Fix congestion_wait() sync/async vs read/write confusion
[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 = 1 */
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 anon rss */
66         MEM_CGROUP_STAT_MAPPED_FILE,  /* # of pages charged as file rss */
67         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
68         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
69
70         MEM_CGROUP_STAT_NSTATS,
71 };
72
73 struct mem_cgroup_stat_cpu {
74         s64 count[MEM_CGROUP_STAT_NSTATS];
75 } ____cacheline_aligned_in_smp;
76
77 struct mem_cgroup_stat {
78         struct mem_cgroup_stat_cpu cpustat[0];
79 };
80
81 /*
82  * For accounting under irq disable, no need for increment preempt count.
83  */
84 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
85                 enum mem_cgroup_stat_index idx, int val)
86 {
87         stat->count[idx] += val;
88 }
89
90 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
91                 enum mem_cgroup_stat_index idx)
92 {
93         int cpu;
94         s64 ret = 0;
95         for_each_possible_cpu(cpu)
96                 ret += stat->cpustat[cpu].count[idx];
97         return ret;
98 }
99
100 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
101 {
102         s64 ret;
103
104         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
105         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
106         return ret;
107 }
108
109 /*
110  * per-zone information in memory controller.
111  */
112 struct mem_cgroup_per_zone {
113         /*
114          * spin_lock to protect the per cgroup LRU
115          */
116         struct list_head        lists[NR_LRU_LISTS];
117         unsigned long           count[NR_LRU_LISTS];
118
119         struct zone_reclaim_stat reclaim_stat;
120 };
121 /* Macro for accessing counter */
122 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
123
124 struct mem_cgroup_per_node {
125         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
126 };
127
128 struct mem_cgroup_lru_info {
129         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
130 };
131
132 /*
133  * The memory controller data structure. The memory controller controls both
134  * page cache and RSS per cgroup. We would eventually like to provide
135  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
136  * to help the administrator determine what knobs to tune.
137  *
138  * TODO: Add a water mark for the memory controller. Reclaim will begin when
139  * we hit the water mark. May be even add a low water mark, such that
140  * no reclaim occurs from a cgroup at it's low water mark, this is
141  * a feature that will be implemented much later in the future.
142  */
143 struct mem_cgroup {
144         struct cgroup_subsys_state css;
145         /*
146          * the counter to account for memory usage
147          */
148         struct res_counter res;
149         /*
150          * the counter to account for mem+swap usage.
151          */
152         struct res_counter memsw;
153         /*
154          * Per cgroup active and inactive list, similar to the
155          * per zone LRU lists.
156          */
157         struct mem_cgroup_lru_info info;
158
159         /*
160           protect against reclaim related member.
161         */
162         spinlock_t reclaim_param_lock;
163
164         int     prev_priority;  /* for recording reclaim priority */
165
166         /*
167          * While reclaiming in a hiearchy, we cache the last child we
168          * reclaimed from.
169          */
170         int last_scanned_child;
171         /*
172          * Should the accounting and control be hierarchical, per subtree?
173          */
174         bool use_hierarchy;
175         unsigned long   last_oom_jiffies;
176         atomic_t        refcnt;
177
178         unsigned int    swappiness;
179
180         /* set when res.limit == memsw.limit */
181         bool            memsw_is_minimum;
182
183         /*
184          * statistics. This must be placed at the end of memcg.
185          */
186         struct mem_cgroup_stat stat;
187 };
188
189 enum charge_type {
190         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
191         MEM_CGROUP_CHARGE_TYPE_MAPPED,
192         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
193         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
194         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
195         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
196         NR_CHARGE_TYPE,
197 };
198
199 /* only for here (for easy reading.) */
200 #define PCGF_CACHE      (1UL << PCG_CACHE)
201 #define PCGF_USED       (1UL << PCG_USED)
202 #define PCGF_LOCK       (1UL << PCG_LOCK)
203 static const unsigned long
204 pcg_default_flags[NR_CHARGE_TYPE] = {
205         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
206         PCGF_USED | PCGF_LOCK, /* Anon */
207         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
208         0, /* FORCE */
209 };
210
211 /* for encoding cft->private value on file */
212 #define _MEM                    (0)
213 #define _MEMSWAP                (1)
214 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
215 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
216 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
217
218 static void mem_cgroup_get(struct mem_cgroup *mem);
219 static void mem_cgroup_put(struct mem_cgroup *mem);
220 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
221
222 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
223                                          struct page_cgroup *pc,
224                                          bool charge)
225 {
226         int val = (charge)? 1 : -1;
227         struct mem_cgroup_stat *stat = &mem->stat;
228         struct mem_cgroup_stat_cpu *cpustat;
229         int cpu = get_cpu();
230
231         cpustat = &stat->cpustat[cpu];
232         if (PageCgroupCache(pc))
233                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
234         else
235                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
236
237         if (charge)
238                 __mem_cgroup_stat_add_safe(cpustat,
239                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
240         else
241                 __mem_cgroup_stat_add_safe(cpustat,
242                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
243         put_cpu();
244 }
245
246 static struct mem_cgroup_per_zone *
247 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
248 {
249         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
250 }
251
252 static struct mem_cgroup_per_zone *
253 page_cgroup_zoneinfo(struct page_cgroup *pc)
254 {
255         struct mem_cgroup *mem = pc->mem_cgroup;
256         int nid = page_cgroup_nid(pc);
257         int zid = page_cgroup_zid(pc);
258
259         if (!mem)
260                 return NULL;
261
262         return mem_cgroup_zoneinfo(mem, nid, zid);
263 }
264
265 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
266                                         enum lru_list idx)
267 {
268         int nid, zid;
269         struct mem_cgroup_per_zone *mz;
270         u64 total = 0;
271
272         for_each_online_node(nid)
273                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
274                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
275                         total += MEM_CGROUP_ZSTAT(mz, idx);
276                 }
277         return total;
278 }
279
280 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
281 {
282         return container_of(cgroup_subsys_state(cont,
283                                 mem_cgroup_subsys_id), struct mem_cgroup,
284                                 css);
285 }
286
287 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
288 {
289         /*
290          * mm_update_next_owner() may clear mm->owner to NULL
291          * if it races with swapoff, page migration, etc.
292          * So this can be called with p == NULL.
293          */
294         if (unlikely(!p))
295                 return NULL;
296
297         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
298                                 struct mem_cgroup, css);
299 }
300
301 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
302 {
303         struct mem_cgroup *mem = NULL;
304
305         if (!mm)
306                 return NULL;
307         /*
308          * Because we have no locks, mm->owner's may be being moved to other
309          * cgroup. We use css_tryget() here even if this looks
310          * pessimistic (rather than adding locks here).
311          */
312         rcu_read_lock();
313         do {
314                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
315                 if (unlikely(!mem))
316                         break;
317         } while (!css_tryget(&mem->css));
318         rcu_read_unlock();
319         return mem;
320 }
321
322 /*
323  * Call callback function against all cgroup under hierarchy tree.
324  */
325 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
326                           int (*func)(struct mem_cgroup *, void *))
327 {
328         int found, ret, nextid;
329         struct cgroup_subsys_state *css;
330         struct mem_cgroup *mem;
331
332         if (!root->use_hierarchy)
333                 return (*func)(root, data);
334
335         nextid = 1;
336         do {
337                 ret = 0;
338                 mem = NULL;
339
340                 rcu_read_lock();
341                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
342                                    &found);
343                 if (css && css_tryget(css))
344                         mem = container_of(css, struct mem_cgroup, css);
345                 rcu_read_unlock();
346
347                 if (mem) {
348                         ret = (*func)(mem, data);
349                         css_put(&mem->css);
350                 }
351                 nextid = found + 1;
352         } while (!ret && css);
353
354         return ret;
355 }
356
357 /*
358  * Following LRU functions are allowed to be used without PCG_LOCK.
359  * Operations are called by routine of global LRU independently from memcg.
360  * What we have to take care of here is validness of pc->mem_cgroup.
361  *
362  * Changes to pc->mem_cgroup happens when
363  * 1. charge
364  * 2. moving account
365  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
366  * It is added to LRU before charge.
367  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
368  * When moving account, the page is not on LRU. It's isolated.
369  */
370
371 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
372 {
373         struct page_cgroup *pc;
374         struct mem_cgroup *mem;
375         struct mem_cgroup_per_zone *mz;
376
377         if (mem_cgroup_disabled())
378                 return;
379         pc = lookup_page_cgroup(page);
380         /* can happen while we handle swapcache. */
381         if (list_empty(&pc->lru) || !pc->mem_cgroup)
382                 return;
383         /*
384          * We don't check PCG_USED bit. It's cleared when the "page" is finally
385          * removed from global LRU.
386          */
387         mz = page_cgroup_zoneinfo(pc);
388         mem = pc->mem_cgroup;
389         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
390         list_del_init(&pc->lru);
391         return;
392 }
393
394 void mem_cgroup_del_lru(struct page *page)
395 {
396         mem_cgroup_del_lru_list(page, page_lru(page));
397 }
398
399 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
400 {
401         struct mem_cgroup_per_zone *mz;
402         struct page_cgroup *pc;
403
404         if (mem_cgroup_disabled())
405                 return;
406
407         pc = lookup_page_cgroup(page);
408         /*
409          * Used bit is set without atomic ops but after smp_wmb().
