435f08dac8bf6c02a2b71580f8de9dc652fa914f
[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
55 /*
56  * Statistics for memory cgroup.
57  */
58 enum mem_cgroup_stat_index {
59         /*
60          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
61          */
62         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
63         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
64         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
65         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
66
67         MEM_CGROUP_STAT_NSTATS,
68 };
69
70 struct mem_cgroup_stat_cpu {
71         s64 count[MEM_CGROUP_STAT_NSTATS];
72 } ____cacheline_aligned_in_smp;
73
74 struct mem_cgroup_stat {
75         struct mem_cgroup_stat_cpu cpustat[0];
76 };
77
78 /*
79  * For accounting under irq disable, no need for increment preempt count.
80  */
81 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
82                 enum mem_cgroup_stat_index idx, int val)
83 {
84         stat->count[idx] += val;
85 }
86
87 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
88                 enum mem_cgroup_stat_index idx)
89 {
90         int cpu;
91         s64 ret = 0;
92         for_each_possible_cpu(cpu)
93                 ret += stat->cpustat[cpu].count[idx];
94         return ret;
95 }
96
97 /*
98  * per-zone information in memory controller.
99  */
100 struct mem_cgroup_per_zone {
101         /*
102          * spin_lock to protect the per cgroup LRU
103          */
104         struct list_head        lists[NR_LRU_LISTS];
105         unsigned long           count[NR_LRU_LISTS];
106
107         struct zone_reclaim_stat reclaim_stat;
108 };
109 /* Macro for accessing counter */
110 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
111
112 struct mem_cgroup_per_node {
113         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
114 };
115
116 struct mem_cgroup_lru_info {
117         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
118 };
119
120 /*
121  * The memory controller data structure. The memory controller controls both
122  * page cache and RSS per cgroup. We would eventually like to provide
123  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
124  * to help the administrator determine what knobs to tune.
125  *
126  * TODO: Add a water mark for the memory controller. Reclaim will begin when
127  * we hit the water mark. May be even add a low water mark, such that
128  * no reclaim occurs from a cgroup at it's low water mark, this is
129  * a feature that will be implemented much later in the future.
130  */
131 struct mem_cgroup {
132         struct cgroup_subsys_state css;
133         /*
134          * the counter to account for memory usage
135          */
136         struct res_counter res;
137         /*
138          * the counter to account for mem+swap usage.
139          */
140         struct res_counter memsw;
141         /*
142          * Per cgroup active and inactive list, similar to the
143          * per zone LRU lists.
144          */
145         struct mem_cgroup_lru_info info;
146
147         /*
148           protect against reclaim related member.
149         */
150         spinlock_t reclaim_param_lock;
151
152         int     prev_priority;  /* for recording reclaim priority */
153
154         /*
155          * While reclaiming in a hiearchy, we cache the last child we
156          * reclaimed from. Protected by cgroup_lock()
157          */
158         struct mem_cgroup *last_scanned_child;
159         /*
160          * Should the accounting and control be hierarchical, per subtree?
161          */
162         bool use_hierarchy;
163         unsigned long   last_oom_jiffies;
164         int             obsolete;
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 /*
286  * Following LRU functions are allowed to be used without PCG_LOCK.
287  * Operations are called by routine of global LRU independently from memcg.
288  * What we have to take care of here is validness of pc->mem_cgroup.
289  *
290  * Changes to pc->mem_cgroup happens when
291  * 1. charge
292  * 2. moving account
293  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
294  * It is added to LRU before charge.
295  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
296  * When moving account, the page is not on LRU. It's isolated.
297  */
298
299 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
300 {
301         struct page_cgroup *pc;
302         struct mem_cgroup *mem;
303         struct mem_cgroup_per_zone *mz;
304
305         if (mem_cgroup_disabled())
306                 return;
307         pc = lookup_page_cgroup(page);
308         /* can happen while we handle swapcache. */
309         if (list_empty(&pc->lru))
310                 return;
311         mz = page_cgroup_zoneinfo(pc);
312         mem = pc->mem_cgroup;
313         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
314         list_del_init(&pc->lru);
315         return;
316 }
317
318 void mem_cgroup_del_lru(struct page *page)
319 {
320         mem_cgroup_del_lru_list(page, page_lru(page));
321 }
322
323 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
324 {
325         struct mem_cgroup_per_zone *mz;
326         struct page_cgroup *pc;
327
328         if (mem_cgroup_disabled())
329                 return;
330
331         pc = lookup_page_cgroup(page);
332         smp_rmb();
333         /* unused page is not rotated. */
334         if (!PageCgroupUsed(pc))
335                 return;
336         mz = page_cgroup_zoneinfo(pc);
337         list_move(&pc->lru, &mz->lists[lru]);
338 }
339
340 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
341 {
342         struct page_cgroup *pc;
343         struct mem_cgroup_per_zone *mz;
344
345         if (mem_cgroup_disabled())
346                 return;
347         pc = lookup_page_cgroup(page);
348         /* barrier to sync with "charge" */
349         smp_rmb();
350         if (!PageCgroupUsed(pc))
351                 return;
352
353         mz = page_cgroup_zoneinfo(pc);
354         MEM_CGROUP_ZSTAT(mz, lru) += 1;
355         list_add(&pc->lru, &mz->lists[lru]);
356 }
357 /*
358  * To add swapcache into LRU. Be careful to all this function.
359  * zone->lru_lock shouldn't be held and irq must not be disabled.
360  */
361 static void mem_cgroup_lru_fixup(struct page *page)
362 {
363         if (!isolate_lru_page(page))
364                 putback_lru_page(page);
365 }
366
367 void mem_cgroup_move_lists(struct page *page,
368                            enum lru_list from, enum lru_list to)
369 {
370         if (mem_cgroup_disabled())
371                 return;
372         mem_cgroup_del_lru_list(page, from);
373         mem_cgroup_add_lru_list(page, to);
374 }
375
376 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
377 {
378         int ret;
379
380         task_lock(task);
381         ret = task->mm && mm_match_cgroup(task->mm, mem);
382         task_unlock(task);
383         return ret;
384 }
385
386 /*
387  * Calculate mapped_ratio under memory controller. This will be used in
388  * vmscan.c for deteremining we have to reclaim mapped pages.
389  */
390 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
391 {
392         long total, rss;
393
394         /*
395          * usage is recorded in bytes. But, here, we assume the number of
396          * physical pages can be represented by "long" on any arch.
397          */
398         total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
399         rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
400         return (int)((rss * 100L) / total);
401 }
402
403 /*
404  * prev_priority control...this will be used in memory reclaim path.
