1417                 break;
1418         }
1419
1420         res_counter_uncharge(&mem->res, PAGE_SIZE);
1421         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1422                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1423         mem_cgroup_charge_statistics(mem, pc, false);
1424
1425         ClearPageCgroupUsed(pc);
1426         /*
1427          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1428          * freed from LRU. This is safe because uncharged page is expected not
1429          * to be reused (freed soon). Exception is SwapCache, it's handled by
1430          * special functions.
1431          */
1432
1433         mz = page_cgroup_zoneinfo(pc);
1434         unlock_page_cgroup(pc);
1435
1436         /* at swapout, this memcg will be accessed to record to swap */
1437         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1438                 css_put(&mem->css);
1439
1440         return mem;
1441
1442 unlock_out:
1443         unlock_page_cgroup(pc);
1444         return NULL;
1445 }
1446
1447 void mem_cgroup_uncharge_page(struct page *page)
1448 {
1449         /* early check. */
1450         if (page_mapped(page))
1451                 return;
1452         if (page->mapping && !PageAnon(page))
1453                 return;
1454         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1455 }
1456
1457 void mem_cgroup_uncharge_cache_page(struct page *page)
1458 {
1459         VM_BUG_ON(page_mapped(page));
1460         VM_BUG_ON(page->mapping);
1461         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1462 }
1463
1464 /*
1465  * called from __delete_from_swap_cache() and drop "page" account.
1466  * memcg information is recorded to swap_cgroup of "ent"
1467  */
1468 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1469 {
1470         struct mem_cgroup *memcg;
1471
1472         memcg = __mem_cgroup_uncharge_common(page,
1473                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1474         /* record memcg information */
1475         if (do_swap_account && memcg) {
1476                 swap_cgroup_record(ent, memcg);
1477                 mem_cgroup_get(memcg);
1478         }
1479         if (memcg)
1480                 css_put(&memcg->css);
1481 }
1482
1483 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1484 /*
1485  * called from swap_entry_free(). remove record in swap_cgroup and
1486  * uncharge "memsw" account.
1487  */
1488 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1489 {
1490         struct mem_cgroup *memcg;
1491
1492         if (!do_swap_account)
1493                 return;
1494
1495         memcg = swap_cgroup_record(ent, NULL);
1496         if (memcg) {
1497                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1498                 mem_cgroup_put(memcg);
1499         }
1500 }
1501 #endif
1502
1503 /*
1504  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1505  * page belongs to.
1506  */
1507 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1508 {
1509         struct page_cgroup *pc;
1510         struct mem_cgroup *mem = NULL;
1511         int ret = 0;
1512
1513         if (mem_cgroup_disabled())
1514                 return 0;
1515
1516         pc = lookup_page_cgroup(page);
1517         lock_page_cgroup(pc);
1518         if (PageCgroupUsed(pc)) {
1519                 mem = pc->mem_cgroup;
1520                 css_get(&mem->css);
1521         }
1522         unlock_page_cgroup(pc);
1523
1524         if (mem) {
1525                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1526                 css_put(&mem->css);
1527         }
1528         *ptr = mem;
1529         return ret;
1530 }
1531
1532 /* remove redundant charge if migration failed*/
1533 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1534                 struct page *oldpage, struct page *newpage)
1535 {
1536         struct page *target, *unused;
1537         struct page_cgroup *pc;
1538         enum charge_type ctype;
1539
1540         if (!mem)
1541                 return;
1542
1543         /* at migration success, oldpage->mapping is NULL. */
1544         if (oldpage->mapping) {
1545                 target = oldpage;
1546                 unused = NULL;
1547         } else {
1548                 target = newpage;
1549                 unused = oldpage;
1550         }
1551
1552         if (PageAnon(target))
1553                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1554         else if (page_is_file_cache(target))
1555                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1556         else
1557                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1558
1559         /* unused page is not on radix-tree now. */
1560         if (unused)
1561                 __mem_cgroup_uncharge_common(unused, ctype);
1562
1563         pc = lookup_page_cgroup(target);
1564         /*
1565          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1566          * So, double-counting is effectively avoided.
