memcg: fix swap accounting
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  */
19
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/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         /*
181          * statistics. This must be placed at the end of memcg.
182          */
183         struct mem_cgroup_stat stat;
184 };
185
186 enum charge_type {
187         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
188         MEM_CGROUP_CHARGE_TYPE_MAPPED,
189         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
190         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
191         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
192         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
193         NR_CHARGE_TYPE,
194 };
195
196 /* only for here (for easy reading.) */
197 #define PCGF_CACHE      (1UL << PCG_CACHE)
198 #define PCGF_USED       (1UL << PCG_USED)
199 #define PCGF_LOCK       (1UL << PCG_LOCK)
200 static const unsigned long
201 pcg_default_flags[NR_CHARGE_TYPE] = {
202         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
203         PCGF_USED | PCGF_LOCK, /* Anon */
204         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
205         0, /* FORCE */
206 };
207
208 /* for encoding cft->private value on file */
209 #define _MEM                    (0)
210 #define _MEMSWAP                (1)
211 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
212 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
213 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
214
215 static void mem_cgroup_get(struct mem_cgroup *mem);
216 static void mem_cgroup_put(struct mem_cgroup *mem);
217 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
218
219 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
220                                          struct page_cgroup *pc,
221                                          bool charge)
222 {
223         int val = (charge)? 1 : -1;
224         struct mem_cgroup_stat *stat = &mem->stat;
225         struct mem_cgroup_stat_cpu *cpustat;
226         int cpu = get_cpu();
227
228         cpustat = &stat->cpustat[cpu];
229         if (PageCgroupCache(pc))
230                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
231         else
232                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
233
234         if (charge)
235                 __mem_cgroup_stat_add_safe(cpustat,
236                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
237         else
238                 __mem_cgroup_stat_add_safe(cpustat,
239                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
240         put_cpu();
241 }
242
243 static struct mem_cgroup_per_zone *
244 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
245 {
246         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
247 }
248
249 static struct mem_cgroup_per_zone *
250 page_cgroup_zoneinfo(struct page_cgroup *pc)
251 {
252         struct mem_cgroup *mem = pc->mem_cgroup;
253         int nid = page_cgroup_nid(pc);
254         int zid = page_cgroup_zid(pc);
255
256         if (!mem)
257                 return NULL;
258
259         return mem_cgroup_zoneinfo(mem, nid, zid);
260 }
261
262 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
263                                         enum lru_list idx)
264 {
265         int nid, zid;
266         struct mem_cgroup_per_zone *mz;
267         u64 total = 0;
268
269         for_each_online_node(nid)
270                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
271                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
272                         total += MEM_CGROUP_ZSTAT(mz, idx);
273                 }
274         return total;
275 }
276
277 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
278 {
279         return container_of(cgroup_subsys_state(cont,
280                                 mem_cgroup_subsys_id), struct mem_cgroup,
281                                 css);
282 }
283
284 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
285 {
286         /*
287          * mm_update_next_owner() may clear mm->owner to NULL
288          * if it races with swapoff, page migration, etc.
289          * So this can be called with p == NULL.
290          */
291         if (unlikely(!p))
292                 return NULL;
293
294         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
295                                 struct mem_cgroup, css);
296 }
297
298 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
299 {
300         struct mem_cgroup *mem = NULL;
301
302         if (!mm)
303                 return NULL;
304         /*
305          * Because we have no locks, mm->owner's may be being moved to other
306          * cgroup. We use css_tryget() here even if this looks
307          * pessimistic (rather than adding locks here).
308          */
309         rcu_read_lock();
310         do {
311                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
312                 if (unlikely(!mem))
313                         break;
314         } while (!css_tryget(&mem->css));
315         rcu_read_unlock();
316         return mem;
317 }
318
319 /*
320  * Call callback function against all cgroup under hierarchy tree.
321  */
322 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
323                           int (*func)(struct mem_cgroup *, void *))
324 {
325         int found, ret, nextid;
326         struct cgroup_subsys_state *css;
327         struct mem_cgroup *mem;
328
329         if (!root->use_hierarchy)
330                 return (*func)(root, data);
331
332         nextid = 1;
333         do {
334                 ret = 0;
335                 mem = NULL;
336
337                 rcu_read_lock();
338                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
339                                    &found);
340                 if (css && css_tryget(css))
341                         mem = container_of(css, struct mem_cgroup, css);
342                 rcu_read_unlock();
343
344                 if (mem) {
345                         ret = (*func)(mem, data);
346                         css_put(&mem->css);
347                 }
348                 nextid = found + 1;
349         } while (!ret && css);
350
351         return ret;
352 }
353
354 /*
355  * Following LRU functions are allowed to be used without PCG_LOCK.
356  * Operations are called by routine of global LRU independently from memcg.
357  * What we have to take care of here is validness of pc->mem_cgroup.
358  *
359  * Changes to pc->mem_cgroup happens when
360  * 1. charge
361  * 2. moving account
362  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
363  * It is added to LRU before charge.
364  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
365  * When moving account, the page is not on LRU. It's isolated.
366  */
367
368 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
369 {
370         struct page_cgroup *pc;
371         struct mem_cgroup *mem;
372         struct mem_cgroup_per_zone *mz;
373
374         if (mem_cgroup_disabled())
375                 return;
376         pc = lookup_page_cgroup(page);
377         /* can happen while we handle swapcache. */
378         if (list_empty(&pc->lru) || !pc->mem_cgroup)
379                 return;
380         /*
381          * We don't check PCG_USED bit. It's cleared when the "page" is finally
382          * removed from global LRU.
383          */
384         mz = page_cgroup_zoneinfo(pc);
385         mem = pc->mem_cgroup;
386         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
387         list_del_init(&pc->lru);
388         return;
389 }
390
391 void mem_cgroup_del_lru(struct page *page)
392 {
393         mem_cgroup_del_lru_list(page, page_lru(page));
394 }
395
396 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
397 {
398         struct mem_cgroup_per_zone *mz;
399         struct page_cgroup *pc;
400
401         if (mem_cgroup_disabled())
402                 return;
403
404         pc = lookup_page_cgroup(page);
405         /*
406          * Used bit is set without atomic ops but after smp_wmb().
407          * For making pc->mem_cgroup visible, insert smp_rmb() here.
408          */
409         smp_rmb();
410         /* unused page is not rotated. */
411         if (!PageCgroupUsed(pc))
412                 return;
413         mz = page_cgroup_zoneinfo(pc);
414         list_move(&pc->lru, &mz->lists[lru]);
415 }
416
417 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
418 {
419         struct page_cgroup *pc;
420         struct mem_cgroup_per_zone *mz;
421
422         if (mem_cgroup_disabled())
423                 return;
424         pc = lookup_page_cgroup(page);
425         /*
426          * Used bit is set without atomic ops but after smp_wmb().
427          * For making pc->mem_cgroup visible, insert smp_rmb() here.
428          */
429         smp_rmb();
430         if (!PageCgroupUsed(pc))
431                 return;
432
433         mz = page_cgroup_zoneinfo(pc);
434         MEM_CGROUP_ZSTAT(mz, lru) += 1;
435         list_add(&pc->lru, &mz->lists[lru]);
436 }
437
438 /*
439  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
440  * lru because the page may.be reused after it's fully uncharged (because of
441  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
442  * it again. This function is only used to charge SwapCache. It's done under
443  * lock_page and expected that zone->lru_lock is never held.
444  */
445 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
446 {
447         unsigned long flags;
448         struct zone *zone = page_zone(page);
449         struct page_cgroup *pc = lookup_page_cgroup(page);
450
451         spin_lock_irqsave(&zone->lru_lock, flags);
452         /*
453          * Forget old LRU when this page_cgroup is *not* used. This Used bit
454          * is guarded by lock_page() because the page is SwapCache.
455          */
456         if (!PageCgroupUsed(pc))
457                 mem_cgroup_del_lru_list(page, page_lru(page));
458         spin_unlock_irqrestore(&zone->lru_lock, flags);
459 }
460
461 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
462 {
463         unsigned long flags;
464         struct zone *zone = page_zone(page);
465         struct page_cgroup *pc = lookup_page_cgroup(page);
466
467         spin_lock_irqsave(&zone->lru_lock, flags);
468         /* link when the page is linked to LRU but page_cgroup isn't */
469         if (PageLRU(page) && list_empty(&pc->lru))
470                 mem_cgroup_add_lru_list(page, page_lru(page));
471         spin_unlock_irqrestore(&zone->lru_lock, flags);
472 }
473
474
475 void mem_cgroup_move_lists(struct page *page,
476                            enum lru_list from, enum lru_list to)
477 {
478         if (mem_cgroup_disabled())
479                 return;
480         mem_cgroup_del_lru_list(page, from);
481         mem_cgroup_add_lru_list(page, to);
482 }
483
484 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
485 {
486         int ret;
487         struct mem_cgroup *curr = NULL;
488
489         task_lock(task);
490         rcu_read_lock();
491         curr = try_get_mem_cgroup_from_mm(task->mm);
492         rcu_read_unlock();
493         task_unlock(task);
494         if (!curr)
495                 return 0;
496         if (curr->use_hierarchy)
497                 ret = css_is_ancestor(&curr->css, &mem->css);
498         else
499                 ret = (curr == mem);
500         css_put(&curr->css);
501         return ret;
502 }
503
504 /*
505  * prev_priority control...this will be used in memory reclaim path.