410          * For making pc->mem_cgroup visible, insert smp_rmb() here.
411          */
412         smp_rmb();
413         /* unused page is not rotated. */
414         if (!PageCgroupUsed(pc))
415                 return;
416         mz = page_cgroup_zoneinfo(pc);
417         list_move(&pc->lru, &mz->lists[lru]);
418 }
419
420 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
421 {
422         struct page_cgroup *pc;
423         struct mem_cgroup_per_zone *mz;
424
425         if (mem_cgroup_disabled())
426                 return;
427         pc = lookup_page_cgroup(page);
428         /*
429          * Used bit is set without atomic ops but after smp_wmb().
430          * For making pc->mem_cgroup visible, insert smp_rmb() here.
431          */
432         smp_rmb();
433         if (!PageCgroupUsed(pc))
434                 return;
435
436         mz = page_cgroup_zoneinfo(pc);
437         MEM_CGROUP_ZSTAT(mz, lru) += 1;
438         list_add(&pc->lru, &mz->lists[lru]);
439 }
440
441 /*
442  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
443  * lru because the page may.be reused after it's fully uncharged (because of
444  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
445  * it again. This function is only used to charge SwapCache. It's done under
446  * lock_page and expected that zone->lru_lock is never held.
447  */
448 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
449 {
450         unsigned long flags;
451         struct zone *zone = page_zone(page);
452         struct page_cgroup *pc = lookup_page_cgroup(page);
453
454         spin_lock_irqsave(&zone->lru_lock, flags);
455         /*
456          * Forget old LRU when this page_cgroup is *not* used. This Used bit
457          * is guarded by lock_page() because the page is SwapCache.
458          */
459         if (!PageCgroupUsed(pc))
460                 mem_cgroup_del_lru_list(page, page_lru(page));
461         spin_unlock_irqrestore(&zone->lru_lock, flags);
462 }
463
464 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
465 {
466         unsigned long flags;
467         struct zone *zone = page_zone(page);
468         struct page_cgroup *pc = lookup_page_cgroup(page);
469
470         spin_lock_irqsave(&zone->lru_lock, flags);
471         /* link when the page is linked to LRU but page_cgroup isn't */
472         if (PageLRU(page) && list_empty(&pc->lru))
473                 mem_cgroup_add_lru_list(page, page_lru(page));
474         spin_unlock_irqrestore(&zone->lru_lock, flags);
475 }
476
477
478 void mem_cgroup_move_lists(struct page *page,
479                            enum lru_list from, enum lru_list to)
480 {
481         if (mem_cgroup_disabled())
482                 return;
483         mem_cgroup_del_lru_list(page, from);
484         mem_cgroup_add_lru_list(page, to);
485 }
486
487 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
488 {
489         int ret;
490         struct mem_cgroup *curr = NULL;
491
492         task_lock(task);
493         rcu_read_lock();
494         curr = try_get_mem_cgroup_from_mm(task->mm);
495         rcu_read_unlock();
496         task_unlock(task);
497         if (!curr)
498                 return 0;
499         if (curr->use_hierarchy)
500                 ret = css_is_ancestor(&curr->css, &mem->css);
501         else
502                 ret = (curr == mem);
503         css_put(&curr->css);
504         return ret;
505 }
506
507 /*
508  * prev_priority control...this will be used in memory reclaim path.
509  */
510 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
511 {
512         int prev_priority;
513
514         spin_lock(&mem->reclaim_param_lock);
515         prev_priority = mem->prev_priority;
516         spin_unlock(&mem->reclaim_param_lock);
517
518         return prev_priority;
519 }
520
521 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
522 {
523         spin_lock(&mem->reclaim_param_lock);
524         if (priority < mem->prev_priority)
525                 mem->prev_priority = priority;
526         spin_unlock(&mem->reclaim_param_lock);
527 }
528
529 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
530 {
531         spin_lock(&mem->reclaim_param_lock);
532         mem->prev_priority = priority;
533         spin_unlock(&mem->reclaim_param_lock);
534 }
535
536 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
537 {
538         unsigned long active;
539         unsigned long inactive;
540         unsigned long gb;
541         unsigned long inactive_ratio;
542
543         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
544         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
545
546         gb = (inactive + active) >> (30 - PAGE_SHIFT);
547         if (gb)
548                 inactive_ratio = int_sqrt(10 * gb);
549         else
550                 inactive_ratio = 1;
551
552         if (present_pages) {
553                 present_pages[0] = inactive;
554                 present_pages[1] = active;
555         }
556
557         return inactive_ratio;
558 }
559
560 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
561 {
562         unsigned long active;
563         unsigned long inactive;
564         unsigned long present_pages[2];
565         unsigned long inactive_ratio;
566
567         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
568
569         inactive = present_pages[0];
570         active = present_pages[1];
571
572         if (inactive * inactive_ratio < active)
573                 return 1;
574
575         return 0;
576 }
577
578 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
579 {
580         unsigned long active;
581         unsigned long inactive;
582
583         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
584         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
585
586         return (active > inactive);
587 }
588
589 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
590                                        struct zone *zone,
591                                        enum lru_list lru)
592 {
593         int nid = zone->zone_pgdat->node_id;
594         int zid = zone_idx(zone);
595         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
596
597         return MEM_CGROUP_ZSTAT(mz, lru);
598 }
599
600 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
601                                                       struct zone *zone)
602 {
603         int nid = zone->zone_pgdat->node_id;
604         int zid = zone_idx(zone);
605         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
606
607         return &mz->reclaim_stat;
608 }
609
610 struct zone_reclaim_stat *
611 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
612 {
613         struct page_cgroup *pc;
614         struct mem_cgroup_per_zone *mz;
615
616         if (mem_cgroup_disabled())
617                 return NULL;
618
619         pc = lookup_page_cgroup(page);
620         /*
621          * Used bit is set without atomic ops but after smp_wmb().
622          * For making pc->mem_cgroup visible, insert smp_rmb() here.
623          */
624         smp_rmb();
625         if (!PageCgroupUsed(pc))
626                 return NULL;
627
628         mz = page_cgroup_zoneinfo(pc);
629         if (!mz)
630                 return NULL;
631
632         return &mz->reclaim_stat;
633 }
634
635 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
636                                         struct list_head *dst,
637                                         unsigned long *scanned, int order,
638                                         int mode, struct zone *z,
639                                         struct mem_cgroup *mem_cont,
640                                         int active, int file)
641 {
642         unsigned long nr_taken = 0;
643         struct page *page;
644         unsigned long scan;
645         LIST_HEAD(pc_list);
646         struct list_head *src;
647         struct page_cgroup *pc, *tmp;
648         int nid = z->zone_pgdat->node_id;
649         int zid = zone_idx(z);
650         struct mem_cgroup_per_zone *mz;
651         int lru = LRU_FILE * !!file + !!active;
652         int ret;
653
654         BUG_ON(!mem_cont);
655         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
656         src = &mz->lists[lru];
657
658         scan = 0;
659         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
660                 if (scan >= nr_to_scan)
661                         break;
662
663                 page = pc->page;
664                 if (unlikely(!PageCgroupUsed(pc)))
665                         continue;
666                 if (unlikely(!PageLRU(page)))
667                         continue;
668
669                 scan++;
670                 ret = __isolate_lru_page(page, mode, file);
671                 switch (ret) {
672                 case 0:
673                         list_move(&page->lru, dst);
674                         mem_cgroup_del_lru(page);
675                         nr_taken++;
676                         break;
677                 case -EBUSY:
678                         /* we don't affect global LRU but rotate in our LRU */
679                         mem_cgroup_rotate_lru_list(page, page_lru(page));
680                         break;
681                 default:
682                         break;
683                 }
684         }
685
686         *scanned = scan;
687         return nr_taken;
688 }
689
690 #define mem_cgroup_from_res_counter(counter, member)    \
691         container_of(counter, struct mem_cgroup, member)
692
693 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
694 {
695         if (do_swap_account) {
696                 if (res_counter_check_under_limit(&mem->res) &&
697                         res_counter_check_under_limit(&mem->memsw))
698                         return true;
699         } else
700                 if (res_counter_check_under_limit(&mem->res))
701                         return true;
702         return false;
703 }
704
705 static unsigned int get_swappiness(struct mem_cgroup *memcg)
706 {
707         struct cgroup *cgrp = memcg->css.cgroup;
708         unsigned int swappiness;
709
710         /* root ? */
711         if (cgrp->parent == NULL)
712                 return vm_swappiness;
713
714         spin_lock(&memcg->reclaim_param_lock);
715         swappiness = memcg->swappiness;
716         spin_unlock(&memcg->reclaim_param_lock);
717
718         return swappiness;
719 }
720
721 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
722 {
723         int *val = data;
724         (*val)++;
725         return 0;
726 }
727
728 /**
729  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
730  * @memcg: The memory cgroup that went over limit
731  * @p: Task that is going to be killed
732  *
733  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
734  * enabled
735  */
736 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
737 {
738         struct cgroup *task_cgrp;
739         struct cgroup *mem_cgrp;
740         /*
741          * Need a buffer in BSS, can't rely on allocations. The code relies
742          * on the assumption that OOM is serialized for memory controller.