405  */
406 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
407 {
408         int prev_priority;
409
410         spin_lock(&mem->reclaim_param_lock);
411         prev_priority = mem->prev_priority;
412         spin_unlock(&mem->reclaim_param_lock);
413
414         return prev_priority;
415 }
416
417 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
418 {
419         spin_lock(&mem->reclaim_param_lock);
420         if (priority < mem->prev_priority)
421                 mem->prev_priority = priority;
422         spin_unlock(&mem->reclaim_param_lock);
423 }
424
425 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
426 {
427         spin_lock(&mem->reclaim_param_lock);
428         mem->prev_priority = priority;
429         spin_unlock(&mem->reclaim_param_lock);
430 }
431
432 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
433 {
434         unsigned long active;
435         unsigned long inactive;
436         unsigned long gb;
437         unsigned long inactive_ratio;
438
439         inactive = mem_cgroup_get_all_zonestat(memcg, LRU_INACTIVE_ANON);
440         active = mem_cgroup_get_all_zonestat(memcg, LRU_ACTIVE_ANON);
441
442         gb = (inactive + active) >> (30 - PAGE_SHIFT);
443         if (gb)
444                 inactive_ratio = int_sqrt(10 * gb);
445         else
446                 inactive_ratio = 1;
447
448         if (present_pages) {
449                 present_pages[0] = inactive;
450                 present_pages[1] = active;
451         }
452
453         return inactive_ratio;
454 }
455
456 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
457 {
458         unsigned long active;
459         unsigned long inactive;
460         unsigned long present_pages[2];
461         unsigned long inactive_ratio;
462
463         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
464
465         inactive = present_pages[0];
466         active = present_pages[1];
467
468         if (inactive * inactive_ratio < active)
469                 return 1;
470
471         return 0;
472 }
473
474 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
475                                        struct zone *zone,
476                                        enum lru_list lru)
477 {
478         int nid = zone->zone_pgdat->node_id;
479         int zid = zone_idx(zone);
480         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
481
482         return MEM_CGROUP_ZSTAT(mz, lru);
483 }
484
485 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
486                                                       struct zone *zone)
487 {
488         int nid = zone->zone_pgdat->node_id;
489         int zid = zone_idx(zone);
490         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
491
492         return &mz->reclaim_stat;
493 }
494
495 struct zone_reclaim_stat *
496 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
497 {
498         struct page_cgroup *pc;
499         struct mem_cgroup_per_zone *mz;
500
501         if (mem_cgroup_disabled())
502                 return NULL;
503
504         pc = lookup_page_cgroup(page);
505         mz = page_cgroup_zoneinfo(pc);
506         if (!mz)
507                 return NULL;
508
509         return &mz->reclaim_stat;
510 }
511
512 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
513                                         struct list_head *dst,
514                                         unsigned long *scanned, int order,
515                                         int mode, struct zone *z,
516                                         struct mem_cgroup *mem_cont,
517                                         int active, int file)
518 {
519         unsigned long nr_taken = 0;
520         struct page *page;
521         unsigned long scan;
522         LIST_HEAD(pc_list);
523         struct list_head *src;
524         struct page_cgroup *pc, *tmp;
525         int nid = z->zone_pgdat->node_id;
526         int zid = zone_idx(z);
527         struct mem_cgroup_per_zone *mz;
528         int lru = LRU_FILE * !!file + !!active;
529
530         BUG_ON(!mem_cont);
531         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
532         src = &mz->lists[lru];
533
534         scan = 0;
535         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
536                 if (scan >= nr_to_scan)
537                         break;
538
539                 page = pc->page;
540                 if (unlikely(!PageCgroupUsed(pc)))
541                         continue;
542                 if (unlikely(!PageLRU(page)))
543                         continue;
544
545                 scan++;
546                 if (__isolate_lru_page(page, mode, file) == 0) {
547                         list_move(&page->lru, dst);
548                         nr_taken++;
549                 }
550         }
551
552         *scanned = scan;
553         return nr_taken;
554 }
555
556 #define mem_cgroup_from_res_counter(counter, member)    \
557         container_of(counter, struct mem_cgroup, member)
558
559 /*
560  * This routine finds the DFS walk successor. This routine should be
561  * called with cgroup_mutex held
562  */
563 static struct mem_cgroup *
564 mem_cgroup_get_next_node(struct mem_cgroup *curr, struct mem_cgroup *root_mem)
565 {
566         struct cgroup *cgroup, *curr_cgroup, *root_cgroup;
567
568         curr_cgroup = curr->css.cgroup;
569         root_cgroup = root_mem->css.cgroup;
570
571         if (!list_empty(&curr_cgroup->children)) {
572                 /*
573                  * Walk down to children
574                  */
575                 mem_cgroup_put(curr);
576                 cgroup = list_entry(curr_cgroup->children.next,
577                                                 struct cgroup, sibling);
578                 curr = mem_cgroup_from_cont(cgroup);
579                 mem_cgroup_get(curr);
580                 goto done;
581         }
582
583 visit_parent:
584         if (curr_cgroup == root_cgroup) {
585                 mem_cgroup_put(curr);
586                 curr = root_mem;
587                 mem_cgroup_get(curr);
588                 goto done;
589         }
590
591         /*
592          * Goto next sibling
593          */
594         if (curr_cgroup->sibling.next != &curr_cgroup->parent->children) {
595                 mem_cgroup_put(curr);
596                 cgroup = list_entry(curr_cgroup->sibling.next, struct cgroup,
597                                                 sibling);
598                 curr = mem_cgroup_from_cont(cgroup);
599                 mem_cgroup_get(curr);
600                 goto done;
601         }
602
603         /*
604          * Go up to next parent and next parent's sibling if need be
605          */
606         curr_cgroup = curr_cgroup->parent;
607         goto visit_parent;
608
609 done:
610         root_mem->last_scanned_child = curr;
611         return curr;
612 }
613
614 /*
615  * Visit the first child (need not be the first child as per the ordering
616  * of the cgroup list, since we track last_scanned_child) of @mem and use
617  * that to reclaim free pages from.
618  */
619 static struct mem_cgroup *
620 mem_cgroup_get_first_node(struct mem_cgroup *root_mem)
621 {
622         struct cgroup *cgroup;
623         struct mem_cgroup *ret;
624         bool obsolete = (root_mem->last_scanned_child &&
625                                 root_mem->last_scanned_child->obsolete);
626
627         /*
628          * Scan all children under the mem_cgroup mem
629          */
630         cgroup_lock();
631         if (list_empty(&root_mem->css.cgroup->children)) {
632                 ret = root_mem;
633                 goto done;
634         }
635
636         if (!root_mem->last_scanned_child || obsolete) {
637
638                 if (obsolete)
639                         mem_cgroup_put(root_mem->last_scanned_child);
640
641                 cgroup = list_first_entry(&root_mem->css.cgroup->children,
642                                 struct cgroup, sibling);
643                 ret = mem_cgroup_from_cont(cgroup);
644                 mem_cgroup_get(ret);
645         } else
646                 ret = mem_cgroup_get_next_node(root_mem->last_scanned_child,
647                                                 root_mem);
648
649 done:
650         root_mem->last_scanned_child = ret;
651         cgroup_unlock();
652         return ret;
653 }
654
655 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
656 {
657         if (do_swap_account) {
658                 if (res_counter_check_under_limit(&mem->res) &&
659                         res_counter_check_under_limit(&mem->memsw))
660                         return true;
661         } else
662                 if (res_counter_check_under_limit(&mem->res))
663                         return true;
664         return false;
665 }
666
667 static unsigned int get_swappiness(struct mem_cgroup *memcg)
668 {
669         struct cgroup *cgrp = memcg->css.cgroup;
670         unsigned int swappiness;
671
672         /* root ? */
673         if (cgrp->parent == NULL)
674                 return vm_swappiness;
675
676         spin_lock(&memcg->reclaim_param_lock);
677         swappiness = memcg->swappiness;
678         spin_unlock(&memcg->reclaim_param_lock);
679
680         return swappiness;
681 }
682
683 /*
684  * Dance down the hierarchy if needed to reclaim memory. We remember the
685  * last child we reclaimed from, so that we don't end up penalizing
686  * one child extensively based on its position in the children list.
687  *
688  * root_mem is the original ancestor that we've been reclaim from.
689  */
690 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
691                                                 gfp_t gfp_mask, bool noswap)
692 {
693         struct mem_cgroup *next_mem;
694         int ret = 0;
695
696         /*
697          * Reclaim unconditionally and don't check for return value.
