memcg: hierarchy avoid unnecessary reclaim
[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         unsigned long   last_oom_jiffies;
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 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
575 {
576         if (do_swap_account) {
577                 if (res_counter_check_under_limit(&mem->res) &&
578                         res_counter_check_under_limit(&mem->memsw))
579                         return true;
580         } else
581                 if (res_counter_check_under_limit(&mem->res))
582                         return true;
583         return false;
584 }
585
586 /*
587  * Dance down the hierarchy if needed to reclaim memory. We remember the
588  * last child we reclaimed from, so that we don't end up penalizing
589  * one child extensively based on its position in the children list.
590  *
591  * root_mem is the original ancestor that we've been reclaim from.
592  */
593 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
594                                                 gfp_t gfp_mask, bool noswap)
595 {
596         struct mem_cgroup *next_mem;
597         int ret = 0;
598
599         /*
600          * Reclaim unconditionally and don't check for return value.
601          * We need to reclaim in the current group and down the tree.
602          * One might think about checking for children before reclaiming,
603          * but there might be left over accounting, even after children
604          * have left.
605          */
606         ret = try_to_free_mem_cgroup_pages(root_mem, gfp_mask, noswap);
607         if (mem_cgroup_check_under_limit(root_mem))
608                 return 0;
609         if (!root_mem->use_hierarchy)
610                 return ret;
611
612         next_mem = mem_cgroup_get_first_node(root_mem);
613
614         while (next_mem != root_mem) {
615                 if (next_mem->obsolete) {
616                         mem_cgroup_put(next_mem);
617                         cgroup_lock();
618                         next_mem = mem_cgroup_get_first_node(root_mem);
619                         cgroup_unlock();
620                         continue;
621                 }
622                 ret = try_to_free_mem_cgroup_pages(next_mem, gfp_mask, noswap);
623                 if (mem_cgroup_check_under_limit(root_mem))
624                         return 0;
625                 cgroup_lock();
626                 next_mem = mem_cgroup_get_next_node(next_mem, root_mem);
627                 cgroup_unlock();
628         }
629         return ret;
630 }
631
632 bool mem_cgroup_oom_called(struct task_struct *task)
633 {
634         bool ret = false;
635         struct mem_cgroup *mem;
636         struct mm_struct *mm;
637
638         rcu_read_lock();
639         mm = task->mm;
640         if (!mm)
641                 mm = &init_mm;
642         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
643         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
644                 ret = true;
645         rcu_read_unlock();
646         return ret;
647 }
648 /*
649  * Unlike exported interface, "oom" parameter is added. if oom==true,
650  * oom-killer can be invoked.
651  */
652 static int __mem_cgroup_try_charge(struct mm_struct *mm,
653                         gfp_t gfp_mask, struct mem_cgroup **memcg,
654                         bool oom)
655 {
656         struct mem_cgroup *mem, *mem_over_limit;
657         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
658         struct res_counter *fail_res;
659
660         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
661                 /* Don't account this! */
662                 *memcg = NULL;
663                 return 0;
664         }
665
666         /*
667          * We always charge the cgroup the mm_struct belongs to.
668          * The mm_struct's mem_cgroup changes on task migration if the
669          * thread group leader migrates. It's possible that mm is not
670          * set, if so charge the init_mm (happens for pagecache usage).
671          */
672         if (likely(!*memcg)) {
673                 rcu_read_lock();
674                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
675                 if (unlikely(!mem)) {
676                         rcu_read_unlock();
677                         return 0;
678                 }
679                 /*
680                  * For every charge from the cgroup, increment reference count
681                  */
682                 css_get(&mem->css);
683                 *memcg = mem;
684                 rcu_read_unlock();
685         } else {
686                 mem = *memcg;
687                 css_get(&mem->css);
688         }
689
690         while (1) {
691                 int ret;
692                 bool noswap = false;
693
694                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
695                 if (likely(!ret)) {
696                         if (!do_swap_account)
697                                 break;
698                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
699                                                         &fail_res);
700                         if (likely(!ret))
701                                 break;
702                         /* mem+swap counter fails */
703                         res_counter_uncharge(&mem->res, PAGE_SIZE);
704                         noswap = true;
705                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
706                                                                         memsw);
707                 } else
708                         /* mem counter fails */
709                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
710                                                                         res);
711
712                 if (!(gfp_mask & __GFP_WAIT))
713                         goto nomem;
714
715                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
716                                                         noswap);
717
718                 /*
719                  * try_to_free_mem_cgroup_pages() might not give us a full
720                  * picture of reclaim. Some pages are reclaimed and might be
721                  * moved to swap cache or just unmapped from the cgroup.