506  */
507 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
508 {
509         int prev_priority;
510
511         spin_lock(&mem->reclaim_param_lock);
512         prev_priority = mem->prev_priority;
513         spin_unlock(&mem->reclaim_param_lock);
514
515         return prev_priority;
516 }
517
518 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
519 {
520         spin_lock(&mem->reclaim_param_lock);
521         if (priority < mem->prev_priority)
522                 mem->prev_priority = priority;
523         spin_unlock(&mem->reclaim_param_lock);
524 }
525
526 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
527 {
528         spin_lock(&mem->reclaim_param_lock);
529         mem->prev_priority = priority;
530         spin_unlock(&mem->reclaim_param_lock);
531 }
532
533 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
534 {
535         unsigned long active;
536         unsigned long inactive;
537         unsigned long gb;
538         unsigned long inactive_ratio;
539
540         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
541         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
542
543         gb = (inactive + active) >> (30 - PAGE_SHIFT);
544         if (gb)
545                 inactive_ratio = int_sqrt(10 * gb);
546         else
547                 inactive_ratio = 1;
548
549         if (present_pages) {
550                 present_pages[0] = inactive;
551                 present_pages[1] = active;
552         }
553
554         return inactive_ratio;
555 }
556
557 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
558 {
559         unsigned long active;
560         unsigned long inactive;
561         unsigned long present_pages[2];
562         unsigned long inactive_ratio;
563
564         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
565
566         inactive = present_pages[0];
567         active = present_pages[1];
568
569         if (inactive * inactive_ratio < active)
570                 return 1;
571
572         return 0;
573 }
574
575 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
576 {
577         unsigned long active;
578         unsigned long inactive;
579
580         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
581         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
582
583         return (active > inactive);
584 }
585
586 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
587                                        struct zone *zone,
588                                        enum lru_list lru)
589 {
590         int nid = zone->zone_pgdat->node_id;
591         int zid = zone_idx(zone);
592         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
593
594         return MEM_CGROUP_ZSTAT(mz, lru);
595 }
596
597 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
598                                                       struct zone *zone)
599 {
600         int nid = zone->zone_pgdat->node_id;
601         int zid = zone_idx(zone);
602         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
603
604         return &mz->reclaim_stat;
605 }
606
607 struct zone_reclaim_stat *
608 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
609 {
610         struct page_cgroup *pc;
611         struct mem_cgroup_per_zone *mz;
612
613         if (mem_cgroup_disabled())
614                 return NULL;
615
616         pc = lookup_page_cgroup(page);
617         /*
618          * Used bit is set without atomic ops but after smp_wmb().
619          * For making pc->mem_cgroup visible, insert smp_rmb() here.
620          */
621         smp_rmb();
622         if (!PageCgroupUsed(pc))
623                 return NULL;
624
625         mz = page_cgroup_zoneinfo(pc);
626         if (!mz)
627                 return NULL;
628
629         return &mz->reclaim_stat;
630 }
631
632 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
633                                         struct list_head *dst,
634                                         unsigned long *scanned, int order,
635                                         int mode, struct zone *z,
636                                         struct mem_cgroup *mem_cont,
637                                         int active, int file)
638 {
639         unsigned long nr_taken = 0;
640         struct page *page;
641         unsigned long scan;
642         LIST_HEAD(pc_list);
643         struct list_head *src;
644         struct page_cgroup *pc, *tmp;
645         int nid = z->zone_pgdat->node_id;
646         int zid = zone_idx(z);
647         struct mem_cgroup_per_zone *mz;
648         int lru = LRU_FILE * !!file + !!active;
649
650         BUG_ON(!mem_cont);
651         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
652         src = &mz->lists[lru];
653
654         scan = 0;
655         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
656                 if (scan >= nr_to_scan)
657                         break;
658
659                 page = pc->page;
660                 if (unlikely(!PageCgroupUsed(pc)))
661                         continue;
662                 if (unlikely(!PageLRU(page)))
663                         continue;
664
665                 scan++;
666                 if (__isolate_lru_page(page, mode, file) == 0) {
667                         list_move(&page->lru, dst);
668                         nr_taken++;
669                 }
670         }
671
672         *scanned = scan;
673         return nr_taken;
674 }
675
676 #define mem_cgroup_from_res_counter(counter, member)    \
677         container_of(counter, struct mem_cgroup, member)
678
679 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
680 {
681         if (do_swap_account) {
682                 if (res_counter_check_under_limit(&mem->res) &&
683                         res_counter_check_under_limit(&mem->memsw))
684                         return true;
685         } else
686                 if (res_counter_check_under_limit(&mem->res))
687                         return true;
688         return false;
689 }
690
691 static unsigned int get_swappiness(struct mem_cgroup *memcg)
692 {
693         struct cgroup *cgrp = memcg->css.cgroup;
694         unsigned int swappiness;
695
696         /* root ? */
697         if (cgrp->parent == NULL)
698                 return vm_swappiness;
699
700         spin_lock(&memcg->reclaim_param_lock);
701         swappiness = memcg->swappiness;
702         spin_unlock(&memcg->reclaim_param_lock);
703
704         return swappiness;
705 }
706
707 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
708 {
709         int *val = data;
710         (*val)++;
711         return 0;
712 }
713
714 /**
715  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
716  * @memcg: The memory cgroup that went over limit
717  * @p: Task that is going to be killed
718  *
719  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
720  * enabled
721  */
722 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
723 {
724         struct cgroup *task_cgrp;
725         struct cgroup *mem_cgrp;
726         /*
727          * Need a buffer in BSS, can't rely on allocations. The code relies
728          * on the assumption that OOM is serialized for memory controller.
729          * If this assumption is broken, revisit this code.
730          */
731         static char memcg_name[PATH_MAX];
732         int ret;
733
734         if (!memcg)
735                 return;
736
737
738         rcu_read_lock();
739
740         mem_cgrp = memcg->css.cgroup;
741         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
742
743         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
744         if (ret < 0) {
745                 /*
746                  * Unfortunately, we are unable to convert to a useful name
747                  * But we'll still print out the usage information
748                  */
749                 rcu_read_unlock();
750                 goto done;
751         }
752         rcu_read_unlock();
753
754         printk(KERN_INFO "Task in %s killed", memcg_name);
755
756         rcu_read_lock();
757         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
758         if (ret < 0) {
759                 rcu_read_unlock();
760                 goto done;
761         }
762         rcu_read_unlock();
763
764         /*
765          * Continues from above, so we don't need an KERN_ level
766          */
767         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
768 done:
769
770         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
771                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
772                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
773                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
774         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
775                 "failcnt %llu\n",
776                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
777                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
778                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
779 }
780
781 /*
782  * This function returns the number of memcg under hierarchy tree. Returns
783  * 1(self count) if no children.
784  */
785 static int mem_cgroup_count_children(struct mem_cgroup *mem)
786 {
787         int num = 0;
788         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
789         return num;
790 }
791
792 /*
793  * Visit the first child (need not be the first child as per the ordering
794  * of the cgroup list, since we track last_scanned_child) of @mem and use
795  * that to reclaim free pages from.
796  */
797 static struct mem_cgroup *
798 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
799 {
800         struct mem_cgroup *ret = NULL;
801         struct cgroup_subsys_state *css;
802         int nextid, found;
803
804         if (!root_mem->use_hierarchy) {
805                 css_get(&root_mem->css);
806                 ret = root_mem;
807         }
808
809         while (!ret) {
810                 rcu_read_lock();
811                 nextid = root_mem->last_scanned_child + 1;
812                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
813                                    &found);
814                 if (css && css_tryget(css))
815                         ret = container_of(css, struct mem_cgroup, css);
816
817                 rcu_read_unlock();
818                 /* Updates scanning parameter */
819                 spin_lock(&root_mem->reclaim_param_lock);
820                 if (!css) {
821                         /* this means start scan from ID:1 */
822                         root_mem->last_scanned_child = 0;
823                 } else
824                         root_mem->last_scanned_child = found;
825                 spin_unlock(&root_mem->reclaim_param_lock);
826         }
827
828         return ret;
829 }
830
831 /*
832  * Scan the hierarchy if needed to reclaim memory. We remember the last child
833  * we reclaimed from, so that we don't end up penalizing one child extensively
834  * based on its position in the children list.