698          * We need to reclaim in the current group and down the tree.
699          * One might think about checking for children before reclaiming,
700          * but there might be left over accounting, even after children
701          * have left.
702          */
703         ret = try_to_free_mem_cgroup_pages(root_mem, gfp_mask, noswap,
704                                            get_swappiness(root_mem));
705         if (mem_cgroup_check_under_limit(root_mem))
706                 return 0;
707         if (!root_mem->use_hierarchy)
708                 return ret;
709
710         next_mem = mem_cgroup_get_first_node(root_mem);
711
712         while (next_mem != root_mem) {
713                 if (next_mem->obsolete) {
714                         mem_cgroup_put(next_mem);
715                         cgroup_lock();
716                         next_mem = mem_cgroup_get_first_node(root_mem);
717                         cgroup_unlock();
718                         continue;
719                 }
720                 ret = try_to_free_mem_cgroup_pages(next_mem, gfp_mask, noswap,
721                                                    get_swappiness(next_mem));
722                 if (mem_cgroup_check_under_limit(root_mem))
723                         return 0;
724                 cgroup_lock();
725                 next_mem = mem_cgroup_get_next_node(next_mem, root_mem);
726                 cgroup_unlock();
727         }
728         return ret;
729 }
730
731 bool mem_cgroup_oom_called(struct task_struct *task)
732 {
733         bool ret = false;
734         struct mem_cgroup *mem;
735         struct mm_struct *mm;
736
737         rcu_read_lock();
738         mm = task->mm;
739         if (!mm)
740                 mm = &init_mm;
741         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
742         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
743                 ret = true;
744         rcu_read_unlock();
745         return ret;
746 }
747 /*
748  * Unlike exported interface, "oom" parameter is added. if oom==true,
749  * oom-killer can be invoked.
750  */
751 static int __mem_cgroup_try_charge(struct mm_struct *mm,
752                         gfp_t gfp_mask, struct mem_cgroup **memcg,
753                         bool oom)
754 {
755         struct mem_cgroup *mem, *mem_over_limit;
756         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
757         struct res_counter *fail_res;
758
759         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
760                 /* Don't account this! */
761                 *memcg = NULL;
762                 return 0;
763         }
764
765         /*
766          * We always charge the cgroup the mm_struct belongs to.
767          * The mm_struct's mem_cgroup changes on task migration if the
768          * thread group leader migrates. It's possible that mm is not
769          * set, if so charge the init_mm (happens for pagecache usage).
770          */
771         if (likely(!*memcg)) {
772                 rcu_read_lock();
773                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
774                 if (unlikely(!mem)) {
775                         rcu_read_unlock();
776                         return 0;
777                 }
778                 /*
779                  * For every charge from the cgroup, increment reference count
780                  */
781                 css_get(&mem->css);
782                 *memcg = mem;
783                 rcu_read_unlock();
784         } else {
785                 mem = *memcg;
786                 css_get(&mem->css);
787         }
788
789         while (1) {
790                 int ret;
791                 bool noswap = false;
792
793                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
794                 if (likely(!ret)) {
795                         if (!do_swap_account)
796                                 break;
797                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
798                                                         &fail_res);
799                         if (likely(!ret))
800                                 break;
801                         /* mem+swap counter fails */
802                         res_counter_uncharge(&mem->res, PAGE_SIZE);
803                         noswap = true;
804                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
805                                                                         memsw);
806                 } else
807                         /* mem counter fails */
808                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
809                                                                         res);
810
811                 if (!(gfp_mask & __GFP_WAIT))
812                         goto nomem;
813
814                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
815                                                         noswap);
816
817                 /*
818                  * try_to_free_mem_cgroup_pages() might not give us a full
819                  * picture of reclaim. Some pages are reclaimed and might be
820                  * moved to swap cache or just unmapped from the cgroup.
821                  * Check the limit again to see if the reclaim reduced the
822                  * current usage of the cgroup before giving up
823                  *
824                  */
825                 if (mem_cgroup_check_under_limit(mem_over_limit))
826                         continue;
827
828                 if (!nr_retries--) {
829                         if (oom) {
830                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
831                                 mem_over_limit->last_oom_jiffies = jiffies;
832                         }
833                         goto nomem;
834                 }
835         }
836         return 0;
837 nomem:
838         css_put(&mem->css);
839         return -ENOMEM;
840 }
841
842 /*
843  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
844  * USED state. If already USED, uncharge and return.
845  */
846
847 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
848                                      struct page_cgroup *pc,
849                                      enum charge_type ctype)
850 {
851         /* try_charge() can return NULL to *memcg, taking care of it. */
852         if (!mem)
853                 return;
854
855         lock_page_cgroup(pc);
856         if (unlikely(PageCgroupUsed(pc))) {
857                 unlock_page_cgroup(pc);
858                 res_counter_uncharge(&mem->res, PAGE_SIZE);
859                 if (do_swap_account)
860                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
861                 css_put(&mem->css);
862                 return;
863         }
864         pc->mem_cgroup = mem;
865         smp_wmb();
866         pc->flags = pcg_default_flags[ctype];
867
868         mem_cgroup_charge_statistics(mem, pc, true);
869
870         unlock_page_cgroup(pc);
871 }
872
873 /**
874  * mem_cgroup_move_account - move account of the page
875  * @pc: page_cgroup of the page.
876  * @from: mem_cgroup which the page is moved from.
877  * @to: mem_cgroup which the page is moved to. @from != @to.
878  *
879  * The caller must confirm following.
880  * - page is not on LRU (isolate_page() is useful.)
881  *
882  * returns 0 at success,
883  * returns -EBUSY when lock is busy or "pc" is unstable.
884  *
885  * This function does "uncharge" from old cgroup but doesn't do "charge" to
886  * new cgroup. It should be done by a caller.
887  */
888
889 static int mem_cgroup_move_account(struct page_cgroup *pc,
890         struct mem_cgroup *from, struct mem_cgroup *to)
891 {
892         struct mem_cgroup_per_zone *from_mz, *to_mz;
893         int nid, zid;
894         int ret = -EBUSY;
895
896         VM_BUG_ON(from == to);
897         VM_BUG_ON(PageLRU(pc->page));
898
899         nid = page_cgroup_nid(pc);
900         zid = page_cgroup_zid(pc);
901         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
902         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
903
904         if (!trylock_page_cgroup(pc))
905                 return ret;
906
907         if (!PageCgroupUsed(pc))
908                 goto out;
909
910         if (pc->mem_cgroup != from)
911                 goto out;
912
913         css_put(&from->css);
914         res_counter_uncharge(&from->res, PAGE_SIZE);
915         mem_cgroup_charge_statistics(from, pc, false);
916         if (do_swap_account)
917                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
918         pc->mem_cgroup = to;
919         mem_cgroup_charge_statistics(to, pc, true);
920         css_get(&to->css);
921         ret = 0;
922 out:
923         unlock_page_cgroup(pc);
924         return ret;
925 }
926
927 /*
928  * move charges to its parent.