722                  * Check the limit again to see if the reclaim reduced the
723                  * current usage of the cgroup before giving up
724                  *
725                  */
726                 if (mem_cgroup_check_under_limit(mem_over_limit))
727                         continue;
728
729                 if (!nr_retries--) {
730                         if (oom) {
731                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
732                                 mem_over_limit->last_oom_jiffies = jiffies;
733                         }
734                         goto nomem;
735                 }
736         }
737         return 0;
738 nomem:
739         css_put(&mem->css);
740         return -ENOMEM;
741 }
742
743 /**
744  * mem_cgroup_try_charge - get charge of PAGE_SIZE.
745  * @mm: an mm_struct which is charged against. (when *memcg is NULL)
746  * @gfp_mask: gfp_mask for reclaim.
747  * @memcg: a pointer to memory cgroup which is charged against.
748  *
749  * charge against memory cgroup pointed by *memcg. if *memcg == NULL, estimated
750  * memory cgroup from @mm is got and stored in *memcg.
751  *
752  * Returns 0 if success. -ENOMEM at failure.
753  * This call can invoke OOM-Killer.
754  */
755
756 int mem_cgroup_try_charge(struct mm_struct *mm,
757                           gfp_t mask, struct mem_cgroup **memcg)
758 {
759         return __mem_cgroup_try_charge(mm, mask, memcg, true);
760 }
761
762 /*
763  * commit a charge got by mem_cgroup_try_charge() and makes page_cgroup to be
764  * USED state. If already USED, uncharge and return.
765  */
766
767 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
768                                      struct page_cgroup *pc,
769                                      enum charge_type ctype)
770 {
771         /* try_charge() can return NULL to *memcg, taking care of it. */
772         if (!mem)
773                 return;
774
775         lock_page_cgroup(pc);
776         if (unlikely(PageCgroupUsed(pc))) {
777                 unlock_page_cgroup(pc);
778                 res_counter_uncharge(&mem->res, PAGE_SIZE);
779                 if (do_swap_account)
780                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
781                 css_put(&mem->css);
782                 return;
783         }
784         pc->mem_cgroup = mem;
785         smp_wmb();
786         pc->flags = pcg_default_flags[ctype];
787
788         mem_cgroup_charge_statistics(mem, pc, true);
789
790         unlock_page_cgroup(pc);
791 }
792
793 /**
794  * mem_cgroup_move_account - move account of the page
795  * @pc: page_cgroup of the page.
796  * @from: mem_cgroup which the page is moved from.
797  * @to: mem_cgroup which the page is moved to. @from != @to.
798  *
799  * The caller must confirm following.
800  * - page is not on LRU (isolate_page() is useful.)
801  *
802  * returns 0 at success,
803  * returns -EBUSY when lock is busy or "pc" is unstable.
804  *
805  * This function does "uncharge" from old cgroup but doesn't do "charge" to
806  * new cgroup. It should be done by a caller.
807  */
808
809 static int mem_cgroup_move_account(struct page_cgroup *pc,
810         struct mem_cgroup *from, struct mem_cgroup *to)
811 {
812         struct mem_cgroup_per_zone *from_mz, *to_mz;
813         int nid, zid;
814         int ret = -EBUSY;
815
816         VM_BUG_ON(from == to);
817         VM_BUG_ON(PageLRU(pc->page));
818
819         nid = page_cgroup_nid(pc);
820         zid = page_cgroup_zid(pc);
821         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
822         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
823
824         if (!trylock_page_cgroup(pc))
825                 return ret;
826
827         if (!PageCgroupUsed(pc))
828                 goto out;
829
830         if (pc->mem_cgroup != from)
831                 goto out;
832
833         css_put(&from->css);
834         res_counter_uncharge(&from->res, PAGE_SIZE);
835         mem_cgroup_charge_statistics(from, pc, false);
836         if (do_swap_account)
837                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
838         pc->mem_cgroup = to;
839         mem_cgroup_charge_statistics(to, pc, true);
840         css_get(&to->css);
841         ret = 0;
842 out:
843         unlock_page_cgroup(pc);
844         return ret;
845 }
846
847 /*
848  * move charges to its parent.
849  */
850
851 static int mem_cgroup_move_parent(struct page_cgroup *pc,
852                                   struct mem_cgroup *child,
853                                   gfp_t gfp_mask)
854 {
855         struct page *page = pc->page;
856         struct cgroup *cg = child->css.cgroup;
857         struct cgroup *pcg = cg->parent;
858         struct mem_cgroup *parent;
859         int ret;
860
861         /* Is ROOT ? */
862         if (!pcg)
863                 return -EINVAL;
864
865
866         parent = mem_cgroup_from_cont(pcg);
867
868
869         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
870         if (ret || !parent)
871                 return ret;
872
873         if (!get_page_unless_zero(page))
874                 return -EBUSY;
875
876         ret = isolate_lru_page(page);
877
878         if (ret)
879                 goto cancel;
880
881         ret = mem_cgroup_move_account(pc, child, parent);
882
883         /* drop extra refcnt by try_charge() (move_account increment one) */
884         css_put(&parent->css);
885         putback_lru_page(page);
886         if (!ret) {
887                 put_page(page);
888                 return 0;
889         }
890         /* uncharge if move fails */
891 cancel:
892         res_counter_uncharge(&parent->res, PAGE_SIZE);
893         if (do_swap_account)
894                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
895         put_page(page);
896         return ret;
897 }
898
899 /*
900  * Charge the memory controller for page usage.