835  *
836  * root_mem is the original ancestor that we've been reclaim from.
837  *
838  * We give up and return to the caller when we visit root_mem twice.
839  * (other groups can be removed while we're walking....)
840  *
841  * If shrink==true, for avoiding to free too much, this returns immedieately.
842  */
843 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
844                                    gfp_t gfp_mask, bool noswap, bool shrink)
845 {
846         struct mem_cgroup *victim;
847         int ret, total = 0;
848         int loop = 0;
849
850         while (loop < 2) {
851                 victim = mem_cgroup_select_victim(root_mem);
852                 if (victim == root_mem)
853                         loop++;
854                 if (!mem_cgroup_local_usage(&victim->stat)) {
855                         /* this cgroup's local usage == 0 */
856                         css_put(&victim->css);
857                         continue;
858                 }
859                 /* we use swappiness of local cgroup */
860                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
861                                                    get_swappiness(victim));
862                 css_put(&victim->css);
863                 /*
864                  * At shrinking usage, we can't check we should stop here or
865                  * reclaim more. It's depends on callers. last_scanned_child
866                  * will work enough for keeping fairness under tree.
867                  */
868                 if (shrink)
869                         return ret;
870                 total += ret;
871                 if (mem_cgroup_check_under_limit(root_mem))
872                         return 1 + total;
873         }
874         return total;
875 }
876
877 bool mem_cgroup_oom_called(struct task_struct *task)
878 {
879         bool ret = false;
880         struct mem_cgroup *mem;
881         struct mm_struct *mm;
882
883         rcu_read_lock();
884         mm = task->mm;
885         if (!mm)
886                 mm = &init_mm;
887         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
888         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
889                 ret = true;
890         rcu_read_unlock();
891         return ret;
892 }
893
894 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
895 {
896         mem->last_oom_jiffies = jiffies;
897         return 0;
898 }
899
900 static void record_last_oom(struct mem_cgroup *mem)
901 {
902         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
903 }
904
905 /*
906  * Currently used to update mapped file statistics, but the routine can be
907  * generalized to update other statistics as well.
908  */
909 void mem_cgroup_update_mapped_file_stat(struct page *page, int val)
910 {
911         struct mem_cgroup *mem;
912         struct mem_cgroup_stat *stat;
913         struct mem_cgroup_stat_cpu *cpustat;
914         int cpu;
915         struct page_cgroup *pc;
916
917         if (!page_is_file_cache(page))
918                 return;
919
920         pc = lookup_page_cgroup(page);
921         if (unlikely(!pc))
922                 return;
923
924         lock_page_cgroup(pc);
925         mem = pc->mem_cgroup;
926         if (!mem)
927                 goto done;
928
929         if (!PageCgroupUsed(pc))
930                 goto done;
931
932         /*
933          * Preemption is already disabled, we don't need get_cpu()
934          */
935         cpu = smp_processor_id();
936         stat = &mem->stat;
937         cpustat = &stat->cpustat[cpu];
938
939         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val);
940 done:
941         unlock_page_cgroup(pc);
942 }
943
944 /*
945  * Unlike exported interface, "oom" parameter is added. if oom==true,
946  * oom-killer can be invoked.
947  */
948 static int __mem_cgroup_try_charge(struct mm_struct *mm,
949                         gfp_t gfp_mask, struct mem_cgroup **memcg,
950                         bool oom)
951 {
952         struct mem_cgroup *mem, *mem_over_limit;
953         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
954         struct res_counter *fail_res;
955
956         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
957                 /* Don't account this! */
958                 *memcg = NULL;
959                 return 0;
960         }
961
962         /*
963          * We always charge the cgroup the mm_struct belongs to.
964          * The mm_struct's mem_cgroup changes on task migration if the
965          * thread group leader migrates. It's possible that mm is not
966          * set, if so charge the init_mm (happens for pagecache usage).
967          */
968         mem = *memcg;
969         if (likely(!mem)) {
970                 mem = try_get_mem_cgroup_from_mm(mm);
971                 *memcg = mem;
972         } else {
973                 css_get(&mem->css);
974         }
975         if (unlikely(!mem))
976                 return 0;
977
978         VM_BUG_ON(css_is_removed(&mem->css));
979
980         while (1) {
981                 int ret;
982                 bool noswap = false;
983
984                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
985                 if (likely(!ret)) {
986                         if (!do_swap_account)
987                                 break;
988                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
989                                                         &fail_res);
990                         if (likely(!ret))
991                                 break;
992                         /* mem+swap counter fails */
993                         res_counter_uncharge(&mem->res, PAGE_SIZE);
994                         noswap = true;
995                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
996                                                                         memsw);
997                 } else
998                         /* mem counter fails */
999                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1000                                                                         res);
1001
1002                 if (!(gfp_mask & __GFP_WAIT))
1003                         goto nomem;
1004
1005                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
1006                                                         noswap, false);
1007                 if (ret)
1008                         continue;
1009
1010                 /*
1011                  * try_to_free_mem_cgroup_pages() might not give us a full
1012                  * picture of reclaim. Some pages are reclaimed and might be
1013                  * moved to swap cache or just unmapped from the cgroup.
1014                  * Check the limit again to see if the reclaim reduced the
1015                  * current usage of the cgroup before giving up
1016                  *
1017                  */
1018                 if (mem_cgroup_check_under_limit(mem_over_limit))
1019                         continue;
1020
1021                 if (!nr_retries--) {
1022                         if (oom) {
1023                                 mutex_lock(&memcg_tasklist);
1024                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1025                                 mutex_unlock(&memcg_tasklist);
1026                                 record_last_oom(mem_over_limit);
1027                         }
1028                         goto nomem;
1029                 }
1030         }
1031         return 0;
1032 nomem:
1033         css_put(&mem->css);
1034         return -ENOMEM;
1035 }
1036
1037
1038 /*
1039  * A helper function to get mem_cgroup from ID. must be called under
1040  * rcu_read_lock(). The caller must check css_is_removed() or some if
1041  * it's concern. (dropping refcnt from swap can be called against removed
1042  * memcg.)
1043  */
1044 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1045 {
1046         struct cgroup_subsys_state *css;
1047
1048         /* ID 0 is unused ID */
1049         if (!id)
1050                 return NULL;
1051         css = css_lookup(&mem_cgroup_subsys, id);
1052         if (!css)
1053                 return NULL;
1054         return container_of(css, struct mem_cgroup, css);
1055 }
1056
1057 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1058 {
1059         struct mem_cgroup *mem;
1060         struct page_cgroup *pc;
1061         unsigned short id;
1062         swp_entry_t ent;
1063
1064         VM_BUG_ON(!PageLocked(page));
1065
1066         if (!PageSwapCache(page))
1067                 return NULL;
1068
1069         pc = lookup_page_cgroup(page);
1070         lock_page_cgroup(pc);
1071         if (PageCgroupUsed(pc)) {
1072                 mem = pc->mem_cgroup;
1073                 if (mem && !css_tryget(&mem->css))
1074                         mem = NULL;
1075         } else {
1076                 ent.val = page_private(page);
1077                 id = lookup_swap_cgroup(ent);
1078                 rcu_read_lock();
1079                 mem = mem_cgroup_lookup(id);
1080                 if (mem && !css_tryget(&mem->css))
1081                         mem = NULL;
1082                 rcu_read_unlock();
1083         }
1084         unlock_page_cgroup(pc);
1085         return mem;
1086 }
1087
1088 /*
1089  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1090  * USED state. If already USED, uncharge and return.
1091  */
1092
1093 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1094                                      struct page_cgroup *pc,
1095                                      enum charge_type ctype)
1096 {
1097         /* try_charge() can return NULL to *memcg, taking care of it. */
1098         if (!mem)
1099                 return;
1100
1101         lock_page_cgroup(pc);
1102         if (unlikely(PageCgroupUsed(pc))) {
1103                 unlock_page_cgroup(pc);
1104                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1105                 if (do_swap_account)
1106                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1107                 css_put(&mem->css);
1108                 return;
1109         }
1110         pc->mem_cgroup = mem;
1111         smp_wmb();
1112         pc->flags = pcg_default_flags[ctype];
1113
1114         mem_cgroup_charge_statistics(mem, pc, true);
1115
1116         unlock_page_cgroup(pc);
1117 }
1118
1119 /**
1120  * mem_cgroup_move_account - move account of the page
1121  * @pc: page_cgroup of the page.
