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