929  */
930
931 static int mem_cgroup_move_parent(struct page_cgroup *pc,
932                                   struct mem_cgroup *child,
933                                   gfp_t gfp_mask)
934 {
935         struct page *page = pc->page;
936         struct cgroup *cg = child->css.cgroup;
937         struct cgroup *pcg = cg->parent;
938         struct mem_cgroup *parent;
939         int ret;
940
941         /* Is ROOT ? */
942         if (!pcg)
943                 return -EINVAL;
944
945
946         parent = mem_cgroup_from_cont(pcg);
947
948
949         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
950         if (ret || !parent)
951                 return ret;
952
953         if (!get_page_unless_zero(page))
954                 return -EBUSY;
955
956         ret = isolate_lru_page(page);
957
958         if (ret)
959                 goto cancel;
960
961         ret = mem_cgroup_move_account(pc, child, parent);
962
963         /* drop extra refcnt by try_charge() (move_account increment one) */
964         css_put(&parent->css);
965         putback_lru_page(page);
966         if (!ret) {
967                 put_page(page);
968                 return 0;
969         }
970         /* uncharge if move fails */
971 cancel:
972         res_counter_uncharge(&parent->res, PAGE_SIZE);
973         if (do_swap_account)
974                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
975         put_page(page);
976         return ret;
977 }
978
979 /*
980  * Charge the memory controller for page usage.
981  * Return
982  * 0 if the charge was successful
983  * < 0 if the cgroup is over its limit
984  */
985 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
986                                 gfp_t gfp_mask, enum charge_type ctype,
987                                 struct mem_cgroup *memcg)
988 {
989         struct mem_cgroup *mem;
990         struct page_cgroup *pc;
991         int ret;
992
993         pc = lookup_page_cgroup(page);
994         /* can happen at boot */
995         if (unlikely(!pc))
996                 return 0;
997         prefetchw(pc);
998
999         mem = memcg;
1000         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1001         if (ret || !mem)
1002                 return ret;
1003
1004         __mem_cgroup_commit_charge(mem, pc, ctype);
1005         return 0;
1006 }
1007
1008 int mem_cgroup_newpage_charge(struct page *page,
1009                               struct mm_struct *mm, gfp_t gfp_mask)
1010 {
1011         if (mem_cgroup_disabled())
1012                 return 0;
1013         if (PageCompound(page))
1014                 return 0;
1015         /*
1016          * If already mapped, we don't have to account.
1017          * If page cache, page->mapping has address_space.
1018          * But page->mapping may have out-of-use anon_vma pointer,
1019          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1020          * is NULL.
1021          */
1022         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1023                 return 0;
1024         if (unlikely(!mm))
1025                 mm = &init_mm;
1026         return mem_cgroup_charge_common(page, mm, gfp_mask,
1027                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1028 }
1029
1030 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1031                                 gfp_t gfp_mask)
1032 {
1033         if (mem_cgroup_disabled())
1034                 return 0;
1035         if (PageCompound(page))
1036                 return 0;
1037         /*
1038          * Corner case handling. This is called from add_to_page_cache()
1039          * in usual. But some FS (shmem) precharges this page before calling it
1040          * and call add_to_page_cache() with GFP_NOWAIT.
1041          *
1042          * For GFP_NOWAIT case, the page may be pre-charged before calling
1043          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1044          * charge twice. (It works but has to pay a bit larger cost.)
1045          */
1046         if (!(gfp_mask & __GFP_WAIT)) {
1047                 struct page_cgroup *pc;
1048
1049
1050                 pc = lookup_page_cgroup(page);
1051                 if (!pc)
1052                         return 0;
1053                 lock_page_cgroup(pc);
1054                 if (PageCgroupUsed(pc)) {
1055                         unlock_page_cgroup(pc);
1056                         return 0;
1057                 }
1058                 unlock_page_cgroup(pc);
1059         }
1060
1061         if (unlikely(!mm))
1062                 mm = &init_mm;
1063
1064         if (page_is_file_cache(page))
1065                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1066                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1067         else
1068                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1069                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, NULL);
1070 }
1071
1072 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1073                                  struct page *page,
1074                                  gfp_t mask, struct mem_cgroup **ptr)
1075 {
1076         struct mem_cgroup *mem;
1077         swp_entry_t     ent;
1078
1079         if (mem_cgroup_disabled())
1080                 return 0;
1081
1082         if (!do_swap_account)
1083                 goto charge_cur_mm;
1084
1085         /*
1086          * A racing thread's fault, or swapoff, may have already updated
1087          * the pte, and even removed page from swap cache: return success
1088          * to go on to do_swap_page()'s pte_same() test, which should fail.
1089          */
1090         if (!PageSwapCache(page))
1091                 return 0;
1092
1093         ent.val = page_private(page);
1094
1095         mem = lookup_swap_cgroup(ent);
1096         if (!mem || mem->obsolete)
1097                 goto charge_cur_mm;
1098         *ptr = mem;
1099         return __mem_cgroup_try_charge(NULL, mask, ptr, true);
1100 charge_cur_mm:
1101         if (unlikely(!mm))
1102                 mm = &init_mm;
1103         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1104 }
1105
1106 #ifdef CONFIG_SWAP
1107
1108 int mem_cgroup_cache_charge_swapin(struct page *page,
1109                         struct mm_struct *mm, gfp_t mask, bool locked)
1110 {
1111         int ret = 0;
1112
1113         if (mem_cgroup_disabled())
1114                 return 0;
1115         if (unlikely(!mm))
1116                 mm = &init_mm;
1117         if (!locked)
1118                 lock_page(page);
1119         /*
1120          * If not locked, the page can be dropped from SwapCache until
1121          * we reach here.
1122          */
1123         if (PageSwapCache(page)) {
1124                 struct mem_cgroup *mem = NULL;
1125                 swp_entry_t ent;
1126
1127                 ent.val = page_private(page);
1128                 if (do_swap_account) {
1129                         mem = lookup_swap_cgroup(ent);
1130                         if (mem && mem->obsolete)
1131                                 mem = NULL;
1132                         if (mem)
1133                                 mm = NULL;
1134                 }
1135                 ret = mem_cgroup_charge_common(page, mm, mask,
1136                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1137
1138                 if (!ret && do_swap_account) {
1139                         /* avoid double counting */
1140                         mem = swap_cgroup_record(ent, NULL);
1141                         if (mem) {
1142                                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1143                                 mem_cgroup_put(mem);
1144                         }
1145                 }
1146         }
1147         if (!locked)
1148                 unlock_page(page);
1149         /* add this page(page_cgroup) to the LRU we want. */
1150         mem_cgroup_lru_fixup(page);
1151
1152         return ret;
1153 }
1154 #endif
1155
1156 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1157 {
1158         struct page_cgroup *pc;
1159
1160         if (mem_cgroup_disabled())
1161                 return;
1162         if (!ptr)
1163                 return;
1164         pc = lookup_page_cgroup(page);
1165         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1166         /*
1167          * Now swap is on-memory. This means this page may be
1168          * counted both as mem and swap....double count.
1169          * Fix it by uncharging from memsw. This SwapCache is stable
1170          * because we're still under lock_page().