743          * If this assumption is broken, revisit this code.
744          */
745         static char memcg_name[PATH_MAX];
746         int ret;
747
748         if (!memcg)
749                 return;
750
751
752         rcu_read_lock();
753
754         mem_cgrp = memcg->css.cgroup;
755         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
756
757         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
758         if (ret < 0) {
759                 /*
760                  * Unfortunately, we are unable to convert to a useful name
761                  * But we'll still print out the usage information
762                  */
763                 rcu_read_unlock();
764                 goto done;
765         }
766         rcu_read_unlock();
767
768         printk(KERN_INFO "Task in %s killed", memcg_name);
769
770         rcu_read_lock();
771         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
772         if (ret < 0) {
773                 rcu_read_unlock();
774                 goto done;
775         }
776         rcu_read_unlock();
777
778         /*
779          * Continues from above, so we don't need an KERN_ level
780          */
781         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
782 done:
783
784         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
785                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
786                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
787                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
788         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
789                 "failcnt %llu\n",
790                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
791                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
792                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
793 }
794
795 /*
796  * This function returns the number of memcg under hierarchy tree. Returns
797  * 1(self count) if no children.
798  */
799 static int mem_cgroup_count_children(struct mem_cgroup *mem)
800 {
801         int num = 0;
802         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
803         return num;
804 }
805
806 /*
807  * Visit the first child (need not be the first child as per the ordering
808  * of the cgroup list, since we track last_scanned_child) of @mem and use
809  * that to reclaim free pages from.
810  */
811 static struct mem_cgroup *
812 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
813 {
814         struct mem_cgroup *ret = NULL;
815         struct cgroup_subsys_state *css;
816         int nextid, found;
817
818         if (!root_mem->use_hierarchy) {
819                 css_get(&root_mem->css);
820                 ret = root_mem;
821         }
822
823         while (!ret) {
824                 rcu_read_lock();
825                 nextid = root_mem->last_scanned_child + 1;
826                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
827                                    &found);
828                 if (css && css_tryget(css))
829                         ret = container_of(css, struct mem_cgroup, css);
830
831                 rcu_read_unlock();
832                 /* Updates scanning parameter */
833                 spin_lock(&root_mem->reclaim_param_lock);
834                 if (!css) {
835                         /* this means start scan from ID:1 */
836                         root_mem->last_scanned_child = 0;
837                 } else
838                         root_mem->last_scanned_child = found;
839                 spin_unlock(&root_mem->reclaim_param_lock);
840         }
841
842         return ret;
843 }
844
845 /*
846  * Scan the hierarchy if needed to reclaim memory. We remember the last child
847  * we reclaimed from, so that we don't end up penalizing one child extensively
848  * based on its position in the children list.
849  *
850  * root_mem is the original ancestor that we've been reclaim from.
851  *
852  * We give up and return to the caller when we visit root_mem twice.
853  * (other groups can be removed while we're walking....)
854  *
855  * If shrink==true, for avoiding to free too much, this returns immedieately.
856  */
857 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
858                                    gfp_t gfp_mask, bool noswap, bool shrink)
859 {
860         struct mem_cgroup *victim;
861         int ret, total = 0;
862         int loop = 0;
863
864         /* If memsw_is_minimum==1, swap-out is of-no-use. */
865         if (root_mem->memsw_is_minimum)
866                 noswap = true;
867
868         while (loop < 2) {
869                 victim = mem_cgroup_select_victim(root_mem);
870                 if (victim == root_mem)
871                         loop++;
872                 if (!mem_cgroup_local_usage(&victim->stat)) {
873                         /* this cgroup's local usage == 0 */
874                         css_put(&victim->css);
875                         continue;
876                 }
877                 /* we use swappiness of local cgroup */
878                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
879                                                    get_swappiness(victim));
880                 css_put(&victim->css);
881                 /*
882                  * At shrinking usage, we can't check we should stop here or
883                  * reclaim more. It's depends on callers. last_scanned_child
884                  * will work enough for keeping fairness under tree.
885                  */
886                 if (shrink)
887                         return ret;
888                 total += ret;
889                 if (mem_cgroup_check_under_limit(root_mem))
890                         return 1 + total;
891         }
892         return total;
893 }
894
895 bool mem_cgroup_oom_called(struct task_struct *task)
896 {
897         bool ret = false;
898         struct mem_cgroup *mem;
899         struct mm_struct *mm;
900
901         rcu_read_lock();
902         mm = task->mm;
903         if (!mm)
904                 mm = &init_mm;
905         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
906         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
907                 ret = true;
908         rcu_read_unlock();
909         return ret;
910 }
911
912 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
913 {
914         mem->last_oom_jiffies = jiffies;
915         return 0;
916 }
917
918 static void record_last_oom(struct mem_cgroup *mem)
919 {
920         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
921 }
922
923 /*
924  * Currently used to update mapped file statistics, but the routine can be
925  * generalized to update other statistics as well.
926  */
927 void mem_cgroup_update_mapped_file_stat(struct page *page, int val)
928 {
929         struct mem_cgroup *mem;
930         struct mem_cgroup_stat *stat;
931         struct mem_cgroup_stat_cpu *cpustat;
932         int cpu;
933         struct page_cgroup *pc;
934
935         if (!page_is_file_cache(page))
936                 return;
937
938         pc = lookup_page_cgroup(page);
939         if (unlikely(!pc))
940                 return;
941
942         lock_page_cgroup(pc);
943         mem = pc->mem_cgroup;
944         if (!mem)
945                 goto done;
946
947         if (!PageCgroupUsed(pc))
948                 goto done;
949
950         /*
951          * Preemption is already disabled, we don't need get_cpu()
952          */
953         cpu = smp_processor_id();
954         stat = &mem->stat;
955         cpustat = &stat->cpustat[cpu];
956
957         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val);
958 done:
959         unlock_page_cgroup(pc);
960 }
961
962 /*
963  * Unlike exported interface, "oom" parameter is added. if oom==true,
964  * oom-killer can be invoked.
965  */
966 static int __mem_cgroup_try_charge(struct mm_struct *mm,
967                         gfp_t gfp_mask, struct mem_cgroup **memcg,
968                         bool oom)
969 {
970         struct mem_cgroup *mem, *mem_over_limit;
971         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
972         struct res_counter *fail_res;
973
974         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
975                 /* Don't account this! */
976                 *memcg = NULL;
977                 return 0;
978         }
979
980         /*
981          * We always charge the cgroup the mm_struct belongs to.
982          * The mm_struct's mem_cgroup changes on task migration if the
983          * thread group leader migrates. It's possible that mm is not
984          * set, if so charge the init_mm (happens for pagecache usage).
985          */
986         mem = *memcg;
987         if (likely(!mem)) {
988                 mem = try_get_mem_cgroup_from_mm(mm);
989                 *memcg = mem;
990         } else {
991                 css_get(&mem->css);
992         }
993         if (unlikely(!mem))
994                 return 0;
995
996         VM_BUG_ON(css_is_removed(&mem->css));
997
998         while (1) {
999                 int ret;
1000                 bool noswap = false;
1001
1002                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
1003                 if (likely(!ret)) {
1004                         if (!do_swap_account)
1005                                 break;
1006                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
1007                                                         &fail_res);
1008                         if (likely(!ret))
1009                                 break;
1010                         /* mem+swap counter fails */
1011                         res_counter_uncharge(&mem->res, PAGE_SIZE);
1012                         noswap = true;
1013                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1014                                                                         memsw);
1015                 } else
1016                         /* mem counter fails */
1017                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1018                                                                         res);
1019
1020                 if (!(gfp_mask & __GFP_WAIT))
1021                         goto nomem;
1022
1023                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
1024                                                         noswap, false);
1025                 if (ret)
1026                         continue;
1027
1028                 /*
1029                  * try_to_free_mem_cgroup_pages() might not give us a full
1030                  * picture of reclaim. Some pages are reclaimed and might be
1031                  * moved to swap cache or just unmapped from the cgroup.
1032                  * Check the limit again to see if the reclaim reduced the
1033                  * current usage of the cgroup before giving up
1034                  *
1035                  */
1036                 if (mem_cgroup_check_under_limit(mem_over_limit))
1037                         continue;
1038
1039                 if (!nr_retries--) {
1040                         if (oom) {
1041                                 mutex_lock(&memcg_tasklist);
1042                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1043                                 mutex_unlock(&memcg_tasklist);
1044                                 record_last_oom(mem_over_limit);
1045                         }
1046                         goto nomem;
1047                 }
1048         }
1049         return 0;
1050 nomem:
1051         css_put(&mem->css);
1052         return -ENOMEM;
1053 }
1054
1055
1056 /*
1057  * A helper function to get mem_cgroup from ID. must be called under
1058  * rcu_read_lock(). The caller must check css_is_removed() or some if
1059  * it's concern. (dropping refcnt from swap can be called against removed
1060  * memcg.)