1122  * @from: mem_cgroup which the page is moved from.
1123  * @to: mem_cgroup which the page is moved to. @from != @to.
1124  *
1125  * The caller must confirm following.
1126  * - page is not on LRU (isolate_page() is useful.)
1127  *
1128  * returns 0 at success,
1129  * returns -EBUSY when lock is busy or "pc" is unstable.
1130  *
1131  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1132  * new cgroup. It should be done by a caller.
1133  */
1134
1135 static int mem_cgroup_move_account(struct page_cgroup *pc,
1136         struct mem_cgroup *from, struct mem_cgroup *to)
1137 {
1138         struct mem_cgroup_per_zone *from_mz, *to_mz;
1139         int nid, zid;
1140         int ret = -EBUSY;
1141         struct page *page;
1142         int cpu;
1143         struct mem_cgroup_stat *stat;
1144         struct mem_cgroup_stat_cpu *cpustat;
1145
1146         VM_BUG_ON(from == to);
1147         VM_BUG_ON(PageLRU(pc->page));
1148
1149         nid = page_cgroup_nid(pc);
1150         zid = page_cgroup_zid(pc);
1151         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1152         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1153
1154         if (!trylock_page_cgroup(pc))
1155                 return ret;
1156
1157         if (!PageCgroupUsed(pc))
1158                 goto out;
1159
1160         if (pc->mem_cgroup != from)
1161                 goto out;
1162
1163         res_counter_uncharge(&from->res, PAGE_SIZE);
1164         mem_cgroup_charge_statistics(from, pc, false);
1165
1166         page = pc->page;
1167         if (page_is_file_cache(page) && page_mapped(page)) {
1168                 cpu = smp_processor_id();
1169                 /* Update mapped_file data for mem_cgroup "from" */
1170                 stat = &from->stat;
1171                 cpustat = &stat->cpustat[cpu];
1172                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1173                                                 -1);
1174
1175                 /* Update mapped_file data for mem_cgroup "to" */
1176                 stat = &to->stat;
1177                 cpustat = &stat->cpustat[cpu];
1178                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1179                                                 1);
1180         }
1181
1182         if (do_swap_account)
1183                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1184         css_put(&from->css);
1185
1186         css_get(&to->css);
1187         pc->mem_cgroup = to;
1188         mem_cgroup_charge_statistics(to, pc, true);
1189         ret = 0;
1190 out:
1191         unlock_page_cgroup(pc);
1192         return ret;
1193 }
1194
1195 /*
1196  * move charges to its parent.
1197  */
1198
1199 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1200                                   struct mem_cgroup *child,
1201                                   gfp_t gfp_mask)
1202 {
1203         struct page *page = pc->page;
1204         struct cgroup *cg = child->css.cgroup;
1205         struct cgroup *pcg = cg->parent;
1206         struct mem_cgroup *parent;
1207         int ret;
1208
1209         /* Is ROOT ? */
1210         if (!pcg)
1211                 return -EINVAL;
1212
1213
1214         parent = mem_cgroup_from_cont(pcg);
1215
1216
1217         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1218         if (ret || !parent)
1219                 return ret;
1220
1221         if (!get_page_unless_zero(page)) {
1222                 ret = -EBUSY;
1223                 goto uncharge;
1224         }
1225
1226         ret = isolate_lru_page(page);
1227
1228         if (ret)
1229                 goto cancel;
1230
1231         ret = mem_cgroup_move_account(pc, child, parent);
1232
1233         putback_lru_page(page);
1234         if (!ret) {
1235                 put_page(page);
1236                 /* drop extra refcnt by try_charge() */
1237                 css_put(&parent->css);
1238                 return 0;
1239         }
1240
1241 cancel:
1242         put_page(page);
1243 uncharge:
1244         /* drop extra refcnt by try_charge() */
1245         css_put(&parent->css);
1246         /* uncharge if move fails */
1247         res_counter_uncharge(&parent->res, PAGE_SIZE);
1248         if (do_swap_account)
1249                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1250         return ret;
1251 }
1252
1253 /*
1254  * Charge the memory controller for page usage.
1255  * Return
1256  * 0 if the charge was successful
1257  * < 0 if the cgroup is over its limit
1258  */
1259 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1260                                 gfp_t gfp_mask, enum charge_type ctype,
1261                                 struct mem_cgroup *memcg)
1262 {
1263         struct mem_cgroup *mem;
1264         struct page_cgroup *pc;
1265         int ret;
1266
1267         pc = lookup_page_cgroup(page);
1268         /* can happen at boot */
1269         if (unlikely(!pc))
1270                 return 0;
1271         prefetchw(pc);
1272
1273         mem = memcg;
1274         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1275         if (ret || !mem)
1276                 return ret;
1277
1278         __mem_cgroup_commit_charge(mem, pc, ctype);
1279         return 0;
1280 }
1281
1282 int mem_cgroup_newpage_charge(struct page *page,
1283                               struct mm_struct *mm, gfp_t gfp_mask)
1284 {
1285         if (mem_cgroup_disabled())
1286                 return 0;
1287         if (PageCompound(page))
1288                 return 0;
1289         /*
1290          * If already mapped, we don't have to account.
1291          * If page cache, page->mapping has address_space.
1292          * But page->mapping may have out-of-use anon_vma pointer,
1293          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1294          * is NULL.
1295          */
1296         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1297                 return 0;
1298         if (unlikely(!mm))
1299                 mm = &init_mm;
1300         return mem_cgroup_charge_common(page, mm, gfp_mask,
1301                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1302 }
1303
1304 static void
1305 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1306                                         enum charge_type ctype);
1307
1308 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1309                                 gfp_t gfp_mask)
1310 {
1311         struct mem_cgroup *mem = NULL;
1312         int ret;
1313
1314         if (mem_cgroup_disabled())
1315                 return 0;
1316         if (PageCompound(page))
1317                 return 0;
1318         /*
1319          * Corner case handling. This is called from add_to_page_cache()
1320          * in usual. But some FS (shmem) precharges this page before calling it
1321          * and call add_to_page_cache() with GFP_NOWAIT.
1322          *
1323          * For GFP_NOWAIT case, the page may be pre-charged before calling
1324          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1325          * charge twice. (It works but has to pay a bit larger cost.)
1326          * And when the page is SwapCache, it should take swap information
1327          * into account. This is under lock_page() now.
1328          */
1329         if (!(gfp_mask & __GFP_WAIT)) {
1330                 struct page_cgroup *pc;
1331
1332
1333                 pc = lookup_page_cgroup(page);
1334                 if (!pc)
1335                         return 0;
1336                 lock_page_cgroup(pc);
1337                 if (PageCgroupUsed(pc)) {
1338                         unlock_page_cgroup(pc);
1339                         return 0;
1340                 }
1341                 unlock_page_cgroup(pc);
1342         }
1343
1344         if (unlikely(!mm && !mem))
1345                 mm = &init_mm;
1346
1347         if (page_is_file_cache(page))
1348                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1349                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1350
1351         /* shmem */
1352         if (PageSwapCache(page)) {
1353                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1354                 if (!ret)
1355                         __mem_cgroup_commit_charge_swapin(page, mem,
1356                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1357         } else
1358                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1359                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1360
1361         return ret;
1362 }
1363
1364 /*
1365  * While swap-in, try_charge -> commit or cancel, the page is locked.
1366  * And when try_charge() successfully returns, one refcnt to memcg without
1367  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1368  * "commit()" or removed by "cancel()"
1369  */
1370 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1371                                  struct page *page,
1372                                  gfp_t mask, struct mem_cgroup **ptr)
1373 {
1374         struct mem_cgroup *mem;
1375         int ret;
1376
1377         if (mem_cgroup_disabled())
1378                 return 0;
1379
1380         if (!do_swap_account)
1381                 goto charge_cur_mm;
1382         /*
1383          * A racing thread's fault, or swapoff, may have already updated
1384          * the pte, and even removed page from swap cache: return success
1385          * to go on to do_swap_page()'s pte_same() test, which should fail.