1171          */
1172         if (do_swap_account) {
1173                 swp_entry_t ent = {.val = page_private(page)};
1174                 struct mem_cgroup *memcg;
1175                 memcg = swap_cgroup_record(ent, NULL);
1176                 if (memcg) {
1177                         /* If memcg is obsolete, memcg can be != ptr */
1178                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1179                         mem_cgroup_put(memcg);
1180                 }
1181
1182         }
1183         /* add this page(page_cgroup) to the LRU we want. */
1184         mem_cgroup_lru_fixup(page);
1185 }
1186
1187 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1188 {
1189         if (mem_cgroup_disabled())
1190                 return;
1191         if (!mem)
1192                 return;
1193         res_counter_uncharge(&mem->res, PAGE_SIZE);
1194         if (do_swap_account)
1195                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1196         css_put(&mem->css);
1197 }
1198
1199
1200 /*
1201  * uncharge if !page_mapped(page)
1202  */
1203 static struct mem_cgroup *
1204 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1205 {
1206         struct page_cgroup *pc;
1207         struct mem_cgroup *mem = NULL;
1208         struct mem_cgroup_per_zone *mz;
1209
1210         if (mem_cgroup_disabled())
1211                 return NULL;
1212
1213         if (PageSwapCache(page))
1214                 return NULL;
1215
1216         /*
1217          * Check if our page_cgroup is valid
1218          */
1219         pc = lookup_page_cgroup(page);
1220         if (unlikely(!pc || !PageCgroupUsed(pc)))
1221                 return NULL;
1222
1223         lock_page_cgroup(pc);
1224
1225         mem = pc->mem_cgroup;
1226
1227         if (!PageCgroupUsed(pc))
1228                 goto unlock_out;
1229
1230         switch (ctype) {
1231         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1232                 if (page_mapped(page))
1233                         goto unlock_out;
1234                 break;
1235         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1236                 if (!PageAnon(page)) {  /* Shared memory */
1237                         if (page->mapping && !page_is_file_cache(page))
1238                                 goto unlock_out;
1239                 } else if (page_mapped(page)) /* Anon */
1240                                 goto unlock_out;
1241                 break;
1242         default:
1243                 break;
1244         }
1245
1246         res_counter_uncharge(&mem->res, PAGE_SIZE);
1247         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1248                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1249
1250         mem_cgroup_charge_statistics(mem, pc, false);
1251         ClearPageCgroupUsed(pc);
1252
1253         mz = page_cgroup_zoneinfo(pc);
1254         unlock_page_cgroup(pc);
1255
1256         /* at swapout, this memcg will be accessed to record to swap */
1257         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1258                 css_put(&mem->css);
1259
1260         return mem;
1261
1262 unlock_out:
1263         unlock_page_cgroup(pc);
1264         return NULL;
1265 }
1266
1267 void mem_cgroup_uncharge_page(struct page *page)
1268 {
1269         /* early check. */
1270         if (page_mapped(page))
1271                 return;
1272         if (page->mapping && !PageAnon(page))
1273                 return;
1274         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1275 }
1276
1277 void mem_cgroup_uncharge_cache_page(struct page *page)
1278 {
1279         VM_BUG_ON(page_mapped(page));
1280         VM_BUG_ON(page->mapping);
1281         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1282 }
1283
1284 /*
1285  * called from __delete_from_swap_cache() and drop "page" account.
1286  * memcg information is recorded to swap_cgroup of "ent"
1287  */
1288 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1289 {
1290         struct mem_cgroup *memcg;
1291
1292         memcg = __mem_cgroup_uncharge_common(page,
1293                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1294         /* record memcg information */
1295         if (do_swap_account && memcg) {
1296                 swap_cgroup_record(ent, memcg);
1297                 mem_cgroup_get(memcg);
1298         }
1299         if (memcg)
1300                 css_put(&memcg->css);
1301 }
1302
1303 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1304 /*
1305  * called from swap_entry_free(). remove record in swap_cgroup and
1306  * uncharge "memsw" account.
1307  */
1308 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1309 {
1310         struct mem_cgroup *memcg;
1311
1312         if (!do_swap_account)
1313                 return;
1314
1315         memcg = swap_cgroup_record(ent, NULL);
1316         if (memcg) {
1317                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1318                 mem_cgroup_put(memcg);
1319         }
1320 }
1321 #endif
1322
1323 /*
1324  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1325  * page belongs to.
1326  */
1327 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1328 {
1329         struct page_cgroup *pc;
1330         struct mem_cgroup *mem = NULL;
1331         int ret = 0;
1332
1333         if (mem_cgroup_disabled())
1334                 return 0;
1335
1336         pc = lookup_page_cgroup(page);
1337         lock_page_cgroup(pc);
1338         if (PageCgroupUsed(pc)) {
1339                 mem = pc->mem_cgroup;
1340                 css_get(&mem->css);
1341         }
1342         unlock_page_cgroup(pc);
1343
1344         if (mem) {
1345                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1346                 css_put(&mem->css);
1347         }
1348         *ptr = mem;
1349         return ret;
1350 }
1351
1352 /* remove redundant charge if migration failed*/
1353 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1354                 struct page *oldpage, struct page *newpage)
1355 {
1356         struct page *target, *unused;
1357         struct page_cgroup *pc;
1358         enum charge_type ctype;
1359
1360         if (!mem)
1361                 return;
1362
1363         /* at migration success, oldpage->mapping is NULL. */
1364         if (oldpage->mapping) {
1365                 target = oldpage;
1366                 unused = NULL;
1367         } else {
1368                 target = newpage;
1369                 unused = oldpage;
1370         }
1371
1372         if (PageAnon(target))
1373                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1374         else if (page_is_file_cache(target))
1375                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1376         else
1377                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1378
1379         /* unused page is not on radix-tree now. */
1380         if (unused)
1381                 __mem_cgroup_uncharge_common(unused, ctype);
1382
1383         pc = lookup_page_cgroup(target);
1384         /*
1385          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1386          * So, double-counting is effectively avoided.
1387          */
1388         __mem_cgroup_commit_charge(mem, pc, ctype);
1389
1390         /*
1391          * Both of oldpage and newpage are still under lock_page().
1392          * Then, we don't have to care about race in radix-tree.
1393          * But we have to be careful that this page is unmapped or not.
1394          *
1395          * There is a case for !page_mapped(). At the start of
1396          * migration, oldpage was mapped. But now, it's zapped.
1397          * But we know *target* page is not freed/reused under us.
1398          * mem_cgroup_uncharge_page() does all necessary checks.
1399          */
1400         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1401                 mem_cgroup_uncharge_page(target);
1402 }
1403
1404 /*
1405  * A call to try to shrink memory usage under specified resource controller.
1406  * This is typically used for page reclaiming for shmem for reducing side
1407  * effect of page allocation from shmem, which is used by some mem_cgroup.
1408  */
1409 int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
1410 {
1411         struct mem_cgroup *mem;
1412         int progress = 0;
1413         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1414
1415         if (mem_cgroup_disabled())
1416                 return 0;
1417         if (!mm)
1418                 return 0;
1419
1420         rcu_read_lock();
1421         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1422         if (unlikely(!mem)) {
1423                 rcu_read_unlock();
1424                 return 0;
1425         }
1426         css_get(&mem->css);
1427         rcu_read_unlock();
1428
1429         do {
1430                 progress = try_to_free_mem_cgroup_pages(mem, gfp_mask, true,
1431                                                         get_swappiness(mem));
1432                 progress += mem_cgroup_check_under_limit(mem);
1433         } while (!progress && --retry);
1434
1435         css_put(&mem->css);
1436         if (!retry)
1437                 return -ENOMEM;
1438         return 0;
1439 }
1440
1441 static DEFINE_MUTEX(set_limit_mutex);
1442
1443 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1444                                 unsigned long long val)
1445 {
1446
1447         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1448         int progress;
1449         u64 memswlimit;
1450         int ret = 0;
1451
1452         while (retry_count) {
1453                 if (signal_pending(current)) {
1454                         ret = -EINTR;
1455                         break;
1456                 }
1457                 /*
1458                  * Rather than hide all in some function, I do this in
1459                  * open coded manner. You see what this really does.