1061  */
1062 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1063 {
1064         struct cgroup_subsys_state *css;
1065
1066         /* ID 0 is unused ID */
1067         if (!id)
1068                 return NULL;
1069         css = css_lookup(&mem_cgroup_subsys, id);
1070         if (!css)
1071                 return NULL;
1072         return container_of(css, struct mem_cgroup, css);
1073 }
1074
1075 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1076 {
1077         struct mem_cgroup *mem;
1078         struct page_cgroup *pc;
1079         unsigned short id;
1080         swp_entry_t ent;
1081
1082         VM_BUG_ON(!PageLocked(page));
1083
1084         if (!PageSwapCache(page))
1085                 return NULL;
1086
1087         pc = lookup_page_cgroup(page);
1088         lock_page_cgroup(pc);
1089         if (PageCgroupUsed(pc)) {
1090                 mem = pc->mem_cgroup;
1091                 if (mem && !css_tryget(&mem->css))
1092                         mem = NULL;
1093         } else {
1094                 ent.val = page_private(page);
1095                 id = lookup_swap_cgroup(ent);
1096                 rcu_read_lock();
1097                 mem = mem_cgroup_lookup(id);
1098                 if (mem && !css_tryget(&mem->css))
1099                         mem = NULL;
1100                 rcu_read_unlock();
1101         }
1102         unlock_page_cgroup(pc);
1103         return mem;
1104 }
1105
1106 /*
1107  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1108  * USED state. If already USED, uncharge and return.
1109  */
1110
1111 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1112                                      struct page_cgroup *pc,
1113                                      enum charge_type ctype)
1114 {
1115         /* try_charge() can return NULL to *memcg, taking care of it. */
1116         if (!mem)
1117                 return;
1118
1119         lock_page_cgroup(pc);
1120         if (unlikely(PageCgroupUsed(pc))) {
1121                 unlock_page_cgroup(pc);
1122                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1123                 if (do_swap_account)
1124                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1125                 css_put(&mem->css);
1126                 return;
1127         }
1128         pc->mem_cgroup = mem;
1129         smp_wmb();
1130         pc->flags = pcg_default_flags[ctype];
1131
1132         mem_cgroup_charge_statistics(mem, pc, true);
1133
1134         unlock_page_cgroup(pc);
1135 }
1136
1137 /**
1138  * mem_cgroup_move_account - move account of the page
1139  * @pc: page_cgroup of the page.
1140  * @from: mem_cgroup which the page is moved from.
1141  * @to: mem_cgroup which the page is moved to. @from != @to.
1142  *
1143  * The caller must confirm following.
1144  * - page is not on LRU (isolate_page() is useful.)
1145  *
1146  * returns 0 at success,
1147  * returns -EBUSY when lock is busy or "pc" is unstable.
1148  *
1149  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1150  * new cgroup. It should be done by a caller.
1151  */
1152
1153 static int mem_cgroup_move_account(struct page_cgroup *pc,
1154         struct mem_cgroup *from, struct mem_cgroup *to)
1155 {
1156         struct mem_cgroup_per_zone *from_mz, *to_mz;
1157         int nid, zid;
1158         int ret = -EBUSY;
1159         struct page *page;
1160         int cpu;
1161         struct mem_cgroup_stat *stat;
1162         struct mem_cgroup_stat_cpu *cpustat;
1163
1164         VM_BUG_ON(from == to);
1165         VM_BUG_ON(PageLRU(pc->page));
1166
1167         nid = page_cgroup_nid(pc);
1168         zid = page_cgroup_zid(pc);
1169         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1170         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1171
1172         if (!trylock_page_cgroup(pc))
1173                 return ret;
1174
1175         if (!PageCgroupUsed(pc))
1176                 goto out;
1177
1178         if (pc->mem_cgroup != from)
1179                 goto out;
1180
1181         res_counter_uncharge(&from->res, PAGE_SIZE);
1182         mem_cgroup_charge_statistics(from, pc, false);
1183
1184         page = pc->page;
1185         if (page_is_file_cache(page) && page_mapped(page)) {
1186                 cpu = smp_processor_id();
1187                 /* Update mapped_file data for mem_cgroup "from" */
1188                 stat = &from->stat;
1189                 cpustat = &stat->cpustat[cpu];
1190                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1191                                                 -1);
1192
1193                 /* Update mapped_file data for mem_cgroup "to" */
1194                 stat = &to->stat;
1195                 cpustat = &stat->cpustat[cpu];
1196                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1197                                                 1);
1198         }
1199
1200         if (do_swap_account)
1201                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1202         css_put(&from->css);
1203
1204         css_get(&to->css);
1205         pc->mem_cgroup = to;
1206         mem_cgroup_charge_statistics(to, pc, true);
1207         ret = 0;
1208 out:
1209         unlock_page_cgroup(pc);
1210         return ret;
1211 }
1212
1213 /*
1214  * move charges to its parent.
1215  */
1216
1217 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1218                                   struct mem_cgroup *child,
1219                                   gfp_t gfp_mask)
1220 {
1221         struct page *page = pc->page;
1222         struct cgroup *cg = child->css.cgroup;
1223         struct cgroup *pcg = cg->parent;
1224         struct mem_cgroup *parent;
1225         int ret;
1226
1227         /* Is ROOT ? */
1228         if (!pcg)
1229                 return -EINVAL;
1230
1231
1232         parent = mem_cgroup_from_cont(pcg);
1233
1234
1235         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1236         if (ret || !parent)
1237                 return ret;
1238
1239         if (!get_page_unless_zero(page)) {
1240                 ret = -EBUSY;
1241                 goto uncharge;
1242         }
1243
1244         ret = isolate_lru_page(page);
1245
1246         if (ret)
1247                 goto cancel;
1248
1249         ret = mem_cgroup_move_account(pc, child, parent);
1250
1251         putback_lru_page(page);
1252         if (!ret) {
1253                 put_page(page);
1254                 /* drop extra refcnt by try_charge() */
1255                 css_put(&parent->css);
1256                 return 0;
1257         }
1258
1259 cancel:
1260         put_page(page);
1261 uncharge:
1262         /* drop extra refcnt by try_charge() */
1263         css_put(&parent->css);
1264         /* uncharge if move fails */
1265         res_counter_uncharge(&parent->res, PAGE_SIZE);
1266         if (do_swap_account)
1267                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1268         return ret;
1269 }
1270
1271 /*
1272  * Charge the memory controller for page usage.
1273  * Return
1274  * 0 if the charge was successful
1275  * < 0 if the cgroup is over its limit
1276  */
1277 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1278                                 gfp_t gfp_mask, enum charge_type ctype,
1279                                 struct mem_cgroup *memcg)
1280 {
1281         struct mem_cgroup *mem;
1282         struct page_cgroup *pc;
1283         int ret;
1284
1285         pc = lookup_page_cgroup(page);
1286         /* can happen at boot */
1287         if (unlikely(!pc))
1288                 return 0;
1289         prefetchw(pc);
1290
1291         mem = memcg;
1292         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1293         if (ret || !mem)
1294                 return ret;
1295
1296         __mem_cgroup_commit_charge(mem, pc, ctype);
1297         return 0;
1298 }
1299
1300 int mem_cgroup_newpage_charge(struct page *page,
1301                               struct mm_struct *mm, gfp_t gfp_mask)
1302 {
1303         if (mem_cgroup_disabled())
1304                 return 0;
1305         if (PageCompound(page))
1306                 return 0;
1307         /*
1308          * If already mapped, we don't have to account.
1309          * If page cache, page->mapping has address_space.
1310          * But page->mapping may have out-of-use anon_vma pointer,
1311          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1312          * is NULL.
1313          */
1314         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1315                 return 0;
1316         if (unlikely(!mm))
1317                 mm = &init_mm;
1318         return mem_cgroup_charge_common(page, mm, gfp_mask,
1319                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1320 }
1321
1322 static void
1323 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1324                                         enum charge_type ctype);
1325
1326 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1327                                 gfp_t gfp_mask)
1328 {
1329         struct mem_cgroup *mem = NULL;
1330         int ret;
1331
1332         if (mem_cgroup_disabled())
1333                 return 0;
1334         if (PageCompound(page))
1335                 return 0;
1336         /*
1337          * Corner case handling. This is called from add_to_page_cache()
1338          * in usual. But some FS (shmem) precharges this page before calling it
1339          * and call add_to_page_cache() with GFP_NOWAIT.
1340          *
1341          * For GFP_NOWAIT case, the page may be pre-charged before calling
1342          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1343          * charge twice. (It works but has to pay a bit larger cost.)
1344          * And when the page is SwapCache, it should take swap information
1345          * into account. This is under lock_page() now.