1567          */
1568         __mem_cgroup_commit_charge(mem, pc, ctype);
1569
1570         /*
1571          * Both of oldpage and newpage are still under lock_page().
1572          * Then, we don't have to care about race in radix-tree.
1573          * But we have to be careful that this page is unmapped or not.
1574          *
1575          * There is a case for !page_mapped(). At the start of
1576          * migration, oldpage was mapped. But now, it's zapped.
1577          * But we know *target* page is not freed/reused under us.
1578          * mem_cgroup_uncharge_page() does all necessary checks.
1579          */
1580         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1581                 mem_cgroup_uncharge_page(target);
1582 }
1583
1584 /*
1585  * A call to try to shrink memory usage under specified resource controller.
1586  * This is typically used for page reclaiming for shmem for reducing side
1587  * effect of page allocation from shmem, which is used by some mem_cgroup.
1588  */
1589 int mem_cgroup_shrink_usage(struct page *page,
1590                             struct mm_struct *mm,
1591                             gfp_t gfp_mask)
1592 {
1593         struct mem_cgroup *mem = NULL;
1594         int progress = 0;
1595         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1596
1597         if (mem_cgroup_disabled())
1598                 return 0;
1599         if (page)
1600                 mem = try_get_mem_cgroup_from_swapcache(page);
1601         if (!mem && mm)
1602                 mem = try_get_mem_cgroup_from_mm(mm);
1603         if (unlikely(!mem))
1604                 return 0;
1605
1606         do {
1607                 progress = mem_cgroup_hierarchical_reclaim(mem,
1608                                         gfp_mask, true, false);
1609                 progress += mem_cgroup_check_under_limit(mem);
1610         } while (!progress && --retry);
1611
1612         css_put(&mem->css);
1613         if (!retry)
1614                 return -ENOMEM;
1615         return 0;
1616 }
1617
1618 static DEFINE_MUTEX(set_limit_mutex);
1619
1620 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1621                                 unsigned long long val)
1622 {
1623         int retry_count;
1624         int progress;
1625         u64 memswlimit;
1626         int ret = 0;
1627         int children = mem_cgroup_count_children(memcg);
1628         u64 curusage, oldusage;
1629
1630         /*
1631          * For keeping hierarchical_reclaim simple, how long we should retry
1632          * is depends on callers. We set our retry-count to be function
1633          * of # of children which we should visit in this loop.
1634          */
1635         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
1636
1637         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1638
1639         while (retry_count) {
1640                 if (signal_pending(current)) {
1641                         ret = -EINTR;
1642                         break;
1643                 }
1644                 /*
1645                  * Rather than hide all in some function, I do this in
1646                  * open coded manner. You see what this really does.
1647                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1648                  */
1649                 mutex_lock(&set_limit_mutex);
1650                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1651                 if (memswlimit < val) {
1652                         ret = -EINVAL;
1653                         mutex_unlock(&set_limit_mutex);
1654                         break;
1655                 }
1656                 ret = res_counter_set_limit(&memcg->res, val);
1657                 mutex_unlock(&set_limit_mutex);
1658
1659                 if (!ret)
1660                         break;
1661
1662                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1663                                                    false, true);
1664                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1665                 /* Usage is reduced ? */
1666                 if (curusage >= oldusage)
1667                         retry_count--;
1668                 else
1669                         oldusage = curusage;
1670         }
1671
1672         return ret;
1673 }
1674
1675 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1676                                 unsigned long long val)
1677 {
1678         int retry_count;
1679         u64 memlimit, oldusage, curusage;
1680         int children = mem_cgroup_count_children(memcg);
1681         int ret = -EBUSY;
1682
1683         if (!do_swap_account)
1684                 return -EINVAL;
1685         /* see mem_cgroup_resize_res_limit */
1686         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
1687         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1688         while (retry_count) {
1689                 if (signal_pending(current)) {
1690                         ret = -EINTR;
1691                         break;
1692                 }
1693                 /*
1694                  * Rather than hide all in some function, I do this in
1695                  * open coded manner. You see what this really does.