1460                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1461                  */
1462                 mutex_lock(&set_limit_mutex);
1463                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1464                 if (memswlimit < val) {
1465                         ret = -EINVAL;
1466                         mutex_unlock(&set_limit_mutex);
1467                         break;
1468                 }
1469                 ret = res_counter_set_limit(&memcg->res, val);
1470                 mutex_unlock(&set_limit_mutex);
1471
1472                 if (!ret)
1473                         break;
1474
1475                 progress = try_to_free_mem_cgroup_pages(memcg,
1476                                                         GFP_KERNEL,
1477                                                         false,
1478                                                         get_swappiness(memcg));
1479                 if (!progress)                  retry_count--;
1480         }
1481
1482         return ret;
1483 }
1484
1485 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1486                                 unsigned long long val)
1487 {
1488         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1489         u64 memlimit, oldusage, curusage;
1490         int ret;
1491
1492         if (!do_swap_account)
1493                 return -EINVAL;
1494
1495         while (retry_count) {
1496                 if (signal_pending(current)) {
1497                         ret = -EINTR;
1498                         break;
1499                 }
1500                 /*
1501                  * Rather than hide all in some function, I do this in
1502                  * open coded manner. You see what this really does.
1503                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1504                  */
1505                 mutex_lock(&set_limit_mutex);
1506                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1507                 if (memlimit > val) {
1508                         ret = -EINVAL;
1509                         mutex_unlock(&set_limit_mutex);
1510                         break;
1511                 }
1512                 ret = res_counter_set_limit(&memcg->memsw, val);
1513                 mutex_unlock(&set_limit_mutex);
1514
1515                 if (!ret)
1516                         break;
1517
1518                 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1519                 try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, true,
1520                                              get_swappiness(memcg));
1521                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1522                 if (curusage >= oldusage)
1523                         retry_count--;
1524         }
1525         return ret;
1526 }
1527
1528 /*
1529  * This routine traverse page_cgroup in given list and drop them all.
1530  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1531  */
1532 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1533                                 int node, int zid, enum lru_list lru)
1534 {
1535         struct zone *zone;
1536         struct mem_cgroup_per_zone *mz;
1537         struct page_cgroup *pc, *busy;
1538         unsigned long flags, loop;
1539         struct list_head *list;
1540         int ret = 0;
1541
1542         zone = &NODE_DATA(node)->node_zones[zid];
1543         mz = mem_cgroup_zoneinfo(mem, node, zid);
1544         list = &mz->lists[lru];
1545
1546         loop = MEM_CGROUP_ZSTAT(mz, lru);
1547         /* give some margin against EBUSY etc...*/
1548         loop += 256;
1549         busy = NULL;
1550         while (loop--) {
1551                 ret = 0;
1552                 spin_lock_irqsave(&zone->lru_lock, flags);
1553                 if (list_empty(list)) {
1554                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1555                         break;
1556                 }
1557                 pc = list_entry(list->prev, struct page_cgroup, lru);
1558                 if (busy == pc) {
1559                         list_move(&pc->lru, list);
1560                         busy = 0;
1561                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1562                         continue;
1563                 }
1564                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1565
1566                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1567                 if (ret == -ENOMEM)
1568                         break;
1569
1570                 if (ret == -EBUSY || ret == -EINVAL) {
1571                         /* found lock contention or "pc" is obsolete. */
1572                         busy = pc;
1573                         cond_resched();
1574                 } else
1575                         busy = NULL;
1576         }
1577
1578         if (!ret && !list_empty(list))
1579                 return -EBUSY;
1580         return ret;
1581 }
1582
1583 /*
1584  * make mem_cgroup's charge to be 0 if there is no task.
1585  * This enables deleting this mem_cgroup.
1586  */
1587 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1588 {
1589         int ret;
1590         int node, zid, shrink;
1591         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1592         struct cgroup *cgrp = mem->css.cgroup;
1593
1594         css_get(&mem->css);
1595
1596         shrink = 0;
1597         /* should free all ? */
1598         if (free_all)
1599                 goto try_to_free;
1600 move_account:
1601         while (mem->res.usage > 0) {
1602                 ret = -EBUSY;
1603                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1604                         goto out;
1605                 ret = -EINTR;
1606                 if (signal_pending(current))
1607                         goto out;
1608                 /* This is for making all *used* pages to be on LRU. */
1609                 lru_add_drain_all();
1610                 ret = 0;
1611                 for_each_node_state(node, N_POSSIBLE) {
1612                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1613                                 enum lru_list l;
1614                                 for_each_lru(l) {
1615                                         ret = mem_cgroup_force_empty_list(mem,
1616                                                         node, zid, l);
1617                                         if (ret)
1618                                                 break;
1619                                 }
1620                         }
1621                         if (ret)
1622                                 break;
1623                 }
1624                 /* it seems parent cgroup doesn't have enough mem */
1625                 if (ret == -ENOMEM)
1626                         goto try_to_free;
1627                 cond_resched();
1628         }
1629         ret = 0;
1630 out:
1631         css_put(&mem->css);
1632         return ret;
1633
1634 try_to_free:
1635         /* returns EBUSY if there is a task or if we come here twice. */
1636         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1637                 ret = -EBUSY;
1638                 goto out;
1639         }
1640         /* we call try-to-free pages for make this cgroup empty */
1641         lru_add_drain_all();
1642         /* try to free all pages in this cgroup */
1643         shrink = 1;
1644         while (nr_retries && mem->res.usage > 0) {
1645                 int progress;
1646
1647                 if (signal_pending(current)) {
1648                         ret = -EINTR;
1649                         goto out;
1650                 }
1651                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1652                                                 false, get_swappiness(mem));
1653                 if (!progress) {
1654                         nr_retries--;
1655                         /* maybe some writeback is necessary */
1656                         congestion_wait(WRITE, HZ/10);
1657                 }
1658
1659         }
1660         lru_add_drain();
1661         /* try move_account...there may be some *locked* pages. */
1662         if (mem->res.usage)
1663                 goto move_account;
1664         ret = 0;
1665         goto out;
1666 }
1667
1668 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1669 {
1670         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1671 }
1672
1673
1674 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1675 {
1676         return mem_cgroup_from_cont(cont)->use_hierarchy;
1677 }
1678
1679 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1680                                         u64 val)
1681 {
1682         int retval = 0;
1683         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1684         struct cgroup *parent = cont->parent;
1685         struct mem_cgroup *parent_mem = NULL;
1686
1687         if (parent)
1688                 parent_mem = mem_cgroup_from_cont(parent);
1689
1690         cgroup_lock();
1691         /*
1692          * If parent's use_hiearchy is set, we can't make any modifications
1693          * in the child subtrees. If it is unset, then the change can
1694          * occur, provided the current cgroup has no children.
1695          *
1696          * For the root cgroup, parent_mem is NULL, we allow value to be
1697          * set if there are no children.
1698          */
1699         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1700                                 (val == 1 || val == 0)) {
1701                 if (list_empty(&cont->children))
1702                         mem->use_hierarchy = val;
1703                 else
1704                         retval = -EBUSY;
1705         } else
1706                 retval = -EINVAL;
1707         cgroup_unlock();
1708
1709         return retval;
1710 }
1711
1712 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1713 {
1714         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1715         u64 val = 0;
1716         int type, name;
1717
1718         type = MEMFILE_TYPE(cft->private);
1719         name = MEMFILE_ATTR(cft->private);
1720         switch (type) {
1721         case _MEM:
1722                 val = res_counter_read_u64(&mem->res, name);
1723                 break;
1724         case _MEMSWAP:
1725                 if (do_swap_account)
1726                         val = res_counter_read_u64(&mem->memsw, name);
1727                 break;
1728         default:
1729                 BUG();
1730                 break;
1731         }
1732         return val;
1733 }
1734 /*
1735  * The user of this function is...