1346          */
1347         if (!(gfp_mask & __GFP_WAIT)) {
1348                 struct page_cgroup *pc;
1349
1350
1351                 pc = lookup_page_cgroup(page);
1352                 if (!pc)
1353                         return 0;
1354                 lock_page_cgroup(pc);
1355                 if (PageCgroupUsed(pc)) {
1356                         unlock_page_cgroup(pc);
1357                         return 0;
1358                 }
1359                 unlock_page_cgroup(pc);
1360         }
1361
1362         if (unlikely(!mm && !mem))
1363                 mm = &init_mm;
1364
1365         if (page_is_file_cache(page))
1366                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1367                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1368
1369         /* shmem */
1370         if (PageSwapCache(page)) {
1371                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1372                 if (!ret)
1373                         __mem_cgroup_commit_charge_swapin(page, mem,
1374                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1375         } else
1376                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1377                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1378
1379         return ret;
1380 }
1381
1382 /*
1383  * While swap-in, try_charge -> commit or cancel, the page is locked.
1384  * And when try_charge() successfully returns, one refcnt to memcg without
1385  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1386  * "commit()" or removed by "cancel()"
1387  */
1388 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1389                                  struct page *page,
1390                                  gfp_t mask, struct mem_cgroup **ptr)
1391 {
1392         struct mem_cgroup *mem;
1393         int ret;
1394
1395         if (mem_cgroup_disabled())
1396                 return 0;
1397
1398         if (!do_swap_account)
1399                 goto charge_cur_mm;
1400         /*
1401          * A racing thread's fault, or swapoff, may have already updated
1402          * the pte, and even removed page from swap cache: return success
1403          * to go on to do_swap_page()'s pte_same() test, which should fail.
1404          */
1405         if (!PageSwapCache(page))
1406                 return 0;
1407         mem = try_get_mem_cgroup_from_swapcache(page);
1408         if (!mem)
1409                 goto charge_cur_mm;
1410         *ptr = mem;
1411         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1412         /* drop extra refcnt from tryget */
1413         css_put(&mem->css);
1414         return ret;
1415 charge_cur_mm:
1416         if (unlikely(!mm))
1417                 mm = &init_mm;
1418         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1419 }
1420
1421 static void
1422 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1423                                         enum charge_type ctype)
1424 {
1425         struct page_cgroup *pc;
1426
1427         if (mem_cgroup_disabled())
1428                 return;
1429         if (!ptr)
1430                 return;
1431         pc = lookup_page_cgroup(page);
1432         mem_cgroup_lru_del_before_commit_swapcache(page);
1433         __mem_cgroup_commit_charge(ptr, pc, ctype);
1434         mem_cgroup_lru_add_after_commit_swapcache(page);
1435         /*
1436          * Now swap is on-memory. This means this page may be
1437          * counted both as mem and swap....double count.
1438          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1439          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1440          * may call delete_from_swap_cache() before reach here.
1441          */
1442         if (do_swap_account && PageSwapCache(page)) {
1443                 swp_entry_t ent = {.val = page_private(page)};
1444                 unsigned short id;
1445                 struct mem_cgroup *memcg;
1446
1447                 id = swap_cgroup_record(ent, 0);
1448                 rcu_read_lock();
1449                 memcg = mem_cgroup_lookup(id);
1450                 if (memcg) {
1451                         /*
1452                          * This recorded memcg can be obsolete one. So, avoid
1453                          * calling css_tryget
1454                          */
1455                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1456                         mem_cgroup_put(memcg);
1457                 }
1458                 rcu_read_unlock();
1459         }
1460         /* add this page(page_cgroup) to the LRU we want. */
1461
1462 }
1463
1464 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1465 {
1466         __mem_cgroup_commit_charge_swapin(page, ptr,
1467                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1468 }
1469
1470 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1471 {
1472         if (mem_cgroup_disabled())
1473                 return;
1474         if (!mem)
1475                 return;
1476         res_counter_uncharge(&mem->res, PAGE_SIZE);
1477         if (do_swap_account)
1478                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1479         css_put(&mem->css);
1480 }
1481
1482
1483 /*
1484  * uncharge if !page_mapped(page)
1485  */
1486 static struct mem_cgroup *
1487 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1488 {
1489         struct page_cgroup *pc;
1490         struct mem_cgroup *mem = NULL;
1491         struct mem_cgroup_per_zone *mz;
1492
1493         if (mem_cgroup_disabled())
1494                 return NULL;
1495
1496         if (PageSwapCache(page))
1497                 return NULL;
1498
1499         /*
1500          * Check if our page_cgroup is valid
1501          */
1502         pc = lookup_page_cgroup(page);
1503         if (unlikely(!pc || !PageCgroupUsed(pc)))
1504                 return NULL;
1505
1506         lock_page_cgroup(pc);
1507
1508         mem = pc->mem_cgroup;
1509
1510         if (!PageCgroupUsed(pc))
1511                 goto unlock_out;
1512
1513         switch (ctype) {
1514         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1515         case MEM_CGROUP_CHARGE_TYPE_DROP:
1516                 if (page_mapped(page))
1517                         goto unlock_out;
1518                 break;
1519         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1520                 if (!PageAnon(page)) {  /* Shared memory */
1521                         if (page->mapping && !page_is_file_cache(page))
1522                                 goto unlock_out;
1523                 } else if (page_mapped(page)) /* Anon */
1524                                 goto unlock_out;
1525                 break;
1526         default:
1527                 break;
1528         }
1529
1530         res_counter_uncharge(&mem->res, PAGE_SIZE);
1531         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1532                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1533         mem_cgroup_charge_statistics(mem, pc, false);
1534
1535         ClearPageCgroupUsed(pc);
1536         /*
1537          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1538          * freed from LRU. This is safe because uncharged page is expected not
1539          * to be reused (freed soon). Exception is SwapCache, it's handled by
1540          * special functions.
1541          */
1542
1543         mz = page_cgroup_zoneinfo(pc);
1544         unlock_page_cgroup(pc);
1545
1546         /* at swapout, this memcg will be accessed to record to swap */
1547         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1548                 css_put(&mem->css);
1549
1550         return mem;
1551
1552 unlock_out:
1553         unlock_page_cgroup(pc);
1554         return NULL;
1555 }
1556
1557 void mem_cgroup_uncharge_page(struct page *page)
1558 {
1559         /* early check. */
1560         if (page_mapped(page))
1561                 return;
1562         if (page->mapping && !PageAnon(page))
1563                 return;
1564         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1565 }
1566
1567 void mem_cgroup_uncharge_cache_page(struct page *page)
1568 {
1569         VM_BUG_ON(page_mapped(page));
1570         VM_BUG_ON(page->mapping);
1571         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1572 }
1573
1574 #ifdef CONFIG_SWAP
1575 /*
1576  * called after __delete_from_swap_cache() and drop "page" account.
1577  * memcg information is recorded to swap_cgroup of "ent"
1578  */
1579 void
1580 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
1581 {
1582         struct mem_cgroup *memcg;
1583         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
1584
1585         if (!swapout) /* this was a swap cache but the swap is unused ! */
1586                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
1587
1588         memcg = __mem_cgroup_uncharge_common(page, ctype);
1589
1590         /* record memcg information */
1591         if (do_swap_account && swapout && memcg) {
1592                 swap_cgroup_record(ent, css_id(&memcg->css));
1593                 mem_cgroup_get(memcg);
1594         }
1595         if (swapout && memcg)
1596                 css_put(&memcg->css);
1597 }
1598 #endif
1599
1600 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1601 /*
1602  * called from swap_entry_free(). remove record in swap_cgroup and
1603  * uncharge "memsw" account.
1604  */
1605 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1606 {
1607         struct mem_cgroup *memcg;
1608         unsigned short id;
1609
1610         if (!do_swap_account)
1611                 return;
1612
1613         id = swap_cgroup_record(ent, 0);
1614         rcu_read_lock();
1615         memcg = mem_cgroup_lookup(id);
1616         if (memcg) {
1617                 /*
1618                  * We uncharge this because swap is freed.
1619                  * This memcg can be obsolete one. We avoid calling css_tryget
1620                  */
1621                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1622                 mem_cgroup_put(memcg);
1623         }
1624         rcu_read_unlock();
1625 }
1626 #endif
1627
1628 /*
1629  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1630  * page belongs to.
1631  */
1632 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1633 {
1634         struct page_cgroup *pc;
1635         struct mem_cgroup *mem = NULL;
1636         int ret = 0;
1637
1638         if (mem_cgroup_disabled())
1639                 return 0;
1640
1641         pc = lookup_page_cgroup(page);
1642         lock_page_cgroup(pc);
1643         if (PageCgroupUsed(pc)) {
1644                 mem = pc->mem_cgroup;
1645                 css_get(&mem->css);
1646         }
1647         unlock_page_cgroup(pc);
1648
1649         if (mem) {
1650                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1651                 css_put(&mem->css);
1652         }
1653         *ptr = mem;
1654         return ret;
1655 }
1656
1657 /* remove redundant charge if migration failed*/
1658 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1659                 struct page *oldpage, struct page *newpage)
1660 {
1661         struct page *target, *unused;
1662         struct page_cgroup *pc;
1663         enum charge_type ctype;
1664
1665         if (!mem)
1666                 return;
1667
1668         /* at migration success, oldpage->mapping is NULL. */
1669         if (oldpage->mapping) {
1670                 target = oldpage;
1671                 unused = NULL;
1672         } else {
1673                 target = newpage;
1674                 unused = oldpage;
1675         }
1676
1677         if (PageAnon(target))
1678                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1679         else if (page_is_file_cache(target))
1680                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1681         else
1682                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1683
1684         /* unused page is not on radix-tree now. */
1685         if (unused)
1686                 __mem_cgroup_uncharge_common(unused, ctype);
1687
1688         pc = lookup_page_cgroup(target);
1689         /*
1690          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1691          * So, double-counting is effectively avoided.