1696                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1697                  */
1698                 mutex_lock(&set_limit_mutex);
1699                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1700                 if (memlimit > val) {
1701                         ret = -EINVAL;
1702                         mutex_unlock(&set_limit_mutex);
1703                         break;
1704                 }
1705                 ret = res_counter_set_limit(&memcg->memsw, val);
1706                 mutex_unlock(&set_limit_mutex);
1707
1708                 if (!ret)
1709                         break;
1710
1711                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
1712                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1713                 /* Usage is reduced ? */
1714                 if (curusage >= oldusage)
1715                         retry_count--;
1716                 else
1717                         oldusage = curusage;
1718         }
1719         return ret;
1720 }
1721
1722 /*
1723  * This routine traverse page_cgroup in given list and drop them all.
1724  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1725  */
1726 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1727                                 int node, int zid, enum lru_list lru)
1728 {
1729         struct zone *zone;
1730         struct mem_cgroup_per_zone *mz;
1731         struct page_cgroup *pc, *busy;
1732         unsigned long flags, loop;
1733         struct list_head *list;
1734         int ret = 0;
1735
1736         zone = &NODE_DATA(node)->node_zones[zid];
1737         mz = mem_cgroup_zoneinfo(mem, node, zid);
1738         list = &mz->lists[lru];
1739
1740         loop = MEM_CGROUP_ZSTAT(mz, lru);
1741         /* give some margin against EBUSY etc...*/
1742         loop += 256;
1743         busy = NULL;
1744         while (loop--) {
1745                 ret = 0;
1746                 spin_lock_irqsave(&zone->lru_lock, flags);
1747                 if (list_empty(list)) {
1748                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1749                         break;
1750                 }
1751                 pc = list_entry(list->prev, struct page_cgroup, lru);
1752                 if (busy == pc) {
1753                         list_move(&pc->lru, list);
1754                         busy = 0;
1755                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1756                         continue;
1757                 }
1758                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1759
1760                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1761                 if (ret == -ENOMEM)
1762                         break;
1763
1764                 if (ret == -EBUSY || ret == -EINVAL) {
1765                         /* found lock contention or "pc" is obsolete. */
1766                         busy = pc;
1767                         cond_resched();
1768                 } else
1769                         busy = NULL;
1770         }
1771
1772         if (!ret && !list_empty(list))
1773                 return -EBUSY;
1774         return ret;
1775 }
1776
1777 /*
1778  * make mem_cgroup's charge to be 0 if there is no task.
1779  * This enables deleting this mem_cgroup.