1386          */
1387         if (!PageSwapCache(page))
1388                 return 0;
1389         mem = try_get_mem_cgroup_from_swapcache(page);
1390         if (!mem)
1391                 goto charge_cur_mm;
1392         *ptr = mem;
1393         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1394         /* drop extra refcnt from tryget */
1395         css_put(&mem->css);
1396         return ret;
1397 charge_cur_mm:
1398         if (unlikely(!mm))
1399                 mm = &init_mm;
1400         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1401 }
1402
1403 static void
1404 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1405                                         enum charge_type ctype)
1406 {
1407         struct page_cgroup *pc;
1408
1409         if (mem_cgroup_disabled())
1410                 return;
1411         if (!ptr)
1412                 return;
1413         pc = lookup_page_cgroup(page);
1414         mem_cgroup_lru_del_before_commit_swapcache(page);
1415         __mem_cgroup_commit_charge(ptr, pc, ctype);
1416         mem_cgroup_lru_add_after_commit_swapcache(page);
1417         /*
1418          * Now swap is on-memory. This means this page may be
1419          * counted both as mem and swap....double count.
1420          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1421          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1422          * may call delete_from_swap_cache() before reach here.
1423          */
1424         if (do_swap_account && PageSwapCache(page)) {
1425                 swp_entry_t ent = {.val = page_private(page)};
1426                 unsigned short id;
1427                 struct mem_cgroup *memcg;
1428
1429                 id = swap_cgroup_record(ent, 0);
1430                 rcu_read_lock();
1431                 memcg = mem_cgroup_lookup(id);
1432                 if (memcg) {
1433                         /*
1434                          * This recorded memcg can be obsolete one. So, avoid
1435                          * calling css_tryget
1436                          */
1437                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1438                         mem_cgroup_put(memcg);
1439                 }
1440                 rcu_read_unlock();
1441         }
1442         /* add this page(page_cgroup) to the LRU we want. */
1443
1444 }
1445
1446 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1447 {
1448         __mem_cgroup_commit_charge_swapin(page, ptr,
1449                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1450 }
1451
1452 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1453 {
1454         if (mem_cgroup_disabled())
1455                 return;
1456         if (!mem)
1457                 return;
1458         res_counter_uncharge(&mem->res, PAGE_SIZE);
1459         if (do_swap_account)
1460                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1461         css_put(&mem->css);
1462 }
1463
1464
1465 /*
1466  * uncharge if !page_mapped(page)
1467  */
1468 static struct mem_cgroup *
1469 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1470 {
1471         struct page_cgroup *pc;
1472         struct mem_cgroup *mem = NULL;
1473         struct mem_cgroup_per_zone *mz;
1474
1475         if (mem_cgroup_disabled())
1476                 return NULL;
1477
1478         if (PageSwapCache(page))
1479                 return NULL;
1480
1481         /*
1482          * Check if our page_cgroup is valid
1483          */
1484         pc = lookup_page_cgroup(page);
1485         if (unlikely(!pc || !PageCgroupUsed(pc)))
1486                 return NULL;
1487
1488         lock_page_cgroup(pc);
1489
1490         mem = pc->mem_cgroup;
1491
1492         if (!PageCgroupUsed(pc))
1493                 goto unlock_out;
1494
1495         switch (ctype) {
1496         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1497         case MEM_CGROUP_CHARGE_TYPE_DROP:
1498                 if (page_mapped(page))
1499                         goto unlock_out;
1500                 break;
1501         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1502                 if (!PageAnon(page)) {  /* Shared memory */
1503                         if (page->mapping && !page_is_file_cache(page))
1504                                 goto unlock_out;
1505                 } else if (page_mapped(page)) /* Anon */
1506                                 goto unlock_out;
1507                 break;
1508         default:
1509                 break;
1510         }
1511
1512         res_counter_uncharge(&mem->res, PAGE_SIZE);
1513         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1514                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1515         mem_cgroup_charge_statistics(mem, pc, false);
1516
1517         ClearPageCgroupUsed(pc);
1518         /*
1519          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1520          * freed from LRU. This is safe because uncharged page is expected not
1521          * to be reused (freed soon). Exception is SwapCache, it's handled by
1522          * special functions.
1523          */
1524
1525         mz = page_cgroup_zoneinfo(pc);
1526         unlock_page_cgroup(pc);
1527
1528         /* at swapout, this memcg will be accessed to record to swap */
1529         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1530                 css_put(&mem->css);
1531
1532         return mem;
1533
1534 unlock_out:
1535         unlock_page_cgroup(pc);
1536         return NULL;
1537 }
1538
1539 void mem_cgroup_uncharge_page(struct page *page)
1540 {
1541         /* early check. */
1542         if (page_mapped(page))
1543                 return;
1544         if (page->mapping && !PageAnon(page))
1545                 return;
1546         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1547 }
1548
1549 void mem_cgroup_uncharge_cache_page(struct page *page)
1550 {
1551         VM_BUG_ON(page_mapped(page));
1552         VM_BUG_ON(page->mapping);
1553         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1554 }
1555
1556 #ifdef CONFIG_SWAP
1557 /*
1558  * called after __delete_from_swap_cache() and drop "page" account.
1559  * memcg information is recorded to swap_cgroup of "ent"
1560  */
1561 void
1562 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
1563 {
1564         struct mem_cgroup *memcg;
1565         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
1566
1567         if (!swapout) /* this was a swap cache but the swap is unused ! */
1568                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
1569
1570         memcg = __mem_cgroup_uncharge_common(page, ctype);
1571
1572         /* record memcg information */
1573         if (do_swap_account && swapout && memcg) {
1574                 swap_cgroup_record(ent, css_id(&memcg->css));
1575                 mem_cgroup_get(memcg);
1576         }
1577         if (swapout && memcg)
1578                 css_put(&memcg->css);
1579 }
1580 #endif
1581
1582 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1583 /*
1584  * called from swap_entry_free(). remove record in swap_cgroup and
1585  * uncharge "memsw" account.
1586  */
1587 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1588 {
1589         struct mem_cgroup *memcg;
1590         unsigned short id;
1591
1592         if (!do_swap_account)
1593                 return;
1594
1595         id = swap_cgroup_record(ent, 0);
1596         rcu_read_lock();
1597         memcg = mem_cgroup_lookup(id);
1598         if (memcg) {
1599                 /*
1600                  * We uncharge this because swap is freed.
1601                  * This memcg can be obsolete one. We avoid calling css_tryget
1602                  */
1603                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1604                 mem_cgroup_put(memcg);
1605         }
1606         rcu_read_unlock();
1607 }
1608 #endif
1609
1610 /*
1611  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1612  * page belongs to.
1613  */
1614 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1615 {
1616         struct page_cgroup *pc;
1617         struct mem_cgroup *mem = NULL;
1618         int ret = 0;
1619
1620         if (mem_cgroup_disabled())
1621                 return 0;
1622
1623         pc = lookup_page_cgroup(page);
1624         lock_page_cgroup(pc);
1625         if (PageCgroupUsed(pc)) {
1626                 mem = pc->mem_cgroup;
1627                 css_get(&mem->css);
1628         }
1629         unlock_page_cgroup(pc);
1630
1631         if (mem) {
1632                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1633                 css_put(&mem->css);
1634         }
1635         *ptr = mem;
1636         return ret;
1637 }
1638
1639 /* remove redundant charge if migration failed*/
1640 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1641                 struct page *oldpage, struct page *newpage)
1642 {
1643         struct page *target, *unused;
1644         struct page_cgroup *pc;
1645         enum charge_type ctype;
1646
1647         if (!mem)
1648                 return;
1649
1650         /* at migration success, oldpage->mapping is NULL. */
1651         if (oldpage->mapping) {
1652                 target = oldpage;
1653                 unused = NULL;
1654         } else {
1655                 target = newpage;
1656                 unused = oldpage;
1657         }
1658
1659         if (PageAnon(target))
1660                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1661         else if (page_is_file_cache(target))
1662                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1663         else
1664                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1665
1666         /* unused page is not on radix-tree now. */
1667         if (unused)
1668                 __mem_cgroup_uncharge_common(unused, ctype);
1669
1670         pc = lookup_page_cgroup(target);
1671         /*
1672          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1673          * So, double-counting is effectively avoided.
1674          */
1675         __mem_cgroup_commit_charge(mem, pc, ctype);
1676
1677         /*
1678          * Both of oldpage and newpage are still under lock_page().
1679          * Then, we don't have to care about race in radix-tree.
1680          * But we have to be careful that this page is unmapped or not.
1681          *
1682          * There is a case for !page_mapped(). At the start of
1683          * migration, oldpage was mapped. But now, it's zapped.
1684          * But we know *target* page is not freed/reused under us.
1685          * mem_cgroup_uncharge_page() does all necessary checks.
1686          */
1687         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1688                 mem_cgroup_uncharge_page(target);
1689 }
1690
1691 /*
1692  * A call to try to shrink memory usage on charge failure at shmem's swapin.