1692          */
1693         __mem_cgroup_commit_charge(mem, pc, ctype);
1694
1695         /*
1696          * Both of oldpage and newpage are still under lock_page().
1697          * Then, we don't have to care about race in radix-tree.
1698          * But we have to be careful that this page is unmapped or not.
1699          *
1700          * There is a case for !page_mapped(). At the start of
1701          * migration, oldpage was mapped. But now, it's zapped.
1702          * But we know *target* page is not freed/reused under us.
1703          * mem_cgroup_uncharge_page() does all necessary checks.
1704          */
1705         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1706                 mem_cgroup_uncharge_page(target);
1707 }
1708
1709 /*
1710  * A call to try to shrink memory usage on charge failure at shmem's swapin.
1711  * Calling hierarchical_reclaim is not enough because we should update
1712  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
1713  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
1714  * not from the memcg which this page would be charged to.
1715  * try_charge_swapin does all of these works properly.
1716  */
1717 int mem_cgroup_shmem_charge_fallback(struct page *page,
1718                             struct mm_struct *mm,
1719                             gfp_t gfp_mask)
1720 {
1721         struct mem_cgroup *mem = NULL;
1722         int ret;
1723
1724         if (mem_cgroup_disabled())
1725                 return 0;
1726
1727         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1728         if (!ret)
1729                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
1730
1731         return ret;
1732 }
1733
1734 static DEFINE_MUTEX(set_limit_mutex);
1735
1736 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1737                                 unsigned long long val)
1738 {
1739         int retry_count;
1740         int progress;
1741         u64 memswlimit;
1742         int ret = 0;
1743         int children = mem_cgroup_count_children(memcg);
1744         u64 curusage, oldusage;
1745
1746         /*
1747          * For keeping hierarchical_reclaim simple, how long we should retry
1748          * is depends on callers. We set our retry-count to be function
1749          * of # of children which we should visit in this loop.
1750          */
1751         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
1752
1753         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1754
1755         while (retry_count) {
1756                 if (signal_pending(current)) {
1757                         ret = -EINTR;
1758                         break;
1759                 }
1760                 /*
1761                  * Rather than hide all in some function, I do this in
1762                  * open coded manner. You see what this really does.
1763                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1764                  */
1765                 mutex_lock(&set_limit_mutex);
1766                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1767                 if (memswlimit < val) {
1768                         ret = -EINVAL;
1769                         mutex_unlock(&set_limit_mutex);
1770                         break;
1771                 }
1772                 ret = res_counter_set_limit(&memcg->res, val);
1773                 if (!ret) {
1774                         if (memswlimit == val)
1775                                 memcg->memsw_is_minimum = true;
1776                         else
1777                                 memcg->memsw_is_minimum = false;
1778                 }
1779                 mutex_unlock(&set_limit_mutex);
1780
1781                 if (!ret)
1782                         break;
1783
1784                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1785                                                    false, true);
1786                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1787                 /* Usage is reduced ? */
1788                 if (curusage >= oldusage)
1789                         retry_count--;
1790                 else
1791                         oldusage = curusage;
1792         }
1793
1794         return ret;
1795 }
1796
1797 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1798                                         unsigned long long val)
1799 {
1800         int retry_count;
1801         u64 memlimit, oldusage, curusage;
1802         int children = mem_cgroup_count_children(memcg);
1803         int ret = -EBUSY;
1804
1805         /* see mem_cgroup_resize_res_limit */
1806         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
1807         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1808         while (retry_count) {
1809                 if (signal_pending(current)) {
1810                         ret = -EINTR;
1811                         break;
1812                 }
1813                 /*
1814                  * Rather than hide all in some function, I do this in
1815                  * open coded manner. You see what this really does.
1816                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1817                  */
1818                 mutex_lock(&set_limit_mutex);
1819                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1820                 if (memlimit > val) {
1821                         ret = -EINVAL;
1822                         mutex_unlock(&set_limit_mutex);
1823                         break;
1824                 }
1825                 ret = res_counter_set_limit(&memcg->memsw, val);
1826                 if (!ret) {
1827                         if (memlimit == val)
1828                                 memcg->memsw_is_minimum = true;
1829                         else
1830                                 memcg->memsw_is_minimum = false;
1831                 }
1832                 mutex_unlock(&set_limit_mutex);
1833
1834                 if (!ret)
1835                         break;
1836
1837                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
1838                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1839                 /* Usage is reduced ? */
1840                 if (curusage >= oldusage)
1841                         retry_count--;
1842                 else
1843                         oldusage = curusage;
1844         }
1845         return ret;
1846 }
1847
1848 /*
1849  * This routine traverse page_cgroup in given list and drop them all.
1850  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1851  */
1852 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1853                                 int node, int zid, enum lru_list lru)
1854 {
1855         struct zone *zone;
1856         struct mem_cgroup_per_zone *mz;
1857         struct page_cgroup *pc, *busy;
1858         unsigned long flags, loop;
1859         struct list_head *list;
1860         int ret = 0;
1861
1862         zone = &NODE_DATA(node)->node_zones[zid];
1863         mz = mem_cgroup_zoneinfo(mem, node, zid);
1864         list = &mz->lists[lru];
1865
1866         loop = MEM_CGROUP_ZSTAT(mz, lru);
1867         /* give some margin against EBUSY etc...*/
1868         loop += 256;
1869         busy = NULL;
1870         while (loop--) {
1871                 ret = 0;
1872                 spin_lock_irqsave(&zone->lru_lock, flags);
1873                 if (list_empty(list)) {
1874                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1875                         break;
1876                 }
1877                 pc = list_entry(list->prev, struct page_cgroup, lru);
1878                 if (busy == pc) {
1879                         list_move(&pc->lru, list);
1880                         busy = 0;
1881                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1882                         continue;
1883                 }
1884                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1885
1886                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1887                 if (ret == -ENOMEM)
1888                         break;
1889
1890                 if (ret == -EBUSY || ret == -EINVAL) {
1891                         /* found lock contention or "pc" is obsolete. */
1892                         busy = pc;
1893                         cond_resched();
1894                 } else
1895                         busy = NULL;
1896         }
1897
1898         if (!ret && !list_empty(list))
1899                 return -EBUSY;
1900         return ret;
1901 }
1902
1903 /*
1904  * make mem_cgroup's charge to be 0 if there is no task.
1905  * This enables deleting this mem_cgroup.
1906  */
1907 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1908 {
1909         int ret;
1910         int node, zid, shrink;
1911         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1912         struct cgroup *cgrp = mem->css.cgroup;
1913
1914         css_get(&mem->css);
1915
1916         shrink = 0;
1917         /* should free all ? */
1918         if (free_all)
1919                 goto try_to_free;
1920 move_account:
1921         while (mem->res.usage > 0) {
1922                 ret = -EBUSY;
1923                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1924                         goto out;
1925                 ret = -EINTR;
1926                 if (signal_pending(current))
1927                         goto out;
1928                 /* This is for making all *used* pages to be on LRU. */
1929                 lru_add_drain_all();
1930                 ret = 0;
1931                 for_each_node_state(node, N_HIGH_MEMORY) {
1932                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1933                                 enum lru_list l;
1934                                 for_each_lru(l) {
1935                                         ret = mem_cgroup_force_empty_list(mem,
1936                                                         node, zid, l);
1937                                         if (ret)
1938                                                 break;
1939                                 }
1940                         }
1941                         if (ret)
1942                                 break;
1943                 }
1944                 /* it seems parent cgroup doesn't have enough mem */
1945                 if (ret == -ENOMEM)
1946                         goto try_to_free;
1947                 cond_resched();
1948         }
1949         ret = 0;
1950 out:
1951         css_put(&mem->css);
1952         return ret;
1953
1954 try_to_free:
1955         /* returns EBUSY if there is a task or if we come here twice. */
1956         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1957                 ret = -EBUSY;
1958                 goto out;
1959         }
1960         /* we call try-to-free pages for make this cgroup empty */
1961         lru_add_drain_all();
1962         /* try to free all pages in this cgroup */
1963         shrink = 1;
1964         while (nr_retries && mem->res.usage > 0) {
1965                 int progress;
1966
1967                 if (signal_pending(current)) {
1968                         ret = -EINTR;
1969                         goto out;
1970                 }
1971                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1972                                                 false, get_swappiness(mem));
1973                 if (!progress) {
1974                         nr_retries--;
1975                         /* maybe some writeback is necessary */
1976                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1977                 }
1978
1979         }
1980         lru_add_drain();
1981         /* try move_account...there may be some *locked* pages. */
1982         if (mem->res.usage)
1983                 goto move_account;
1984         ret = 0;
1985         goto out;
1986 }
1987
1988 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1989 {
1990         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1991 }
1992
1993
1994 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1995 {
1996         return mem_cgroup_from_cont(cont)->use_hierarchy;
1997 }
1998
1999 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2000                                         u64 val)
2001 {
2002         int retval = 0;
2003         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2004         struct cgroup *parent = cont->parent;
2005         struct mem_cgroup *parent_mem = NULL;
2006
2007         if (parent)
2008                 parent_mem = mem_cgroup_from_cont(parent);
2009
2010         cgroup_lock();
2011         /*
2012          * If parent's use_hiearchy is set, we can't make any modifications
2013          * in the child subtrees. If it is unset, then the change can
2014          * occur, provided the current cgroup has no children.