1780  */
1781 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1782 {
1783         int ret;
1784         int node, zid, shrink;
1785         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1786         struct cgroup *cgrp = mem->css.cgroup;
1787
1788         css_get(&mem->css);
1789
1790         shrink = 0;
1791         /* should free all ? */
1792         if (free_all)
1793                 goto try_to_free;
1794 move_account:
1795         while (mem->res.usage > 0) {
1796                 ret = -EBUSY;
1797                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1798                         goto out;
1799                 ret = -EINTR;
1800                 if (signal_pending(current))
1801                         goto out;
1802                 /* This is for making all *used* pages to be on LRU. */
1803                 lru_add_drain_all();
1804                 ret = 0;
1805                 for_each_node_state(node, N_HIGH_MEMORY) {
1806                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1807                                 enum lru_list l;
1808                                 for_each_lru(l) {
1809                                         ret = mem_cgroup_force_empty_list(mem,
1810                                                         node, zid, l);
1811                                         if (ret)
1812                                                 break;
1813                                 }
1814                         }
1815                         if (ret)
1816                                 break;
1817                 }
1818                 /* it seems parent cgroup doesn't have enough mem */
1819                 if (ret == -ENOMEM)
1820                         goto try_to_free;
1821                 cond_resched();
1822         }
1823         ret = 0;
1824 out:
1825         css_put(&mem->css);
1826         return ret;
1827
1828 try_to_free:
1829         /* returns EBUSY if there is a task or if we come here twice. */
1830         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1831                 ret = -EBUSY;
1832                 goto out;
1833         }
1834         /* we call try-to-free pages for make this cgroup empty */
1835         lru_add_drain_all();
1836         /* try to free all pages in this cgroup */
1837         shrink = 1;
1838         while (nr_retries && mem->res.usage > 0) {
1839                 int progress;
1840
1841                 if (signal_pending(current)) {
1842                         ret = -EINTR;
1843                         goto out;
1844                 }
1845                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1846                                                 false, get_swappiness(mem));
1847                 if (!progress) {
1848                         nr_retries--;
1849                         /* maybe some writeback is necessary */
1850                         congestion_wait(WRITE, HZ/10);
1851                 }
1852
1853         }
1854         lru_add_drain();
1855         /* try move_account...there may be some *locked* pages. */
1856         if (mem->res.usage)
1857                 goto move_account;
1858         ret = 0;
1859         goto out;
1860 }
1861
1862 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1863 {
1864         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1865 }
1866
1867
1868 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1869 {
1870         return mem_cgroup_from_cont(cont)->use_hierarchy;
1871 }
1872
1873 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1874                                         u64 val)
1875 {
1876         int retval = 0;
1877         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1878         struct cgroup *parent = cont->parent;
1879         struct mem_cgroup *parent_mem = NULL;
1880
1881         if (parent)
1882                 parent_mem = mem_cgroup_from_cont(parent);
1883
1884         cgroup_lock();
1885         /*
1886          * If parent's use_hiearchy is set, we can't make any modifications
1887          * in the child subtrees. If it is unset, then the change can
1888          * occur, provided the current cgroup has no children.
1889          *
1890          * For the root cgroup, parent_mem is NULL, we allow value to be
1891          * set if there are no children.
1892          */
1893         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1894                                 (val == 1 || val == 0)) {
1895                 if (list_empty(&cont->children))
1896                         mem->use_hierarchy = val;
1897                 else
1898                         retval = -EBUSY;
1899         } else
1900                 retval = -EINVAL;
1901         cgroup_unlock();
1902
1903         return retval;
1904 }
1905
1906 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1907 {
1908         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1909         u64 val = 0;
1910         int type, name;
1911
1912         type = MEMFILE_TYPE(cft->private);
1913         name = MEMFILE_ATTR(cft->private);
1914         switch (type) {
1915         case _MEM:
1916                 val = res_counter_read_u64(&mem->res, name);
1917                 break;
1918         case _MEMSWAP:
1919                 if (do_swap_account)
1920                         val = res_counter_read_u64(&mem->memsw, name);
1921                 break;
1922         default:
1923                 BUG();
1924                 break;
1925         }
1926         return val;
1927 }
1928 /*
1929  * The user of this function is...
1930  * RES_LIMIT.