901  * Return
902  * 0 if the charge was successful
903  * < 0 if the cgroup is over its limit
904  */
905 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
906                                 gfp_t gfp_mask, enum charge_type ctype,
907                                 struct mem_cgroup *memcg)
908 {
909         struct mem_cgroup *mem;
910         struct page_cgroup *pc;
911         int ret;
912
913         pc = lookup_page_cgroup(page);
914         /* can happen at boot */
915         if (unlikely(!pc))
916                 return 0;
917         prefetchw(pc);
918
919         mem = memcg;
920         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
921         if (ret || !mem)
922                 return ret;
923
924         __mem_cgroup_commit_charge(mem, pc, ctype);
925         return 0;
926 }
927
928 int mem_cgroup_newpage_charge(struct page *page,
929                               struct mm_struct *mm, gfp_t gfp_mask)
930 {
931         if (mem_cgroup_disabled())
932                 return 0;
933         if (PageCompound(page))
934                 return 0;
935         /*
936          * If already mapped, we don't have to account.
937          * If page cache, page->mapping has address_space.
938          * But page->mapping may have out-of-use anon_vma pointer,
939          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
940          * is NULL.
941          */
942         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
943                 return 0;
944         if (unlikely(!mm))
945                 mm = &init_mm;
946         return mem_cgroup_charge_common(page, mm, gfp_mask,
947                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
948 }
949
950 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
951                                 gfp_t gfp_mask)
952 {
953         if (mem_cgroup_disabled())
954                 return 0;
955         if (PageCompound(page))
956                 return 0;
957         /*
958          * Corner case handling. This is called from add_to_page_cache()
959          * in usual. But some FS (shmem) precharges this page before calling it
960          * and call add_to_page_cache() with GFP_NOWAIT.
961          *
962          * For GFP_NOWAIT case, the page may be pre-charged before calling
963          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
964          * charge twice. (It works but has to pay a bit larger cost.)
965          */
966         if (!(gfp_mask & __GFP_WAIT)) {
967                 struct page_cgroup *pc;
968
969
970                 pc = lookup_page_cgroup(page);
971                 if (!pc)
972                         return 0;
973                 lock_page_cgroup(pc);
974                 if (PageCgroupUsed(pc)) {
975                         unlock_page_cgroup(pc);
976                         return 0;
977                 }
978                 unlock_page_cgroup(pc);
979         }
980
981         if (unlikely(!mm))
982                 mm = &init_mm;
983
984         if (page_is_file_cache(page))
985                 return mem_cgroup_charge_common(page, mm, gfp_mask,
986                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
987         else
988                 return mem_cgroup_charge_common(page, mm, gfp_mask,
989                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, NULL);
990 }
991
992 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
993                                  struct page *page,
994                                  gfp_t mask, struct mem_cgroup **ptr)
995 {
996         struct mem_cgroup *mem;
997         swp_entry_t     ent;
998
999         if (mem_cgroup_disabled())
1000                 return 0;
1001
1002         if (!do_swap_account)
1003                 goto charge_cur_mm;
1004
1005         /*
1006          * A racing thread's fault, or swapoff, may have already updated
1007          * the pte, and even removed page from swap cache: return success
1008          * to go on to do_swap_page()'s pte_same() test, which should fail.
1009          */
1010         if (!PageSwapCache(page))
1011                 return 0;
1012
1013         ent.val = page_private(page);
1014
1015         mem = lookup_swap_cgroup(ent);
1016         if (!mem || mem->obsolete)
1017                 goto charge_cur_mm;
1018         *ptr = mem;
1019         return __mem_cgroup_try_charge(NULL, mask, ptr, true);
1020 charge_cur_mm:
1021         if (unlikely(!mm))
1022                 mm = &init_mm;
1023         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1024 }
1025
1026 #ifdef CONFIG_SWAP
1027
1028 int mem_cgroup_cache_charge_swapin(struct page *page,
1029                         struct mm_struct *mm, gfp_t mask, bool locked)
1030 {
1031         int ret = 0;
1032
1033         if (mem_cgroup_disabled())
1034                 return 0;
1035         if (unlikely(!mm))
1036                 mm = &init_mm;
1037         if (!locked)
1038                 lock_page(page);
1039         /*
1040          * If not locked, the page can be dropped from SwapCache until
1041          * we reach here.