1736  * RES_LIMIT.
1737  */
1738 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1739                             const char *buffer)
1740 {
1741         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1742         int type, name;
1743         unsigned long long val;
1744         int ret;
1745
1746         type = MEMFILE_TYPE(cft->private);
1747         name = MEMFILE_ATTR(cft->private);
1748         switch (name) {
1749         case RES_LIMIT:
1750                 /* This function does all necessary parse...reuse it */
1751                 ret = res_counter_memparse_write_strategy(buffer, &val);
1752                 if (ret)
1753                         break;
1754                 if (type == _MEM)
1755                         ret = mem_cgroup_resize_limit(memcg, val);
1756                 else
1757                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1758                 break;
1759         default:
1760                 ret = -EINVAL; /* should be BUG() ? */
1761                 break;
1762         }
1763         return ret;
1764 }
1765
1766 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1767                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1768 {
1769         struct cgroup *cgroup;
1770         unsigned long long min_limit, min_memsw_limit, tmp;
1771
1772         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1773         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1774         cgroup = memcg->css.cgroup;
1775         if (!memcg->use_hierarchy)
1776                 goto out;
1777
1778         while (cgroup->parent) {
1779                 cgroup = cgroup->parent;
1780                 memcg = mem_cgroup_from_cont(cgroup);
1781                 if (!memcg->use_hierarchy)
1782                         break;
1783                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
1784                 min_limit = min(min_limit, tmp);
1785                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1786                 min_memsw_limit = min(min_memsw_limit, tmp);
1787         }
1788 out:
1789         *mem_limit = min_limit;
1790         *memsw_limit = min_memsw_limit;
1791         return;
1792 }
1793
1794 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1795 {
1796         struct mem_cgroup *mem;
1797         int type, name;
1798
1799         mem = mem_cgroup_from_cont(cont);
1800         type = MEMFILE_TYPE(event);
1801         name = MEMFILE_ATTR(event);
1802         switch (name) {
1803         case RES_MAX_USAGE:
1804                 if (type == _MEM)
1805                         res_counter_reset_max(&mem->res);
1806                 else
1807                         res_counter_reset_max(&mem->memsw);
1808                 break;
1809         case RES_FAILCNT:
1810                 if (type == _MEM)
1811                         res_counter_reset_failcnt(&mem->res);
1812                 else
1813                         res_counter_reset_failcnt(&mem->memsw);
1814                 break;
1815         }
1816         return 0;
1817 }
1818
1819 static const struct mem_cgroup_stat_desc {
1820         const char *msg;
1821         u64 unit;
1822 } mem_cgroup_stat_desc[] = {
1823         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
1824         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
1825         [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
1826         [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
1827 };
1828
1829 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
1830                                  struct cgroup_map_cb *cb)
1831 {
1832         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
1833         struct mem_cgroup_stat *stat = &mem_cont->stat;
1834         int i;
1835
1836         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
1837                 s64 val;
1838
1839                 val = mem_cgroup_read_stat(stat, i);
1840                 val *= mem_cgroup_stat_desc[i].unit;
1841                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
1842         }
1843         /* showing # of active pages */
1844         {
1845                 unsigned long active_anon, inactive_anon;
1846                 unsigned long active_file, inactive_file;
1847                 unsigned long unevictable;
1848
1849                 inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
1850                                                 LRU_INACTIVE_ANON);
1851                 active_anon = mem_cgroup_get_all_zonestat(mem_cont,
1852                                                 LRU_ACTIVE_ANON);
1853                 inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
1854                                                 LRU_INACTIVE_FILE);
1855                 active_file = mem_cgroup_get_all_zonestat(mem_cont,
1856                                                 LRU_ACTIVE_FILE);
1857                 unevictable = mem_cgroup_get_all_zonestat(mem_cont,
1858                                                         LRU_UNEVICTABLE);
1859
1860                 cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
1861                 cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
1862                 cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
1863                 cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
1864                 cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
1865
1866         }
1867         {
1868                 unsigned long long limit, memsw_limit;
1869                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
1870                 cb->fill(cb, "hierarchical_memory_limit", limit);
1871                 if (do_swap_account)
1872                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
1873         }
1874
1875 #ifdef CONFIG_DEBUG_VM
1876         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
1877
1878         {
1879                 int nid, zid;
1880                 struct mem_cgroup_per_zone *mz;
1881                 unsigned long recent_rotated[2] = {0, 0};
1882                 unsigned long recent_scanned[2] = {0, 0};
1883
1884                 for_each_online_node(nid)
1885                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1886                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1887
1888                                 recent_rotated[0] +=
1889                                         mz->reclaim_stat.recent_rotated[0];
1890                                 recent_rotated[1] +=
1891                                         mz->reclaim_stat.recent_rotated[1];
1892                                 recent_scanned[0] +=
1893                                         mz->reclaim_stat.recent_scanned[0];
1894                                 recent_scanned[1] +=
1895                                         mz->reclaim_stat.recent_scanned[1];
1896                         }
1897                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
1898                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
1899                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
1900                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
1901         }
1902 #endif
1903
1904         return 0;
1905 }
1906
1907 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
1908 {
1909         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1910
1911         return get_swappiness(memcg);
1912 }
1913
1914 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
1915                                        u64 val)
1916 {
1917         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
1918         struct mem_cgroup *parent;
1919         if (val > 100)
1920                 return -EINVAL;
1921
1922         if (cgrp->parent == NULL)
1923                 return -EINVAL;
1924
1925         parent = mem_cgroup_from_cont(cgrp->parent);
1926         /* If under hierarchy, only empty-root can set this value */
1927         if ((parent->use_hierarchy) ||
1928             (memcg->use_hierarchy && !list_empty(&cgrp->children)))
1929                 return -EINVAL;
1930
1931         spin_lock(&memcg->reclaim_param_lock);
1932         memcg->swappiness = val;
1933         spin_unlock(&memcg->reclaim_param_lock);
1934
1935         return 0;
1936 }
1937
1938
1939 static struct cftype mem_cgroup_files[] = {
1940         {
1941                 .name = "usage_in_bytes",
1942                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
1943                 .read_u64 = mem_cgroup_read,
1944         },
1945         {
1946                 .name = "max_usage_in_bytes",
1947                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
1948                 .trigger = mem_cgroup_reset,
1949                 .read_u64 = mem_cgroup_read,
1950         },
1951         {
1952                 .name = "limit_in_bytes",
1953                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
1954                 .write_string = mem_cgroup_write,
1955                 .read_u64 = mem_cgroup_read,
1956         },
1957         {
1958                 .name = "failcnt",
1959                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
1960                 .trigger = mem_cgroup_reset,
1961                 .read_u64 = mem_cgroup_read,
1962         },
1963         {
1964                 .name = "stat",
1965                 .read_map = mem_control_stat_show,
1966         },
1967         {
1968                 .name = "force_empty",
1969                 .trigger = mem_cgroup_force_empty_write,
1970         },
1971         {
1972                 .name = "use_hierarchy",
1973                 .write_u64 = mem_cgroup_hierarchy_write,
1974                 .read_u64 = mem_cgroup_hierarchy_read,
1975         },
1976         {
1977                 .name = "swappiness",
1978                 .read_u64 = mem_cgroup_swappiness_read,
1979                 .write_u64 = mem_cgroup_swappiness_write,
1980         },
1981 };
1982
1983 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1984 static struct cftype memsw_cgroup_files[] = {
1985         {
1986                 .name = "memsw.usage_in_bytes",
1987                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
1988                 .read_u64 = mem_cgroup_read,
1989         },
1990         {
1991                 .name = "memsw.max_usage_in_bytes",
1992                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
1993                 .trigger = mem_cgroup_reset,
1994                 .read_u64 = mem_cgroup_read,
1995         },
1996         {
1997                 .name = "memsw.limit_in_bytes",
1998                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
1999                 .write_string = mem_cgroup_write,
2000                 .read_u64 = mem_cgroup_read,
2001         },
2002         {
2003                 .name = "memsw.failcnt",
2004                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2005                 .trigger = mem_cgroup_reset,
2006                 .read_u64 = mem_cgroup_read,
2007         },
2008 };
2009
2010 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2011 {
2012         if (!do_swap_account)
2013                 return 0;
2014         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2015                                 ARRAY_SIZE(memsw_cgroup_files));
2016 };
2017 #else
2018 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2019 {
2020         return 0;
2021 }
2022 #endif
2023
2024 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2025 {
2026         struct mem_cgroup_per_node *pn;
2027         struct mem_cgroup_per_zone *mz;
2028         enum lru_list l;
2029         int zone, tmp = node;
2030         /*
2031          * This routine is called against possible nodes.