1931  */
1932 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1933                             const char *buffer)
1934 {
1935         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1936         int type, name;
1937         unsigned long long val;
1938         int ret;
1939
1940         type = MEMFILE_TYPE(cft->private);
1941         name = MEMFILE_ATTR(cft->private);
1942         switch (name) {
1943         case RES_LIMIT:
1944                 /* This function does all necessary parse...reuse it */
1945                 ret = res_counter_memparse_write_strategy(buffer, &val);
1946                 if (ret)
1947                         break;
1948                 if (type == _MEM)
1949                         ret = mem_cgroup_resize_limit(memcg, val);
1950                 else
1951                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1952                 break;
1953         default:
1954                 ret = -EINVAL; /* should be BUG() ? */
1955                 break;
1956         }
1957         return ret;
1958 }
1959
1960 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1961                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1962 {
1963         struct cgroup *cgroup;
1964         unsigned long long min_limit, min_memsw_limit, tmp;
1965
1966         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1967         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1968         cgroup = memcg->css.cgroup;
1969         if (!memcg->use_hierarchy)
1970                 goto out;
1971
1972         while (cgroup->parent) {
1973                 cgroup = cgroup->parent;
1974                 memcg = mem_cgroup_from_cont(cgroup);
1975                 if (!memcg->use_hierarchy)
1976                         break;
1977                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
1978                 min_limit = min(min_limit, tmp);
1979                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1980                 min_memsw_limit = min(min_memsw_limit, tmp);
1981         }
1982 out:
1983         *mem_limit = min_limit;
1984         *memsw_limit = min_memsw_limit;
1985         return;
1986 }
1987
1988 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1989 {
1990         struct mem_cgroup *mem;
1991         int type, name;
1992
1993         mem = mem_cgroup_from_cont(cont);
1994         type = MEMFILE_TYPE(event);
1995         name = MEMFILE_ATTR(event);
1996         switch (name) {
1997         case RES_MAX_USAGE:
1998                 if (type == _MEM)
1999                         res_counter_reset_max(&mem->res);
2000                 else
2001                         res_counter_reset_max(&mem->memsw);
2002                 break;
2003         case RES_FAILCNT:
2004                 if (type == _MEM)
2005                         res_counter_reset_failcnt(&mem->res);
2006                 else
2007                         res_counter_reset_failcnt(&mem->memsw);
2008                 break;
2009         }
2010         return 0;
2011 }
2012
2013
2014 /* For read statistics */
2015 enum {
2016         MCS_CACHE,
2017         MCS_RSS,
2018         MCS_PGPGIN,
2019         MCS_PGPGOUT,
2020         MCS_INACTIVE_ANON,
2021         MCS_ACTIVE_ANON,
2022         MCS_INACTIVE_FILE,
2023         MCS_ACTIVE_FILE,
2024         MCS_UNEVICTABLE,
2025         NR_MCS_STAT,
2026 };
2027
2028 struct mcs_total_stat {
2029         s64 stat[NR_MCS_STAT];
2030 };
2031
2032 struct {
2033         char *local_name;
2034         char *total_name;
2035 } memcg_stat_strings[NR_MCS_STAT] = {
2036         {"cache", "total_cache"},
2037         {"rss", "total_rss"},
2038         {"pgpgin", "total_pgpgin"},
2039         {"pgpgout", "total_pgpgout"},
2040         {"inactive_anon", "total_inactive_anon"},
2041         {"active_anon", "total_active_anon"},
2042         {"inactive_file", "total_inactive_file"},
2043         {"active_file", "total_active_file"},
2044         {"unevictable", "total_unevictable"}
2045 };
2046
2047
2048 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2049 {
2050         struct mcs_total_stat *s = data;
2051         s64 val;
2052
2053         /* per cpu stat */
2054         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2055         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2056         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2057         s->stat[MCS_RSS] += val * PAGE_SIZE;
2058         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2059         s->stat[MCS_PGPGIN] += val;
2060         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2061         s->stat[MCS_PGPGOUT] += val;
2062
2063         /* per zone stat */
2064         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2065         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2066         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2067         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2068         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2069         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2070         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2071         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2072         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2073         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2074         return 0;
2075 }
2076
2077 static void
2078 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2079 {
2080         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2081 }
2082
2083 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2084                                  struct cgroup_map_cb *cb)
2085 {
2086         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2087         struct mcs_total_stat mystat;
2088         int i;
2089
2090         memset(&mystat, 0, sizeof(mystat));
2091         mem_cgroup_get_local_stat(mem_cont, &mystat);
2092
2093         for (i = 0; i < NR_MCS_STAT; i++)
2094                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2095
2096         /* Hierarchical information */
2097         {
2098                 unsigned long long limit, memsw_limit;
2099                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2100                 cb->fill(cb, "hierarchical_memory_limit", limit);
2101                 if (do_swap_account)
2102                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2103         }