1042          */
1043         if (PageSwapCache(page)) {
1044                 struct mem_cgroup *mem = NULL;
1045                 swp_entry_t ent;
1046
1047                 ent.val = page_private(page);
1048                 if (do_swap_account) {
1049                         mem = lookup_swap_cgroup(ent);
1050                         if (mem && mem->obsolete)
1051                                 mem = NULL;
1052                         if (mem)
1053                                 mm = NULL;
1054                 }
1055                 ret = mem_cgroup_charge_common(page, mm, mask,
1056                                 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1057
1058                 if (!ret && do_swap_account) {
1059                         /* avoid double counting */
1060                         mem = swap_cgroup_record(ent, NULL);
1061                         if (mem) {
1062                                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1063                                 mem_cgroup_put(mem);
1064                         }
1065                 }
1066         }
1067         if (!locked)
1068                 unlock_page(page);
1069         /* add this page(page_cgroup) to the LRU we want. */
1070         mem_cgroup_lru_fixup(page);
1071
1072         return ret;
1073 }
1074 #endif
1075
1076 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1077 {
1078         struct page_cgroup *pc;
1079
1080         if (mem_cgroup_disabled())
1081                 return;
1082         if (!ptr)
1083                 return;
1084         pc = lookup_page_cgroup(page);
1085         __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1086         /*
1087          * Now swap is on-memory. This means this page may be
1088          * counted both as mem and swap....double count.
1089          * Fix it by uncharging from memsw. This SwapCache is stable
1090          * because we're still under lock_page().
1091          */
1092         if (do_swap_account) {
1093                 swp_entry_t ent = {.val = page_private(page)};
1094                 struct mem_cgroup *memcg;
1095                 memcg = swap_cgroup_record(ent, NULL);
1096                 if (memcg) {
1097                         /* If memcg is obsolete, memcg can be != ptr */
1098                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1099                         mem_cgroup_put(memcg);
1100                 }
1101
1102         }
1103         /* add this page(page_cgroup) to the LRU we want. */
1104         mem_cgroup_lru_fixup(page);
1105 }
1106
1107 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1108 {
1109         if (mem_cgroup_disabled())
1110                 return;
1111         if (!mem)
1112                 return;
1113         res_counter_uncharge(&mem->res, PAGE_SIZE);
1114         if (do_swap_account)
1115                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1116         css_put(&mem->css);
1117 }
1118
1119
1120 /*
1121  * uncharge if !page_mapped(page)
1122  */
1123 static struct mem_cgroup *
1124 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1125 {
1126         struct page_cgroup *pc;
1127         struct mem_cgroup *mem = NULL;
1128         struct mem_cgroup_per_zone *mz;
1129
1130         if (mem_cgroup_disabled())
1131                 return NULL;
1132
1133         if (PageSwapCache(page))
1134                 return NULL;
1135
1136         /*
1137          * Check if our page_cgroup is valid
1138          */
1139         pc = lookup_page_cgroup(page);
1140         if (unlikely(!pc || !PageCgroupUsed(pc)))
1141                 return NULL;
1142
1143         lock_page_cgroup(pc);
1144
1145         mem = pc->mem_cgroup;
1146
1147         if (!PageCgroupUsed(pc))
1148                 goto unlock_out;
1149
1150         switch (ctype) {
1151         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1152                 if (page_mapped(page))
1153                         goto unlock_out;
1154                 break;
1155         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1156                 if (!PageAnon(page)) {  /* Shared memory */
1157                         if (page->mapping && !page_is_file_cache(page))
1158                                 goto unlock_out;
1159                 } else if (page_mapped(page)) /* Anon */
1160                                 goto unlock_out;
1161                 break;
1162         default:
1163                 break;
1164         }
1165
1166         res_counter_uncharge(&mem->res, PAGE_SIZE);
1167         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1168                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1169
1170         mem_cgroup_charge_statistics(mem, pc, false);
1171         ClearPageCgroupUsed(pc);
1172
1173         mz = page_cgroup_zoneinfo(pc);
1174         unlock_page_cgroup(pc);
1175
1176         /* at swapout, this memcg will be accessed to record to swap */
1177         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1178                 css_put(&mem->css);
1179
1180         return mem;
1181
1182 unlock_out:
1183         unlock_page_cgroup(pc);
1184         return NULL;
1185 }
1186
1187 void mem_cgroup_uncharge_page(struct page *page)
1188 {
1189         /* early check. */
1190         if (page_mapped(page))
1191                 return;
1192         if (page->mapping && !PageAnon(page))
1193                 return;
1194         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1195 }
1196
1197 void mem_cgroup_uncharge_cache_page(struct page *page)
1198 {
1199         VM_BUG_ON(page_mapped(page));
1200         VM_BUG_ON(page->mapping);
1201         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1202 }
1203
1204 /*
1205  * called from __delete_from_swap_cache() and drop "page" account.
