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