2104
2105         memset(&mystat, 0, sizeof(mystat));
2106         mem_cgroup_get_total_stat(mem_cont, &mystat);
2107         for (i = 0; i < NR_MCS_STAT; i++)
2108                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2109
2110
2111 #ifdef CONFIG_DEBUG_VM
2112         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2113
2114         {
2115                 int nid, zid;
2116                 struct mem_cgroup_per_zone *mz;
2117                 unsigned long recent_rotated[2] = {0, 0};
2118                 unsigned long recent_scanned[2] = {0, 0};
2119
2120                 for_each_online_node(nid)
2121                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2122                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2123
2124                                 recent_rotated[0] +=
2125                                         mz->reclaim_stat.recent_rotated[0];
2126                                 recent_rotated[1] +=
2127                                         mz->reclaim_stat.recent_rotated[1];
2128                                 recent_scanned[0] +=
2129                                         mz->reclaim_stat.recent_scanned[0];
2130                                 recent_scanned[1] +=
2131                                         mz->reclaim_stat.recent_scanned[1];
2132                         }
2133                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2134                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2135                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2136                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2137         }
2138 #endif
2139
2140         return 0;
2141 }
2142
2143 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2144 {
2145         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2146
2147         return get_swappiness(memcg);
2148 }
2149
2150 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2151                                        u64 val)
2152 {
2153         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2154         struct mem_cgroup *parent;
2155
2156         if (val > 100)
2157                 return -EINVAL;
2158
2159         if (cgrp->parent == NULL)
2160                 return -EINVAL;
2161
2162         parent = mem_cgroup_from_cont(cgrp->parent);
2163
2164         cgroup_lock();
2165
2166         /* If under hierarchy, only empty-root can set this value */
2167         if ((parent->use_hierarchy) ||
2168             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2169                 cgroup_unlock();
2170                 return -EINVAL;
2171         }
2172
2173         spin_lock(&memcg->reclaim_param_lock);
2174         memcg->swappiness = val;
2175         spin_unlock(&memcg->reclaim_param_lock);
2176
2177         cgroup_unlock();
2178
2179         return 0;
2180 }
2181
2182
2183 static struct cftype mem_cgroup_files[] = {
2184         {
2185                 .name = "usage_in_bytes",
2186                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2187                 .read_u64 = mem_cgroup_read,
2188         },
2189         {
2190                 .name = "max_usage_in_bytes",
2191                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2192                 .trigger = mem_cgroup_reset,
2193                 .read_u64 = mem_cgroup_read,
2194         },
2195         {
2196                 .name = "limit_in_bytes",
2197                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2198                 .write_string = mem_cgroup_write,
2199                 .read_u64 = mem_cgroup_read,
2200         },
2201         {
2202                 .name = "failcnt",
2203                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2204                 .trigger = mem_cgroup_reset,
2205                 .read_u64 = mem_cgroup_read,
2206         },
2207         {
2208                 .name = "stat",
2209                 .read_map = mem_control_stat_show,
2210         },
2211         {
2212                 .name = "force_empty",
2213                 .trigger = mem_cgroup_force_empty_write,
2214         },
2215         {
2216                 .name = "use_hierarchy",
2217                 .write_u64 = mem_cgroup_hierarchy_write,
2218                 .read_u64 = mem_cgroup_hierarchy_read,
2219         },
2220         {
2221                 .name = "swappiness",
2222                 .read_u64 = mem_cgroup_swappiness_read,
2223                 .write_u64 = mem_cgroup_swappiness_write,
2224         },
2225 };
2226
2227 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2228 static struct cftype memsw_cgroup_files[] = {
2229         {
2230                 .name = "memsw.usage_in_bytes",
2231                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2232                 .read_u64 = mem_cgroup_read,
2233         },
2234         {
2235                 .name = "memsw.max_usage_in_bytes",
2236                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2237                 .trigger = mem_cgroup_reset,
2238                 .read_u64 = mem_cgroup_read,
2239         },
2240         {
2241                 .name = "memsw.limit_in_bytes",
2242                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2243                 .write_string = mem_cgroup_write,
2244                 .read_u64 = mem_cgroup_read,
2245         },
2246         {
2247                 .name = "memsw.failcnt",
2248                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2249                 .trigger = mem_cgroup_reset,
2250                 .read_u64 = mem_cgroup_read,
2251         },
2252 };
2253
2254 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2255 {
2256         if (!do_swap_account)
2257                 return 0;
2258         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2259                                 ARRAY_SIZE(memsw_cgroup_files));
2260 };
2261 #else
2262 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2263 {
2264         return 0;
2265 }
2266 #endif
2267
2268 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2269 {
2270         struct mem_cgroup_per_node *pn;
2271         struct mem_cgroup_per_zone *mz;
2272         enum lru_list l;
2273         int zone, tmp = node;
2274         /*
2275          * This routine is called against possible nodes.