1206  * memcg information is recorded to swap_cgroup of "ent"
1207  */
1208 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1209 {
1210         struct mem_cgroup *memcg;
1211
1212         memcg = __mem_cgroup_uncharge_common(page,
1213                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1214         /* record memcg information */
1215         if (do_swap_account && memcg) {
1216                 swap_cgroup_record(ent, memcg);
1217                 mem_cgroup_get(memcg);
1218         }
1219         if (memcg)
1220                 css_put(&memcg->css);
1221 }
1222
1223 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1224 /*
1225  * called from swap_entry_free(). remove record in swap_cgroup and
1226  * uncharge "memsw" account.
1227  */
1228 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1229 {
1230         struct mem_cgroup *memcg;
1231
1232         if (!do_swap_account)
1233                 return;
1234
1235         memcg = swap_cgroup_record(ent, NULL);
1236         if (memcg) {
1237                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1238                 mem_cgroup_put(memcg);
1239         }
1240 }
1241 #endif
1242
1243 /*
1244  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1245  * page belongs to.
1246  */
1247 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1248 {
1249         struct page_cgroup *pc;
1250         struct mem_cgroup *mem = NULL;
1251         int ret = 0;
1252
1253         if (mem_cgroup_disabled())
1254                 return 0;
1255
1256         pc = lookup_page_cgroup(page);
1257         lock_page_cgroup(pc);
1258         if (PageCgroupUsed(pc)) {
1259                 mem = pc->mem_cgroup;
1260                 css_get(&mem->css);
1261         }
1262         unlock_page_cgroup(pc);
1263
1264         if (mem) {
1265                 ret = mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem);
1266                 css_put(&mem->css);
1267         }
1268         *ptr = mem;
1269         return ret;
1270 }
1271
1272 /* remove redundant charge if migration failed*/
1273 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1274                 struct page *oldpage, struct page *newpage)
1275 {
1276         struct page *target, *unused;
1277         struct page_cgroup *pc;
1278         enum charge_type ctype;
1279
1280         if (!mem)
1281                 return;
1282
1283         /* at migration success, oldpage->mapping is NULL. */
1284         if (oldpage->mapping) {
1285                 target = oldpage;
1286                 unused = NULL;
1287         } else {
1288                 target = newpage;
1289                 unused = oldpage;
1290         }
1291
1292         if (PageAnon(target))
1293                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1294         else if (page_is_file_cache(target))
1295                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1296         else
1297                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1298
1299         /* unused page is not on radix-tree now. */
1300         if (unused)
1301                 __mem_cgroup_uncharge_common(unused, ctype);
1302
1303         pc = lookup_page_cgroup(target);
1304         /*
1305          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1306          * So, double-counting is effectively avoided.
1307          */
1308         __mem_cgroup_commit_charge(mem, pc, ctype);
1309
1310         /*
1311          * Both of oldpage and newpage are still under lock_page().
1312          * Then, we don't have to care about race in radix-tree.
1313          * But we have to be careful that this page is unmapped or not.
1314          *
1315          * There is a case for !page_mapped(). At the start of
1316          * migration, oldpage was mapped. But now, it's zapped.
1317          * But we know *target* page is not freed/reused under us.
1318          * mem_cgroup_uncharge_page() does all necessary checks.
1319          */
1320         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1321                 mem_cgroup_uncharge_page(target);
1322 }
1323
1324 /*
1325  * A call to try to shrink memory usage under specified resource controller.
1326  * This is typically used for page reclaiming for shmem for reducing side
1327  * effect of page allocation from shmem, which is used by some mem_cgroup.
1328  */
1329 int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
1330 {
1331         struct mem_cgroup *mem;
1332         int progress = 0;
1333         int retry = MEM_CGROUP_RECLAIM_RETRIES;
1334
1335         if (mem_cgroup_disabled())
1336                 return 0;
1337         if (!mm)
1338                 return 0;
1339
1340         rcu_read_lock();
1341         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1342         if (unlikely(!mem)) {
1343                 rcu_read_unlock();
1344                 return 0;
1345         }
1346         css_get(&mem->css);
1347         rcu_read_unlock();
1348
1349         do {
1350                 progress = try_to_free_mem_cgroup_pages(mem, gfp_mask, true);
1351                 progress += mem_cgroup_check_under_limit(mem);
1352         } while (!progress && --retry);
1353
1354         css_put(&mem->css);
1355         if (!retry)
1356                 return -ENOMEM;
1357         return 0;
1358 }
1359
1360 static DEFINE_MUTEX(set_limit_mutex);
1361
1362 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1363                                 unsigned long long val)
1364 {
1365
1366         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1367         int progress;
1368         u64 memswlimit;
1369         int ret = 0;
1370
1371         while (retry_count) {
1372                 if (signal_pending(current)) {
1373                         ret = -EINTR;
1374                         break;
1375                 }
1376                 /*
1377                  * Rather than hide all in some function, I do this in
1378                  * open coded manner. You see what this really does.