1693  * Calling hierarchical_reclaim is not enough because we should update
1694  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
1695  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
1696  * not from the memcg which this page would be charged to.
1697  * try_charge_swapin does all of these works properly.
1698  */
1699 int mem_cgroup_shmem_charge_fallback(struct page *page,
1700                             struct mm_struct *mm,
1701                             gfp_t gfp_mask)
1702 {
1703         struct mem_cgroup *mem = NULL;
1704         int ret;
1705
1706         if (mem_cgroup_disabled())
1707                 return 0;
1708
1709         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1710         if (!ret)
1711                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
1712
1713         return ret;
1714 }
1715
1716 static DEFINE_MUTEX(set_limit_mutex);
1717
1718 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1719                                 unsigned long long val)
1720 {
1721         int retry_count;
1722         int progress;
1723         u64 memswlimit;
1724         int ret = 0;
1725         int children = mem_cgroup_count_children(memcg);
1726         u64 curusage, oldusage;
1727
1728         /*
1729          * For keeping hierarchical_reclaim simple, how long we should retry
1730          * is depends on callers. We set our retry-count to be function
1731          * of # of children which we should visit in this loop.
1732          */
1733         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
1734
1735         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1736
1737         while (retry_count) {
1738                 if (signal_pending(current)) {
1739                         ret = -EINTR;
1740                         break;
1741                 }
1742                 /*
1743                  * Rather than hide all in some function, I do this in
1744                  * open coded manner. You see what this really does.
1745                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1746                  */
1747                 mutex_lock(&set_limit_mutex);
1748                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1749                 if (memswlimit < val) {
1750                         ret = -EINVAL;
1751                         mutex_unlock(&set_limit_mutex);
1752                         break;
1753                 }
1754                 ret = res_counter_set_limit(&memcg->res, val);
1755                 mutex_unlock(&set_limit_mutex);
1756
1757                 if (!ret)
1758                         break;
1759
1760                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1761                                                    false, true);
1762                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1763                 /* Usage is reduced ? */
1764                 if (curusage >= oldusage)
1765                         retry_count--;
1766                 else
1767                         oldusage = curusage;
1768         }
1769
1770         return ret;
1771 }
1772
1773 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1774                                         unsigned long long val)
1775 {
1776         int retry_count;
1777         u64 memlimit, oldusage, curusage;
1778         int children = mem_cgroup_count_children(memcg);
1779         int ret = -EBUSY;
1780
1781         /* see mem_cgroup_resize_res_limit */
1782         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
1783         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1784         while (retry_count) {
1785                 if (signal_pending(current)) {
1786                         ret = -EINTR;
1787                         break;
1788                 }
1789                 /*
1790                  * Rather than hide all in some function, I do this in
1791                  * open coded manner. You see what this really does.
1792                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1793                  */
1794                 mutex_lock(&set_limit_mutex);
1795                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1796                 if (memlimit > val) {
1797                         ret = -EINVAL;
1798                         mutex_unlock(&set_limit_mutex);
1799                         break;
1800                 }
1801                 ret = res_counter_set_limit(&memcg->memsw, val);
1802                 mutex_unlock(&set_limit_mutex);
1803
1804                 if (!ret)
1805                         break;
1806
1807                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
1808                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1809                 /* Usage is reduced ? */
1810                 if (curusage >= oldusage)
1811                         retry_count--;
1812                 else
1813                         oldusage = curusage;
1814         }
1815         return ret;
1816 }
1817
1818 /*
1819  * This routine traverse page_cgroup in given list and drop them all.
1820  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1821  */
1822 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1823                                 int node, int zid, enum lru_list lru)
1824 {
1825         struct zone *zone;
1826         struct mem_cgroup_per_zone *mz;
1827         struct page_cgroup *pc, *busy;
1828         unsigned long flags, loop;
1829         struct list_head *list;
1830         int ret = 0;
1831
1832         zone = &NODE_DATA(node)->node_zones[zid];
1833         mz = mem_cgroup_zoneinfo(mem, node, zid);
1834         list = &mz->lists[lru];
1835
1836         loop = MEM_CGROUP_ZSTAT(mz, lru);
1837         /* give some margin against EBUSY etc...*/
1838         loop += 256;
1839         busy = NULL;
1840         while (loop--) {
1841                 ret = 0;
1842                 spin_lock_irqsave(&zone->lru_lock, flags);
1843                 if (list_empty(list)) {
1844                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1845                         break;
1846                 }
1847                 pc = list_entry(list->prev, struct page_cgroup, lru);
1848                 if (busy == pc) {
1849                         list_move(&pc->lru, list);
1850                         busy = 0;
1851                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1852                         continue;
1853                 }
1854                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1855
1856                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1857                 if (ret == -ENOMEM)
1858                         break;
1859
1860                 if (ret == -EBUSY || ret == -EINVAL) {
1861                         /* found lock contention or "pc" is obsolete. */
1862                         busy = pc;
1863                         cond_resched();
1864                 } else
1865                         busy = NULL;
1866         }
1867
1868         if (!ret && !list_empty(list))
1869                 return -EBUSY;
1870         return ret;
1871 }
1872
1873 /*
1874  * make mem_cgroup's charge to be 0 if there is no task.
1875  * This enables deleting this mem_cgroup.
1876  */
1877 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1878 {
1879         int ret;
1880         int node, zid, shrink;
1881         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1882         struct cgroup *cgrp = mem->css.cgroup;
1883
1884         css_get(&mem->css);
1885
1886         shrink = 0;
1887         /* should free all ? */
1888         if (free_all)
1889                 goto try_to_free;
1890 move_account:
1891         while (mem->res.usage > 0) {
1892                 ret = -EBUSY;
1893                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1894                         goto out;
1895                 ret = -EINTR;
1896                 if (signal_pending(current))
1897                         goto out;
1898                 /* This is for making all *used* pages to be on LRU. */
1899                 lru_add_drain_all();
1900                 ret = 0;
1901                 for_each_node_state(node, N_HIGH_MEMORY) {
1902                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1903                                 enum lru_list l;
1904                                 for_each_lru(l) {
1905                                         ret = mem_cgroup_force_empty_list(mem,
1906                                                         node, zid, l);
1907                                         if (ret)
1908                                                 break;
1909                                 }
1910                         }
1911                         if (ret)
1912                                 break;
1913                 }
1914                 /* it seems parent cgroup doesn't have enough mem */
1915                 if (ret == -ENOMEM)
1916                         goto try_to_free;
1917                 cond_resched();
1918         }
1919         ret = 0;
1920 out:
1921         css_put(&mem->css);
1922         return ret;
1923
1924 try_to_free:
1925         /* returns EBUSY if there is a task or if we come here twice. */
1926         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1927                 ret = -EBUSY;
1928                 goto out;
1929         }
1930         /* we call try-to-free pages for make this cgroup empty */
1931         lru_add_drain_all();
1932         /* try to free all pages in this cgroup */
1933         shrink = 1;
1934         while (nr_retries && mem->res.usage > 0) {
1935                 int progress;
1936
1937                 if (signal_pending(current)) {
1938                         ret = -EINTR;
1939                         goto out;
1940                 }
1941                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1942                                                 false, get_swappiness(mem));
1943                 if (!progress) {
1944                         nr_retries--;
1945                         /* maybe some writeback is necessary */
1946                         congestion_wait(WRITE, HZ/10);
1947                 }
1948
1949         }
1950         lru_add_drain();
1951         /* try move_account...there may be some *locked* pages. */
1952         if (mem->res.usage)
1953                 goto move_account;
1954         ret = 0;
1955         goto out;
1956 }
1957
1958 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1959 {
1960         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1961 }
1962
1963
1964 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1965 {
1966         return mem_cgroup_from_cont(cont)->use_hierarchy;
1967 }
1968
1969 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1970                                         u64 val)
1971 {
1972         int retval = 0;
1973         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1974         struct cgroup *parent = cont->parent;
1975         struct mem_cgroup *parent_mem = NULL;
1976
1977         if (parent)
1978                 parent_mem = mem_cgroup_from_cont(parent);
1979
1980         cgroup_lock();
1981         /*
1982          * If parent's use_hiearchy is set, we can't make any modifications
1983          * in the child subtrees. If it is unset, then the change can
1984          * occur, provided the current cgroup has no children.
1985          *
1986          * For the root cgroup, parent_mem is NULL, we allow value to be
1987          * set if there are no children.