2276          * But it's BUG to call kmalloc() against offline node.
2277          *
2278          * TODO: this routine can waste much memory for nodes which will
2279          *       never be onlined. It's better to use memory hotplug callback
2280          *       function.
2281          */
2282         if (!node_state(node, N_NORMAL_MEMORY))
2283                 tmp = -1;
2284         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2285         if (!pn)
2286                 return 1;
2287
2288         mem->info.nodeinfo[node] = pn;
2289         memset(pn, 0, sizeof(*pn));
2290
2291         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2292                 mz = &pn->zoneinfo[zone];
2293                 for_each_lru(l)
2294                         INIT_LIST_HEAD(&mz->lists[l]);
2295         }
2296         return 0;
2297 }
2298
2299 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2300 {
2301         kfree(mem->info.nodeinfo[node]);
2302 }
2303
2304 static int mem_cgroup_size(void)
2305 {
2306         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2307         return sizeof(struct mem_cgroup) + cpustat_size;
2308 }
2309
2310 static struct mem_cgroup *mem_cgroup_alloc(void)
2311 {
2312         struct mem_cgroup *mem;
2313         int size = mem_cgroup_size();
2314
2315         if (size < PAGE_SIZE)
2316                 mem = kmalloc(size, GFP_KERNEL);
2317         else
2318                 mem = vmalloc(size);
2319
2320         if (mem)
2321                 memset(mem, 0, size);
2322         return mem;
2323 }
2324
2325 /*
2326  * At destroying mem_cgroup, references from swap_cgroup can remain.
2327  * (scanning all at force_empty is too costly...)
2328  *
2329  * Instead of clearing all references at force_empty, we remember
2330  * the number of reference from swap_cgroup and free mem_cgroup when
2331  * it goes down to 0.
2332  *
2333  * Removal of cgroup itself succeeds regardless of refs from swap.