2015          *
2016          * For the root cgroup, parent_mem is NULL, we allow value to be
2017          * set if there are no children.
2018          */
2019         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2020                                 (val == 1 || val == 0)) {
2021                 if (list_empty(&cont->children))
2022                         mem->use_hierarchy = val;
2023                 else
2024                         retval = -EBUSY;
2025         } else
2026                 retval = -EINVAL;
2027         cgroup_unlock();
2028
2029         return retval;
2030 }
2031
2032 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2033 {
2034         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2035         u64 val = 0;
2036         int type, name;
2037
2038         type = MEMFILE_TYPE(cft->private);
2039         name = MEMFILE_ATTR(cft->private);
2040         switch (type) {
2041         case _MEM:
2042                 val = res_counter_read_u64(&mem->res, name);
2043                 break;
2044         case _MEMSWAP:
2045                 val = res_counter_read_u64(&mem->memsw, name);
2046                 break;
2047         default:
2048                 BUG();
2049                 break;
2050         }
2051         return val;
2052 }
2053 /*
2054  * The user of this function is...
2055  * RES_LIMIT.
2056  */
2057 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2058                             const char *buffer)
2059 {
2060         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2061         int type, name;
2062         unsigned long long val;
2063         int ret;
2064
2065         type = MEMFILE_TYPE(cft->private);
2066         name = MEMFILE_ATTR(cft->private);
2067         switch (name) {
2068         case RES_LIMIT:
2069                 /* This function does all necessary parse...reuse it */
2070                 ret = res_counter_memparse_write_strategy(buffer, &val);
2071                 if (ret)
2072                         break;
2073                 if (type == _MEM)
2074                         ret = mem_cgroup_resize_limit(memcg, val);
2075                 else
2076                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2077                 break;
2078         default:
2079                 ret = -EINVAL; /* should be BUG() ? */
2080                 break;
2081         }
2082         return ret;
2083 }
2084
2085 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2086                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2087 {
2088         struct cgroup *cgroup;
2089         unsigned long long min_limit, min_memsw_limit, tmp;
2090
2091         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2092         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2093         cgroup = memcg->css.cgroup;
2094         if (!memcg->use_hierarchy)
2095                 goto out;
2096
2097         while (cgroup->parent) {
2098                 cgroup = cgroup->parent;
2099                 memcg = mem_cgroup_from_cont(cgroup);
2100                 if (!memcg->use_hierarchy)
2101                         break;
2102                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2103                 min_limit = min(min_limit, tmp);
2104                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2105                 min_memsw_limit = min(min_memsw_limit, tmp);
2106         }
2107 out:
2108         *mem_limit = min_limit;
2109         *memsw_limit = min_memsw_limit;
2110         return;
2111 }
2112
2113 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2114 {
2115         struct mem_cgroup *mem;
2116         int type, name;
2117
2118         mem = mem_cgroup_from_cont(cont);
2119         type = MEMFILE_TYPE(event);
2120         name = MEMFILE_ATTR(event);
2121         switch (name) {
2122         case RES_MAX_USAGE:
2123                 if (type == _MEM)
2124                         res_counter_reset_max(&mem->res);
2125                 else
2126                         res_counter_reset_max(&mem->memsw);
2127                 break;
2128         case RES_FAILCNT:
2129                 if (type == _MEM)
2130                         res_counter_reset_failcnt(&mem->res);
2131                 else
2132                         res_counter_reset_failcnt(&mem->memsw);
2133                 break;
2134         }
2135         return 0;
2136 }
2137
2138
2139 /* For read statistics */
2140 enum {
2141         MCS_CACHE,
2142         MCS_RSS,
2143         MCS_MAPPED_FILE,
2144         MCS_PGPGIN,
2145         MCS_PGPGOUT,
2146         MCS_INACTIVE_ANON,
2147         MCS_ACTIVE_ANON,
2148         MCS_INACTIVE_FILE,
2149         MCS_ACTIVE_FILE,
2150         MCS_UNEVICTABLE,
2151         NR_MCS_STAT,
2152 };
2153
2154 struct mcs_total_stat {
2155         s64 stat[NR_MCS_STAT];
2156 };
2157
2158 struct {
2159         char *local_name;
2160         char *total_name;
2161 } memcg_stat_strings[NR_MCS_STAT] = {
2162         {"cache", "total_cache"},
2163         {"rss", "total_rss"},
2164         {"mapped_file", "total_mapped_file"},
2165         {"pgpgin", "total_pgpgin"},
2166         {"pgpgout", "total_pgpgout"},
2167         {"inactive_anon", "total_inactive_anon"},
2168         {"active_anon", "total_active_anon"},
2169         {"inactive_file", "total_inactive_file"},
2170         {"active_file", "total_active_file"},
2171         {"unevictable", "total_unevictable"}
2172 };
2173
2174
2175 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2176 {
2177         struct mcs_total_stat *s = data;
2178         s64 val;
2179
2180         /* per cpu stat */
2181         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2182         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2183         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2184         s->stat[MCS_RSS] += val * PAGE_SIZE;
2185         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE);
2186         s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE;
2187         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2188         s->stat[MCS_PGPGIN] += val;
2189         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2190         s->stat[MCS_PGPGOUT] += val;
2191
2192         /* per zone stat */
2193         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2194         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2195         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2196         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2197         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2198         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2199         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2200         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2201         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2202         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2203         return 0;
2204 }
2205
2206 static void
2207 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2208 {
2209         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2210 }
2211
2212 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2213                                  struct cgroup_map_cb *cb)
2214 {
2215         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2216         struct mcs_total_stat mystat;
2217         int i;
2218
2219         memset(&mystat, 0, sizeof(mystat));
2220         mem_cgroup_get_local_stat(mem_cont, &mystat);
2221
2222         for (i = 0; i < NR_MCS_STAT; i++)
2223                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2224
2225         /* Hierarchical information */
2226         {
2227                 unsigned long long limit, memsw_limit;
2228                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2229                 cb->fill(cb, "hierarchical_memory_limit", limit);
2230                 if (do_swap_account)
2231                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2232         }
2233
2234         memset(&mystat, 0, sizeof(mystat));
2235         mem_cgroup_get_total_stat(mem_cont, &mystat);
2236         for (i = 0; i < NR_MCS_STAT; i++)
2237                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2238
2239
2240 #ifdef CONFIG_DEBUG_VM
2241         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2242
2243         {
2244                 int nid, zid;
2245                 struct mem_cgroup_per_zone *mz;
2246                 unsigned long recent_rotated[2] = {0, 0};
2247                 unsigned long recent_scanned[2] = {0, 0};
2248
2249                 for_each_online_node(nid)
2250                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2251                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2252
2253                                 recent_rotated[0] +=
2254                                         mz->reclaim_stat.recent_rotated[0];
2255                                 recent_rotated[1] +=
2256                                         mz->reclaim_stat.recent_rotated[1];
2257                                 recent_scanned[0] +=
2258                                         mz->reclaim_stat.recent_scanned[0];
2259                                 recent_scanned[1] +=
2260                                         mz->reclaim_stat.recent_scanned[1];
2261                         }
2262                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2263                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2264                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2265                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2266         }
2267 #endif
2268
2269         return 0;
2270 }
2271
2272 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2273 {
2274         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2275
2276         return get_swappiness(memcg);
2277 }
2278
2279 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2280                                        u64 val)
2281 {
2282         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2283         struct mem_cgroup *parent;
2284
2285         if (val > 100)
2286                 return -EINVAL;
2287
2288         if (cgrp->parent == NULL)
2289                 return -EINVAL;
2290
2291         parent = mem_cgroup_from_cont(cgrp->parent);
2292
2293         cgroup_lock();
2294
2295         /* If under hierarchy, only empty-root can set this value */
2296         if ((parent->use_hierarchy) ||
2297             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2298                 cgroup_unlock();
2299                 return -EINVAL;
2300         }
2301
2302         spin_lock(&memcg->reclaim_param_lock);
2303         memcg->swappiness = val;
2304         spin_unlock(&memcg->reclaim_param_lock);
2305
2306         cgroup_unlock();
2307
2308         return 0;
2309 }
2310
2311
2312 static struct cftype mem_cgroup_files[] = {
2313         {
2314                 .name = "usage_in_bytes",
2315                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2316                 .read_u64 = mem_cgroup_read,
2317         },
2318         {
2319                 .name = "max_usage_in_bytes",
2320                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2321                 .trigger = mem_cgroup_reset,
2322                 .read_u64 = mem_cgroup_read,
2323         },
2324         {
2325                 .name = "limit_in_bytes",
2326                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2327                 .write_string = mem_cgroup_write,
2328                 .read_u64 = mem_cgroup_read,
2329         },
2330         {
2331                 .name = "failcnt",
2332                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2333                 .trigger = mem_cgroup_reset,
2334                 .read_u64 = mem_cgroup_read,
2335         },
2336         {
2337                 .name = "stat",
2338                 .read_map = mem_control_stat_show,
2339         },
2340         {
2341                 .name = "force_empty",
2342                 .trigger = mem_cgroup_force_empty_write,
2343         },
2344         {
2345                 .name = "use_hierarchy",
2346                 .write_u64 = mem_cgroup_hierarchy_write,
2347                 .read_u64 = mem_cgroup_hierarchy_read,
2348         },
2349         {
2350                 .name = "swappiness",
2351                 .read_u64 = mem_cgroup_swappiness_read,
2352                 .write_u64 = mem_cgroup_swappiness_write,
2353         },
2354 };
2355
2356 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2357 static struct cftype memsw_cgroup_files[] = {
2358         {
2359                 .name = "memsw.usage_in_bytes",
2360                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2361                 .read_u64 = mem_cgroup_read,
2362         },
2363         {
2364                 .name = "memsw.max_usage_in_bytes",
2365                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2366                 .trigger = mem_cgroup_reset,
2367                 .read_u64 = mem_cgroup_read,
2368         },
2369         {
2370                 .name = "memsw.limit_in_bytes",
2371                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2372                 .write_string = mem_cgroup_write,
2373                 .read_u64 = mem_cgroup_read,
2374         },
2375         {
2376                 .name = "memsw.failcnt",
2377                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2378                 .trigger = mem_cgroup_reset,
2379                 .read_u64 = mem_cgroup_read,
2380         },
2381 };
2382
2383 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2384 {
2385         if (!do_swap_account)
2386                 return 0;
2387         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2388                                 ARRAY_SIZE(memsw_cgroup_files));
2389 };
2390 #else
2391 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2392 {
2393         return 0;
2394 }
2395 #endif
2396
2397 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2398 {
2399         struct mem_cgroup_per_node *pn;
2400         struct mem_cgroup_per_zone *mz;
2401         enum lru_list l;
2402         int zone, tmp = node;
2403         /*
2404          * This routine is called against possible nodes.