1988          */
1989         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1990                                 (val == 1 || val == 0)) {
1991                 if (list_empty(&cont->children))
1992                         mem->use_hierarchy = val;
1993                 else
1994                         retval = -EBUSY;
1995         } else
1996                 retval = -EINVAL;
1997         cgroup_unlock();
1998
1999         return retval;
2000 }
2001
2002 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2003 {
2004         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2005         u64 val = 0;
2006         int type, name;
2007
2008         type = MEMFILE_TYPE(cft->private);
2009         name = MEMFILE_ATTR(cft->private);
2010         switch (type) {
2011         case _MEM:
2012                 val = res_counter_read_u64(&mem->res, name);
2013                 break;
2014         case _MEMSWAP:
2015                 val = res_counter_read_u64(&mem->memsw, name);
2016                 break;
2017         default:
2018                 BUG();
2019                 break;
2020         }
2021         return val;
2022 }
2023 /*
2024  * The user of this function is...
2025  * RES_LIMIT.
2026  */
2027 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2028                             const char *buffer)
2029 {
2030         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2031         int type, name;
2032         unsigned long long val;
2033         int ret;
2034
2035         type = MEMFILE_TYPE(cft->private);
2036         name = MEMFILE_ATTR(cft->private);
2037         switch (name) {
2038         case RES_LIMIT:
2039                 /* This function does all necessary parse...reuse it */
2040                 ret = res_counter_memparse_write_strategy(buffer, &val);
2041                 if (ret)
2042                         break;
2043                 if (type == _MEM)
2044                         ret = mem_cgroup_resize_limit(memcg, val);
2045                 else
2046                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2047                 break;
2048         default:
2049                 ret = -EINVAL; /* should be BUG() ? */
2050                 break;
2051         }
2052         return ret;
2053 }
2054
2055 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2056                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2057 {
2058         struct cgroup *cgroup;
2059         unsigned long long min_limit, min_memsw_limit, tmp;
2060
2061         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2062         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2063         cgroup = memcg->css.cgroup;
2064         if (!memcg->use_hierarchy)
2065                 goto out;
2066
2067         while (cgroup->parent) {
2068                 cgroup = cgroup->parent;
2069                 memcg = mem_cgroup_from_cont(cgroup);
2070                 if (!memcg->use_hierarchy)
2071                         break;
2072                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2073                 min_limit = min(min_limit, tmp);
2074                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2075                 min_memsw_limit = min(min_memsw_limit, tmp);
2076         }
2077 out:
2078         *mem_limit = min_limit;
2079         *memsw_limit = min_memsw_limit;
2080         return;
2081 }
2082
2083 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2084 {
2085         struct mem_cgroup *mem;
2086         int type, name;
2087
2088         mem = mem_cgroup_from_cont(cont);
2089         type = MEMFILE_TYPE(event);
2090         name = MEMFILE_ATTR(event);
2091         switch (name) {
2092         case RES_MAX_USAGE:
2093                 if (type == _MEM)
2094                         res_counter_reset_max(&mem->res);
2095                 else
2096                         res_counter_reset_max(&mem->memsw);
2097                 break;
2098         case RES_FAILCNT:
2099                 if (type == _MEM)
2100                         res_counter_reset_failcnt(&mem->res);
2101                 else
2102                         res_counter_reset_failcnt(&mem->memsw);
2103                 break;
2104         }
2105         return 0;
2106 }
2107
2108
2109 /* For read statistics */
2110 enum {
2111         MCS_CACHE,
2112         MCS_RSS,
2113         MCS_MAPPED_FILE,
2114         MCS_PGPGIN,
2115         MCS_PGPGOUT,
2116         MCS_INACTIVE_ANON,
2117         MCS_ACTIVE_ANON,
2118         MCS_INACTIVE_FILE,
2119         MCS_ACTIVE_FILE,
2120         MCS_UNEVICTABLE,
2121         NR_MCS_STAT,
2122 };
2123
2124 struct mcs_total_stat {
2125         s64 stat[NR_MCS_STAT];
2126 };
2127
2128 struct {
2129         char *local_name;
2130         char *total_name;
2131 } memcg_stat_strings[NR_MCS_STAT] = {
2132         {"cache", "total_cache"},
2133         {"rss", "total_rss"},
2134         {"mapped_file", "total_mapped_file"},
2135         {"pgpgin", "total_pgpgin"},
2136         {"pgpgout", "total_pgpgout"},
2137         {"inactive_anon", "total_inactive_anon"},
2138         {"active_anon", "total_active_anon"},
2139         {"inactive_file", "total_inactive_file"},
2140         {"active_file", "total_active_file"},
2141         {"unevictable", "total_unevictable"}
2142 };
2143
2144
2145 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2146 {
2147         struct mcs_total_stat *s = data;
2148         s64 val;
2149
2150         /* per cpu stat */
2151         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2152         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2153         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2154         s->stat[MCS_RSS] += val * PAGE_SIZE;
2155         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE);
2156         s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE;
2157         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2158         s->stat[MCS_PGPGIN] += val;
2159         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2160         s->stat[MCS_PGPGOUT] += val;
2161
2162         /* per zone stat */
2163         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2164         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2165         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2166         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2167         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2168         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2169         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2170         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2171         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2172         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2173         return 0;
2174 }
2175
2176 static void
2177 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2178 {
2179         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2180 }
2181
2182 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2183                                  struct cgroup_map_cb *cb)
2184 {
2185         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2186         struct mcs_total_stat mystat;
2187         int i;
2188
2189         memset(&mystat, 0, sizeof(mystat));
2190         mem_cgroup_get_local_stat(mem_cont, &mystat);
2191
2192         for (i = 0; i < NR_MCS_STAT; i++)
2193                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2194
2195         /* Hierarchical information */
2196         {
2197                 unsigned long long limit, memsw_limit;
2198                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2199                 cb->fill(cb, "hierarchical_memory_limit", limit);
2200                 if (do_swap_account)
2201                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2202         }
2203
2204         memset(&mystat, 0, sizeof(mystat));
2205         mem_cgroup_get_total_stat(mem_cont, &mystat);
2206         for (i = 0; i < NR_MCS_STAT; i++)
2207                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2208
2209
2210 #ifdef CONFIG_DEBUG_VM
2211         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2212
2213         {
2214                 int nid, zid;
2215                 struct mem_cgroup_per_zone *mz;
2216                 unsigned long recent_rotated[2] = {0, 0};
2217                 unsigned long recent_scanned[2] = {0, 0};
2218
2219                 for_each_online_node(nid)
2220                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2221                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2222
2223                                 recent_rotated[0] +=
2224                                         mz->reclaim_stat.recent_rotated[0];
2225                                 recent_rotated[1] +=
2226                                         mz->reclaim_stat.recent_rotated[1];
2227                                 recent_scanned[0] +=
2228                                         mz->reclaim_stat.recent_scanned[0];
2229                                 recent_scanned[1] +=
2230                                         mz->reclaim_stat.recent_scanned[1];
2231                         }
2232                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2233                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2234                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2235                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2236         }
2237 #endif
2238
2239         return 0;
2240 }
2241
2242 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2243 {
2244         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2245
2246         return get_swappiness(memcg);
2247 }
2248
2249 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2250                                        u64 val)
2251 {
2252         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2253         struct mem_cgroup *parent;
2254
2255         if (val > 100)
2256                 return -EINVAL;
2257
2258         if (cgrp->parent == NULL)
2259                 return -EINVAL;
2260
2261         parent = mem_cgroup_from_cont(cgrp->parent);
2262
2263         cgroup_lock();
2264
2265         /* If under hierarchy, only empty-root can set this value */
2266         if ((parent->use_hierarchy) ||
2267             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2268                 cgroup_unlock();
2269                 return -EINVAL;
2270         }
2271
2272         spin_lock(&memcg->reclaim_param_lock);
2273         memcg->swappiness = val;
2274         spin_unlock(&memcg->reclaim_param_lock);
2275
2276         cgroup_unlock();
2277
2278         return 0;
2279 }
2280
2281
2282 static struct cftype mem_cgroup_files[] = {
2283         {
2284                 .name = "usage_in_bytes",
2285                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2286                 .read_u64 = mem_cgroup_read,
2287         },
2288         {
2289                 .name = "max_usage_in_bytes",
2290                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2291                 .trigger = mem_cgroup_reset,
2292                 .read_u64 = mem_cgroup_read,
2293         },
2294         {
2295                 .name = "limit_in_bytes",
2296                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2297                 .write_string = mem_cgroup_write,
2298                 .read_u64 = mem_cgroup_read,
2299         },
2300         {
2301                 .name = "failcnt",
2302                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2303                 .trigger = mem_cgroup_reset,
2304                 .read_u64 = mem_cgroup_read,
2305         },
2306         {
2307                 .name = "stat",
2308                 .read_map = mem_control_stat_show,
2309         },
2310         {
2311                 .name = "force_empty",
2312                 .trigger = mem_cgroup_force_empty_write,
2313         },
2314         {
2315                 .name = "use_hierarchy",
2316                 .write_u64 = mem_cgroup_hierarchy_write,
2317                 .read_u64 = mem_cgroup_hierarchy_read,
2318         },
2319         {
2320                 .name = "swappiness",
2321                 .read_u64 = mem_cgroup_swappiness_read,
2322                 .write_u64 = mem_cgroup_swappiness_write,
2323         },
2324 };
2325
2326 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2327 static struct cftype memsw_cgroup_files[] = {
2328         {
2329                 .name = "memsw.usage_in_bytes",
2330                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2331                 .read_u64 = mem_cgroup_read,
2332         },
2333         {
2334                 .name = "memsw.max_usage_in_bytes",
2335                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2336                 .trigger = mem_cgroup_reset,
2337                 .read_u64 = mem_cgroup_read,
2338         },
2339         {
2340                 .name = "memsw.limit_in_bytes",
2341                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2342                 .write_string = mem_cgroup_write,
2343                 .read_u64 = mem_cgroup_read,
2344         },
2345         {
2346                 .name = "memsw.failcnt",
2347                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2348                 .trigger = mem_cgroup_reset,
2349                 .read_u64 = mem_cgroup_read,
2350         },
2351 };
2352
2353 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2354 {
2355         if (!do_swap_account)
2356                 return 0;
2357         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2358                                 ARRAY_SIZE(memsw_cgroup_files));
2359 };
2360 #else
2361 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2362 {
2363         return 0;
2364 }
2365 #endif
2366
2367 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2368 {
2369         struct mem_cgroup_per_node *pn;
2370         struct mem_cgroup_per_zone *mz;
2371         enum lru_list l;
2372         int zone, tmp = node;
2373         /*
2374          * This routine is called against possible nodes.