2334  */
2335
2336 static void __mem_cgroup_free(struct mem_cgroup *mem)
2337 {
2338         int node;
2339
2340         free_css_id(&mem_cgroup_subsys, &mem->css);
2341
2342         for_each_node_state(node, N_POSSIBLE)
2343                 free_mem_cgroup_per_zone_info(mem, node);
2344
2345         if (mem_cgroup_size() < PAGE_SIZE)
2346                 kfree(mem);
2347         else
2348                 vfree(mem);
2349 }
2350
2351 static void mem_cgroup_get(struct mem_cgroup *mem)
2352 {
2353         atomic_inc(&mem->refcnt);
2354 }
2355
2356 static void mem_cgroup_put(struct mem_cgroup *mem)
2357 {
2358         if (atomic_dec_and_test(&mem->refcnt)) {
2359                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2360                 __mem_cgroup_free(mem);
2361                 if (parent)
2362                         mem_cgroup_put(parent);
2363         }
2364 }
2365
2366 /*
2367  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2368  */
2369 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2370 {
2371         if (!mem->res.parent)
2372                 return NULL;
2373         return mem_cgroup_from_res_counter(mem->res.parent, res);
2374 }
2375
2376 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2377 static void __init enable_swap_cgroup(void)
2378 {
2379         if (!mem_cgroup_disabled() && really_do_swap_account)
2380                 do_swap_account = 1;
2381 }
2382 #else
2383 static void __init enable_swap_cgroup(void)
2384 {
2385 }
2386 #endif
2387
2388 static struct cgroup_subsys_state * __ref
2389 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2390 {
2391         struct mem_cgroup *mem, *parent;
2392         long error = -ENOMEM;
2393         int node;
2394
2395         mem = mem_cgroup_alloc();
2396         if (!mem)
2397                 return ERR_PTR(error);
2398
2399         for_each_node_state(node, N_POSSIBLE)
2400                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2401                         goto free_out;
2402         /* root ? */
2403         if (cont->parent == NULL) {
2404                 enable_swap_cgroup();
2405                 parent = NULL;
2406         } else {
2407                 parent = mem_cgroup_from_cont(cont->parent);
2408                 mem->use_hierarchy = parent->use_hierarchy;
2409         }
2410
2411         if (parent && parent->use_hierarchy) {
2412                 res_counter_init(&mem->res, &parent->res);
2413                 res_counter_init(&mem->memsw, &parent->memsw);
2414                 /*
2415                  * We increment refcnt of the parent to ensure that we can
2416                  * safely access it on res_counter_charge/uncharge.
2417                  * This refcnt will be decremented when freeing this
2418                  * mem_cgroup(see mem_cgroup_put).
2419                  */
2420                 mem_cgroup_get(parent);
2421         } else {
2422                 res_counter_init(&mem->res, NULL);
2423                 res_counter_init(&mem->memsw, NULL);
2424         }
2425         mem->last_scanned_child = 0;
2426         spin_lock_init(&mem->reclaim_param_lock);
2427
2428         if (parent)
2429                 mem->swappiness = get_swappiness(parent);
2430         atomic_set(&mem->refcnt, 1);
2431         return &mem->css;
2432 free_out:
2433         __mem_cgroup_free(mem);
2434         return ERR_PTR(error);
2435 }
2436
2437 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2438                                         struct cgroup *cont)
2439 {
2440         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2441
2442         return mem_cgroup_force_empty(mem, false);
2443 }
2444
2445 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2446                                 struct cgroup *cont)
2447 {
2448         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2449
2450         mem_cgroup_put(mem);
2451 }
2452
2453 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2454                                 struct cgroup *cont)
2455 {
2456         int ret;
2457
2458         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2459                                 ARRAY_SIZE(mem_cgroup_files));
2460
2461         if (!ret)
2462                 ret = register_memsw_files(cont, ss);
2463         return ret;
2464 }
2465
2466 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2467                                 struct cgroup *cont,
2468                                 struct cgroup *old_cont,
2469                                 struct task_struct *p)
2470 {
2471         mutex_lock(&memcg_tasklist);
2472         /*
2473          * FIXME: It's better to move charges of this process from old
2474          * memcg to new memcg. But it's just on TODO-List now.
2475          */
2476         mutex_unlock(&memcg_tasklist);
2477 }
2478
2479 struct cgroup_subsys mem_cgroup_subsys = {
2480         .name = "memory",
2481         .subsys_id = mem_cgroup_subsys_id,
2482         .create = mem_cgroup_create,
2483         .pre_destroy = mem_cgroup_pre_destroy,
2484         .destroy = mem_cgroup_destroy,
2485         .populate = mem_cgroup_populate,
2486         .attach = mem_cgroup_move_task,
2487         .early_init = 0,
2488         .use_id = 1,
2489 };
2490
2491 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2492
2493 static int __init disable_swap_account(char *s)
2494 {
2495         really_do_swap_account = 0;
2496         return 1;
2497 }
2498 __setup("noswapaccount", disable_swap_account);
2499 #endif