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