2375          * But it's BUG to call kmalloc() against offline node.
2376          *
2377          * TODO: this routine can waste much memory for nodes which will
2378          *       never be onlined. It's better to use memory hotplug callback
2379          *       function.
2380          */
2381         if (!node_state(node, N_NORMAL_MEMORY))
2382                 tmp = -1;
2383         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2384         if (!pn)
2385                 return 1;
2386
2387         mem->info.nodeinfo[node] = pn;
2388         memset(pn, 0, sizeof(*pn));
2389
2390         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2391                 mz = &pn->zoneinfo[zone];
2392                 for_each_lru(l)
2393                         INIT_LIST_HEAD(&mz->lists[l]);
2394         }
2395         return 0;
2396 }
2397
2398 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2399 {
2400         kfree(mem->info.nodeinfo[node]);
2401 }
2402
2403 static int mem_cgroup_size(void)
2404 {
2405         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2406         return sizeof(struct mem_cgroup) + cpustat_size;
2407 }
2408
2409 static struct mem_cgroup *mem_cgroup_alloc(void)
2410 {
2411         struct mem_cgroup *mem;
2412         int size = mem_cgroup_size();
2413
2414         if (size < PAGE_SIZE)
2415                 mem = kmalloc(size, GFP_KERNEL);
2416         else
2417                 mem = vmalloc(size);
2418
2419         if (mem)
2420                 memset(mem, 0, size);
2421         return mem;
2422 }
2423
2424 /*
2425  * At destroying mem_cgroup, references from swap_cgroup can remain.
2426  * (scanning all at force_empty is too costly...)
2427  *
2428  * Instead of clearing all references at force_empty, we remember
2429  * the number of reference from swap_cgroup and free mem_cgroup when
2430  * it goes down to 0.
2431  *
2432  * Removal of cgroup itself succeeds regardless of refs from swap.
2433  */
2434
2435 static void __mem_cgroup_free(struct mem_cgroup *mem)
2436 {
2437         int node;
2438
2439         free_css_id(&mem_cgroup_subsys, &mem->css);
2440
2441         for_each_node_state(node, N_POSSIBLE)
2442                 free_mem_cgroup_per_zone_info(mem, node);
2443
2444         if (mem_cgroup_size() < PAGE_SIZE)
2445                 kfree(mem);
2446         else
2447                 vfree(mem);
2448 }
2449
2450 static void mem_cgroup_get(struct mem_cgroup *mem)
2451 {
2452         atomic_inc(&mem->refcnt);
2453 }
2454
2455 static void mem_cgroup_put(struct mem_cgroup *mem)
2456 {
2457         if (atomic_dec_and_test(&mem->refcnt)) {
2458                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2459                 __mem_cgroup_free(mem);
2460                 if (parent)
2461                         mem_cgroup_put(parent);
2462         }
2463 }
2464
2465 /*
2466  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2467  */
2468 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2469 {
2470         if (!mem->res.parent)
2471                 return NULL;
2472         return mem_cgroup_from_res_counter(mem->res.parent, res);
2473 }
2474
2475 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2476 static void __init enable_swap_cgroup(void)
2477 {
2478         if (!mem_cgroup_disabled() && really_do_swap_account)
2479                 do_swap_account = 1;
2480 }
2481 #else
2482 static void __init enable_swap_cgroup(void)
2483 {
2484 }
2485 #endif
2486
2487 static struct cgroup_subsys_state * __ref
2488 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2489 {
2490         struct mem_cgroup *mem, *parent;
2491         long error = -ENOMEM;
2492         int node;
2493
2494         mem = mem_cgroup_alloc();
2495         if (!mem)
2496                 return ERR_PTR(error);
2497
2498         for_each_node_state(node, N_POSSIBLE)
2499                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2500                         goto free_out;
2501         /* root ? */
2502         if (cont->parent == NULL) {
2503                 enable_swap_cgroup();
2504                 parent = NULL;
2505         } else {
2506                 parent = mem_cgroup_from_cont(cont->parent);
2507                 mem->use_hierarchy = parent->use_hierarchy;
2508         }
2509
2510         if (parent && parent->use_hierarchy) {
2511                 res_counter_init(&mem->res, &parent->res);
2512                 res_counter_init(&mem->memsw, &parent->memsw);
2513                 /*
2514                  * We increment refcnt of the parent to ensure that we can
2515                  * safely access it on res_counter_charge/uncharge.
2516                  * This refcnt will be decremented when freeing this
2517                  * mem_cgroup(see mem_cgroup_put).
2518                  */
2519                 mem_cgroup_get(parent);
2520         } else {
2521                 res_counter_init(&mem->res, NULL);
2522                 res_counter_init(&mem->memsw, NULL);
2523         }
2524         mem->last_scanned_child = 0;
2525         spin_lock_init(&mem->reclaim_param_lock);
2526
2527         if (parent)
2528                 mem->swappiness = get_swappiness(parent);
2529         atomic_set(&mem->refcnt, 1);
2530         return &mem->css;
2531 free_out:
2532         __mem_cgroup_free(mem);
2533         return ERR_PTR(error);
2534 }
2535
2536 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2537                                         struct cgroup *cont)
2538 {
2539         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2540
2541         return mem_cgroup_force_empty(mem, false);
2542 }
2543
2544 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2545                                 struct cgroup *cont)
2546 {
2547         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2548
2549         mem_cgroup_put(mem);
2550 }
2551
2552 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2553                                 struct cgroup *cont)
2554 {
2555         int ret;
2556
2557         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2558                                 ARRAY_SIZE(mem_cgroup_files));
2559
2560         if (!ret)
2561                 ret = register_memsw_files(cont, ss);
2562         return ret;
2563 }
2564
2565 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2566                                 struct cgroup *cont,
2567                                 struct cgroup *old_cont,
2568                                 struct task_struct *p)
2569 {
2570         mutex_lock(&memcg_tasklist);
2571         /*
2572          * FIXME: It's better to move charges of this process from old
2573          * memcg to new memcg. But it's just on TODO-List now.
2574          */
2575         mutex_unlock(&memcg_tasklist);
2576 }
2577
2578 struct cgroup_subsys mem_cgroup_subsys = {
2579         .name = "memory",
2580         .subsys_id = mem_cgroup_subsys_id,
2581         .create = mem_cgroup_create,
2582         .pre_destroy = mem_cgroup_pre_destroy,
2583         .destroy = mem_cgroup_destroy,
2584         .populate = mem_cgroup_populate,
2585         .attach = mem_cgroup_move_task,
2586         .early_init = 0,
2587         .use_id = 1,
2588 };
2589
2590 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2591
2592 static int __init disable_swap_account(char *s)
2593 {
2594         really_do_swap_account = 0;
2595         return 1;
2596 }
2597 __setup("noswapaccount", disable_swap_account);
2598 #endif