2405          * But it's BUG to call kmalloc() against offline node.
2406          *
2407          * TODO: this routine can waste much memory for nodes which will
2408          *       never be onlined. It's better to use memory hotplug callback
2409          *       function.
2410          */
2411         if (!node_state(node, N_NORMAL_MEMORY))
2412                 tmp = -1;
2413         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2414         if (!pn)
2415                 return 1;
2416
2417         mem->info.nodeinfo[node] = pn;
2418         memset(pn, 0, sizeof(*pn));
2419
2420         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2421                 mz = &pn->zoneinfo[zone];
2422                 for_each_lru(l)
2423                         INIT_LIST_HEAD(&mz->lists[l]);
2424         }
2425         return 0;
2426 }
2427
2428 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2429 {
2430         kfree(mem->info.nodeinfo[node]);
2431 }
2432
2433 static int mem_cgroup_size(void)
2434 {
2435         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2436         return sizeof(struct mem_cgroup) + cpustat_size;
2437 }
2438
2439 static struct mem_cgroup *mem_cgroup_alloc(void)
2440 {
2441         struct mem_cgroup *mem;
2442         int size = mem_cgroup_size();
2443
2444         if (size < PAGE_SIZE)
2445                 mem = kmalloc(size, GFP_KERNEL);
2446         else
2447                 mem = vmalloc(size);
2448
2449         if (mem)
2450                 memset(mem, 0, size);
2451         return mem;
2452 }
2453
2454 /*
2455  * At destroying mem_cgroup, references from swap_cgroup can remain.
2456  * (scanning all at force_empty is too costly...)
2457  *
2458  * Instead of clearing all references at force_empty, we remember
2459  * the number of reference from swap_cgroup and free mem_cgroup when
2460  * it goes down to 0.
2461  *
2462  * Removal of cgroup itself succeeds regardless of refs from swap.
2463  */
2464
2465 static void __mem_cgroup_free(struct mem_cgroup *mem)
2466 {
2467         int node;
2468
2469         free_css_id(&mem_cgroup_subsys, &mem->css);
2470
2471         for_each_node_state(node, N_POSSIBLE)
2472                 free_mem_cgroup_per_zone_info(mem, node);
2473
2474         if (mem_cgroup_size() < PAGE_SIZE)
2475                 kfree(mem);
2476         else
2477                 vfree(mem);
2478 }
2479
2480 static void mem_cgroup_get(struct mem_cgroup *mem)
2481 {
2482         atomic_inc(&mem->refcnt);
2483 }
2484
2485 static void mem_cgroup_put(struct mem_cgroup *mem)
2486 {
2487         if (atomic_dec_and_test(&mem->refcnt)) {
2488                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2489                 __mem_cgroup_free(mem);
2490                 if (parent)
2491                         mem_cgroup_put(parent);
2492         }
2493 }
2494
2495 /*
2496  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2497  */
2498 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2499 {
2500         if (!mem->res.parent)
2501                 return NULL;
2502         return mem_cgroup_from_res_counter(mem->res.parent, res);
2503 }
2504
2505 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2506 static void __init enable_swap_cgroup(void)
2507 {
2508         if (!mem_cgroup_disabled() && really_do_swap_account)
2509                 do_swap_account = 1;
2510 }
2511 #else
2512 static void __init enable_swap_cgroup(void)
2513 {
2514 }
2515 #endif
2516
2517 static struct cgroup_subsys_state * __ref
2518 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2519 {
2520         struct mem_cgroup *mem, *parent;
2521         long error = -ENOMEM;
2522         int node;
2523
2524         mem = mem_cgroup_alloc();
2525         if (!mem)
2526                 return ERR_PTR(error);
2527
2528         for_each_node_state(node, N_POSSIBLE)
2529                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2530                         goto free_out;
2531         /* root ? */
2532         if (cont->parent == NULL) {
2533                 enable_swap_cgroup();
2534                 parent = NULL;
2535         } else {
2536                 parent = mem_cgroup_from_cont(cont->parent);
2537                 mem->use_hierarchy = parent->use_hierarchy;
2538         }
2539
2540         if (parent && parent->use_hierarchy) {
2541                 res_counter_init(&mem->res, &parent->res);
2542                 res_counter_init(&mem->memsw, &parent->memsw);
2543                 /*
2544                  * We increment refcnt of the parent to ensure that we can
2545                  * safely access it on res_counter_charge/uncharge.
2546                  * This refcnt will be decremented when freeing this
2547                  * mem_cgroup(see mem_cgroup_put).
2548                  */
2549                 mem_cgroup_get(parent);
2550         } else {
2551                 res_counter_init(&mem->res, NULL);
2552                 res_counter_init(&mem->memsw, NULL);
2553         }
2554         mem->last_scanned_child = 0;
2555         spin_lock_init(&mem->reclaim_param_lock);
2556
2557         if (parent)
2558                 mem->swappiness = get_swappiness(parent);
2559         atomic_set(&mem->refcnt, 1);
2560         return &mem->css;
2561 free_out:
2562         __mem_cgroup_free(mem);
2563         return ERR_PTR(error);
2564 }
2565
2566 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2567                                         struct cgroup *cont)
2568 {
2569         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2570
2571         return mem_cgroup_force_empty(mem, false);
2572 }
2573
2574 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2575                                 struct cgroup *cont)
2576 {
2577         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2578
2579         mem_cgroup_put(mem);
2580 }
2581
2582 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2583                                 struct cgroup *cont)
2584 {
2585         int ret;
2586
2587         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2588                                 ARRAY_SIZE(mem_cgroup_files));
2589
2590         if (!ret)
2591                 ret = register_memsw_files(cont, ss);
2592         return ret;
2593 }
2594
2595 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2596                                 struct cgroup *cont,
2597                                 struct cgroup *old_cont,
2598                                 struct task_struct *p)
2599 {
2600         mutex_lock(&memcg_tasklist);
2601         /*
2602          * FIXME: It's better to move charges of this process from old
2603          * memcg to new memcg. But it's just on TODO-List now.
2604          */
2605         mutex_unlock(&memcg_tasklist);
2606 }
2607
2608 struct cgroup_subsys mem_cgroup_subsys = {
2609         .name = "memory",
2610         .subsys_id = mem_cgroup_subsys_id,
2611         .create = mem_cgroup_create,
2612         .pre_destroy = mem_cgroup_pre_destroy,
2613         .destroy = mem_cgroup_destroy,
2614         .populate = mem_cgroup_populate,
2615         .attach = mem_cgroup_move_task,
2616         .early_init = 0,
2617         .use_id = 1,
2618 };
2619
2620 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2621
2622 static int __init disable_swap_account(char *s)
2623 {
2624         really_do_swap_account = 0;
2625         return 1;
2626 }
2627 __setup("noswapaccount", disable_swap_account);
2628 #endif