2032          * But it's BUG to call kmalloc() against offline node.
2033          *
2034          * TODO: this routine can waste much memory for nodes which will
2035          *       never be onlined. It's better to use memory hotplug callback
2036          *       function.
2037          */
2038         if (!node_state(node, N_NORMAL_MEMORY))
2039                 tmp = -1;
2040         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2041         if (!pn)
2042                 return 1;
2043
2044         mem->info.nodeinfo[node] = pn;
2045         memset(pn, 0, sizeof(*pn));
2046
2047         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2048                 mz = &pn->zoneinfo[zone];
2049                 for_each_lru(l)
2050                         INIT_LIST_HEAD(&mz->lists[l]);
2051         }
2052         return 0;
2053 }
2054
2055 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2056 {
2057         kfree(mem->info.nodeinfo[node]);
2058 }
2059
2060 static int mem_cgroup_size(void)
2061 {
2062         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2063         return sizeof(struct mem_cgroup) + cpustat_size;
2064 }
2065
2066 static struct mem_cgroup *mem_cgroup_alloc(void)
2067 {
2068         struct mem_cgroup *mem;
2069         int size = mem_cgroup_size();
2070
2071         if (size < PAGE_SIZE)
2072                 mem = kmalloc(size, GFP_KERNEL);
2073         else
2074                 mem = vmalloc(size);
2075
2076         if (mem)
2077                 memset(mem, 0, size);
2078         return mem;
2079 }
2080
2081 /*
2082  * At destroying mem_cgroup, references from swap_cgroup can remain.
2083  * (scanning all at force_empty is too costly...)
2084  *
2085  * Instead of clearing all references at force_empty, we remember
2086  * the number of reference from swap_cgroup and free mem_cgroup when
2087  * it goes down to 0.
2088  *
2089  * When mem_cgroup is destroyed, mem->obsolete will be set to 0 and
2090  * entry which points to this memcg will be ignore at swapin.
2091  *
2092  * Removal of cgroup itself succeeds regardless of refs from swap.
2093  */
2094
2095 static void mem_cgroup_free(struct mem_cgroup *mem)
2096 {
2097         int node;
2098
2099         if (atomic_read(&mem->refcnt) > 0)
2100                 return;
2101
2102
2103         for_each_node_state(node, N_POSSIBLE)
2104                 free_mem_cgroup_per_zone_info(mem, node);
2105
2106         if (mem_cgroup_size() < PAGE_SIZE)
2107                 kfree(mem);
2108         else
2109                 vfree(mem);
2110 }
2111
2112 static void mem_cgroup_get(struct mem_cgroup *mem)
2113 {
2114         atomic_inc(&mem->refcnt);
2115 }
2116
2117 static void mem_cgroup_put(struct mem_cgroup *mem)
2118 {
2119         if (atomic_dec_and_test(&mem->refcnt)) {
2120                 if (!mem->obsolete)
2121                         return;
2122                 mem_cgroup_free(mem);
2123         }
2124 }
2125
2126
2127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2128 static void __init enable_swap_cgroup(void)
2129 {
2130         if (!mem_cgroup_disabled() && really_do_swap_account)
2131                 do_swap_account = 1;
2132 }
2133 #else
2134 static void __init enable_swap_cgroup(void)
2135 {
2136 }
2137 #endif
2138
2139 static struct cgroup_subsys_state *
2140 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2141 {
2142         struct mem_cgroup *mem, *parent;
2143         int node;
2144
2145         mem = mem_cgroup_alloc();
2146         if (!mem)
2147                 return ERR_PTR(-ENOMEM);
2148
2149         for_each_node_state(node, N_POSSIBLE)
2150                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2151                         goto free_out;
2152         /* root ? */
2153         if (cont->parent == NULL) {
2154                 enable_swap_cgroup();
2155                 parent = NULL;
2156         } else {
2157                 parent = mem_cgroup_from_cont(cont->parent);
2158                 mem->use_hierarchy = parent->use_hierarchy;
2159         }
2160
2161         if (parent && parent->use_hierarchy) {
2162                 res_counter_init(&mem->res, &parent->res);
2163                 res_counter_init(&mem->memsw, &parent->memsw);
2164         } else {
2165                 res_counter_init(&mem->res, NULL);
2166                 res_counter_init(&mem->memsw, NULL);
2167         }
2168         mem->last_scanned_child = NULL;
2169         spin_lock_init(&mem->reclaim_param_lock);
2170
2171         if (parent)
2172                 mem->swappiness = get_swappiness(parent);
2173
2174         return &mem->css;
2175 free_out:
2176         for_each_node_state(node, N_POSSIBLE)
2177                 free_mem_cgroup_per_zone_info(mem, node);
2178         mem_cgroup_free(mem);
2179         return ERR_PTR(-ENOMEM);
2180 }
2181
2182 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2183                                         struct cgroup *cont)
2184 {
2185         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2186         mem->obsolete = 1;
2187         mem_cgroup_force_empty(mem, false);
2188 }
2189
2190 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2191                                 struct cgroup *cont)
2192 {
2193         mem_cgroup_free(mem_cgroup_from_cont(cont));
2194 }
2195
2196 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2197                                 struct cgroup *cont)
2198 {
2199         int ret;
2200
2201         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2202                                 ARRAY_SIZE(mem_cgroup_files));
2203
2204         if (!ret)
2205                 ret = register_memsw_files(cont, ss);
2206         return ret;
2207 }
2208
2209 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2210                                 struct cgroup *cont,
2211                                 struct cgroup *old_cont,
2212                                 struct task_struct *p)
2213 {
2214         /*
2215          * FIXME: It's better to move charges of this process from old
2216          * memcg to new memcg. But it's just on TODO-List now.
2217          */
2218 }
2219
2220 struct cgroup_subsys mem_cgroup_subsys = {
2221         .name = "memory",
2222         .subsys_id = mem_cgroup_subsys_id,
2223         .create = mem_cgroup_create,
2224         .pre_destroy = mem_cgroup_pre_destroy,
2225         .destroy = mem_cgroup_destroy,
2226         .populate = mem_cgroup_populate,
2227         .attach = mem_cgroup_move_task,
2228         .early_init = 0,
2229 };
2230
2231 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2232
2233 static int __init disable_swap_account(char *s)
2234 {
2235         really_do_swap_account = 0;
2236         return 1;
2237 }
2238 __setup("noswapaccount", disable_swap_account);
2239 #endif