1379                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1380                  */
1381                 mutex_lock(&set_limit_mutex);
1382                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1383                 if (memswlimit < val) {
1384                         ret = -EINVAL;
1385                         mutex_unlock(&set_limit_mutex);
1386                         break;
1387                 }
1388                 ret = res_counter_set_limit(&memcg->res, val);
1389                 mutex_unlock(&set_limit_mutex);
1390
1391                 if (!ret)
1392                         break;
1393
1394                 progress = try_to_free_mem_cgroup_pages(memcg,
1395                                 GFP_KERNEL, false);
1396                 if (!progress)                  retry_count--;
1397         }
1398         return ret;
1399 }
1400
1401 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1402                                 unsigned long long val)
1403 {
1404         int retry_count = MEM_CGROUP_RECLAIM_RETRIES;
1405         u64 memlimit, oldusage, curusage;
1406         int ret;
1407
1408         if (!do_swap_account)
1409                 return -EINVAL;
1410
1411         while (retry_count) {
1412                 if (signal_pending(current)) {
1413                         ret = -EINTR;
1414                         break;
1415                 }
1416                 /*
1417                  * Rather than hide all in some function, I do this in
1418                  * open coded manner. You see what this really does.
1419                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1420                  */
1421                 mutex_lock(&set_limit_mutex);
1422                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1423                 if (memlimit > val) {
1424                         ret = -EINVAL;
1425                         mutex_unlock(&set_limit_mutex);
1426                         break;
1427                 }
1428                 ret = res_counter_set_limit(&memcg->memsw, val);
1429                 mutex_unlock(&set_limit_mutex);
1430
1431                 if (!ret)
1432                         break;
1433
1434                 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1435                 try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, true);
1436                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1437                 if (curusage >= oldusage)
1438                         retry_count--;
1439         }
1440         return ret;
1441 }
1442
1443 /*
1444  * This routine traverse page_cgroup in given list and drop them all.
1445  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1446  */
1447 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1448                                 int node, int zid, enum lru_list lru)
1449 {
1450         struct zone *zone;
1451         struct mem_cgroup_per_zone *mz;
1452         struct page_cgroup *pc, *busy;
1453         unsigned long flags, loop;
1454         struct list_head *list;
1455         int ret = 0;
1456
1457         zone = &NODE_DATA(node)->node_zones[zid];
1458         mz = mem_cgroup_zoneinfo(mem, node, zid);
1459         list = &mz->lists[lru];
1460
1461         loop = MEM_CGROUP_ZSTAT(mz, lru);
1462         /* give some margin against EBUSY etc...*/
1463         loop += 256;
1464         busy = NULL;
1465         while (loop--) {
1466                 ret = 0;
1467                 spin_lock_irqsave(&zone->lru_lock, flags);
1468                 if (list_empty(list)) {
1469                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1470                         break;
1471                 }
1472                 pc = list_entry(list->prev, struct page_cgroup, lru);
1473                 if (busy == pc) {
1474                         list_move(&pc->lru, list);
1475                         busy = 0;
1476                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1477                         continue;
1478                 }
1479                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1480
1481                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1482                 if (ret == -ENOMEM)
1483                         break;
1484
1485                 if (ret == -EBUSY || ret == -EINVAL) {
1486                         /* found lock contention or "pc" is obsolete. */
1487                         busy = pc;
1488                         cond_resched();
1489                 } else
1490                         busy = NULL;
1491         }
1492
1493         if (!ret && !list_empty(list))
1494                 return -EBUSY;
1495         return ret;
1496 }
1497
1498 /*
1499  * make mem_cgroup's charge to be 0 if there is no task.
1500  * This enables deleting this mem_cgroup.
1501  */
1502 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1503 {
1504         int ret;
1505         int node, zid, shrink;
1506         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1507         struct cgroup *cgrp = mem->css.cgroup;
1508
1509         css_get(&mem->css);
1510
1511         shrink = 0;
1512         /* should free all ? */
1513         if (free_all)
1514                 goto try_to_free;
1515 move_account:
1516         while (mem->res.usage > 0) {
1517                 ret = -EBUSY;
1518                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1519                         goto out;
1520                 ret = -EINTR;
1521                 if (signal_pending(current))
1522                         goto out;
1523                 /* This is for making all *used* pages to be on LRU. */
1524                 lru_add_drain_all();
1525                 ret = 0;
1526                 for_each_node_state(node, N_POSSIBLE) {
1527                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1528                                 enum lru_list l;
1529                                 for_each_lru(l) {
1530                                         ret = mem_cgroup_force_empty_list(mem,
1531                                                         node, zid, l);
1532                                         if (ret)
1533                                                 break;
1534                                 }
1535                         }
1536                         if (ret)
1537                                 break;
1538                 }
1539                 /* it seems parent cgroup doesn't have enough mem */
1540                 if (ret == -ENOMEM)
1541                         goto try_to_free;
1542                 cond_resched();
1543         }
1544         ret = 0;
1545 out:
1546         css_put(&mem->css);
1547         return ret;
1548
1549 try_to_free:
1550         /* returns EBUSY if there is a task or if we come here twice. */
1551         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1552                 ret = -EBUSY;
1553                 goto out;
1554         }
1555         /* we call try-to-free pages for make this cgroup empty */
1556         lru_add_drain_all();
1557         /* try to free all pages in this cgroup */
1558         shrink = 1;
1559         while (nr_retries && mem->res.usage > 0) {
1560                 int progress;
1561
1562                 if (signal_pending(current)) {
1563                         ret = -EINTR;
1564                         goto out;
1565                 }
1566                 progress = try_to_free_mem_cgroup_pages(mem,
1567                                                   GFP_KERNEL, false);
1568                 if (!progress) {
1569                         nr_retries--;
1570                         /* maybe some writeback is necessary */
1571                         congestion_wait(WRITE, HZ/10);
1572                 }
1573
1574         }
1575         lru_add_drain();
1576         /* try move_account...there may be some *locked* pages. */
1577         if (mem->res.usage)
1578                 goto move_account;
1579         ret = 0;
1580         goto out;
1581 }
1582
1583 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1584 {
1585         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1586 }
1587
1588
1589 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1590 {
1591         return mem_cgroup_from_cont(cont)->use_hierarchy;
1592 }
1593
1594 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1595                                         u64 val)
1596 {
1597         int retval = 0;
1598         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1599         struct cgroup *parent = cont->parent;
1600         struct mem_cgroup *parent_mem = NULL;
1601
1602         if (parent)
1603                 parent_mem = mem_cgroup_from_cont(parent);
1604
1605         cgroup_lock();
1606         /*
1607          * If parent's use_hiearchy is set, we can't make any modifications
1608          * in the child subtrees. If it is unset, then the change can
1609          * occur, provided the current cgroup has no children.
1610          *
1611          * For the root cgroup, parent_mem is NULL, we allow value to be
1612          * set if there are no children.
1613          */
1614         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1615                                 (val == 1 || val == 0)) {
1616                 if (list_empty(&cont->children))
1617                         mem->use_hierarchy = val;
1618                 else
1619                         retval = -EBUSY;
1620         } else
1621                 retval = -EINVAL;
1622         cgroup_unlock();
1623
1624         return retval;
1625 }
1626
1627 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1628 {
1629         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1630         u64 val = 0;
1631         int type, name;
1632
1633         type = MEMFILE_TYPE(cft->private);
1634         name = MEMFILE_ATTR(cft->private);
1635         switch (type) {
1636         case _MEM:
1637                 val = res_counter_read_u64(&mem->res, name);
1638                 break;
1639         case _MEMSWAP:
1640                 if (do_swap_account)
1641                         val = res_counter_read_u64(&mem->memsw, name);
1642                 break;
1643         default:
1644                 BUG();
1645                 break;
1646         }
1647         return val;
1648 }
1649 /*
1650  * The user of this function is...
1651  * RES_LIMIT.
1652  */
1653 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1654                             const char *buffer)
1655 {
1656         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1657         int type, name;
1658         unsigned long long val;
1659         int ret;
1660
1661         type = MEMFILE_TYPE(cft->private);
1662         name = MEMFILE_ATTR(cft->private);
1663         switch (name) {
1664         case RES_LIMIT:
1665                 /* This function does all necessary parse...reuse it */
1666                 ret = res_counter_memparse_write_strategy(buffer, &val);
1667                 if (ret)
1668                         break;
1669                 if (type == _MEM)
1670                         ret = mem_cgroup_resize_limit(memcg, val);
1671                 else
1672                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1673                 break;
1674         default:
1675                 ret = -EINVAL; /* should be BUG() ? */
1676                 break;
1677         }
1678         return ret;
1679 }
1680
1681 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
1682 {
1683         struct mem_cgroup *mem;
1684         int type, name;
1685
1686         mem = mem_cgroup_from_cont(cont);
1687         type = MEMFILE_TYPE(event);
1688         name = MEMFILE_ATTR(event);
1689         switch (name) {
1690         case RES_MAX_USAGE:
1691                 if (type == _MEM)
1692                         res_counter_reset_max(&mem->res);
1693                 else
1694                         res_counter_reset_max(&mem->memsw);
1695                 break;
1696         case RES_FAILCNT:
1697                 if (type == _MEM)
1698                         res_counter_reset_failcnt(&mem->res);
1699                 else
1700                         res_counter_reset_failcnt(&mem->memsw);
1701                 break;
1702         }
1703         return 0;
1704 }
1705
1706 static const struct mem_cgroup_stat_desc {
1707         const char *msg;
1708         u64 unit;
1709 } mem_cgroup_stat_desc[] = {
1710         [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
1711         [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
1712         [MEM_CGROUP_STAT_PGPGIN_COUNT] = {"pgpgin", 1, },
1713         [MEM_CGROUP_STAT_PGPGOUT_COUNT] = {"pgpgout", 1, },
1714 };
1715
1716 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
1717                                  struct cgroup_map_cb *cb)
1718 {
1719         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
1720         struct mem_cgroup_stat *stat = &mem_cont->stat;
1721         int i;
1722
1723         for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
1724                 s64 val;
1725
1726                 val = mem_cgroup_read_stat(stat, i);
1727                 val *= mem_cgroup_stat_desc[i].unit;
1728                 cb->fill(cb, mem_cgroup_stat_desc[i].msg, val);
1729         }
1730         /* showing # of active pages */
1731         {
1732                 unsigned long active_anon, inactive_anon;
1733                 unsigned long active_file, inactive_file;
1734                 unsigned long unevictable;
1735
1736                 inactive_anon = mem_cgroup_get_all_zonestat(mem_cont,
1737                                                 LRU_INACTIVE_ANON);
1738                 active_anon = mem_cgroup_get_all_zonestat(mem_cont,
1739                                                 LRU_ACTIVE_ANON);
1740                 inactive_file = mem_cgroup_get_all_zonestat(mem_cont,
1741                                                 LRU_INACTIVE_FILE);
1742                 active_file = mem_cgroup_get_all_zonestat(mem_cont,
1743                                                 LRU_ACTIVE_FILE);
1744                 unevictable = mem_cgroup_get_all_zonestat(mem_cont,
1745                                                         LRU_UNEVICTABLE);
1746
1747                 cb->fill(cb, "active_anon", (active_anon) * PAGE_SIZE);
1748                 cb->fill(cb, "inactive_anon", (inactive_anon) * PAGE_SIZE);
1749                 cb->fill(cb, "active_file", (active_file) * PAGE_SIZE);
1750                 cb->fill(cb, "inactive_file", (inactive_file) * PAGE_SIZE);
1751                 cb->fill(cb, "unevictable", unevictable * PAGE_SIZE);
1752
1753         }
1754         return 0;
1755 }
1756
1757
1758 static struct cftype mem_cgroup_files[] = {
1759         {
1760                 .name = "usage_in_bytes",
1761                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
1762                 .read_u64 = mem_cgroup_read,
1763         },
1764         {
1765                 .name = "max_usage_in_bytes",
1766                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
1767                 .trigger = mem_cgroup_reset,
1768                 .read_u64 = mem_cgroup_read,
1769         },
1770         {
1771                 .name = "limit_in_bytes",
1772                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
1773                 .write_string = mem_cgroup_write,
1774                 .read_u64 = mem_cgroup_read,
1775         },
1776         {
1777                 .name = "failcnt",
1778                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
1779                 .trigger = mem_cgroup_reset,
1780                 .read_u64 = mem_cgroup_read,
1781         },
1782         {
1783                 .name = "stat",
1784                 .read_map = mem_control_stat_show,
1785         },
1786         {
1787                 .name = "force_empty",
1788                 .trigger = mem_cgroup_force_empty_write,
1789         },
1790         {
1791                 .name = "use_hierarchy",
1792                 .write_u64 = mem_cgroup_hierarchy_write,
1793                 .read_u64 = mem_cgroup_hierarchy_read,
1794         },
1795 };
1796
1797 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1798 static struct cftype memsw_cgroup_files[] = {
1799         {
1800                 .name = "memsw.usage_in_bytes",
1801                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
1802                 .read_u64 = mem_cgroup_read,
1803         },
1804         {
1805                 .name = "memsw.max_usage_in_bytes",
1806                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
1807                 .trigger = mem_cgroup_reset,
1808                 .read_u64 = mem_cgroup_read,
1809         },
1810         {
1811                 .name = "memsw.limit_in_bytes",
1812                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
1813                 .write_string = mem_cgroup_write,
1814                 .read_u64 = mem_cgroup_read,
1815         },
1816         {
1817                 .name = "memsw.failcnt",
1818                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
1819                 .trigger = mem_cgroup_reset,
1820                 .read_u64 = mem_cgroup_read,
1821         },
1822 };
1823
1824 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
1825 {
1826         if (!do_swap_account)
1827                 return 0;
1828         return cgroup_add_files(cont, ss, memsw_cgroup_files,
1829                                 ARRAY_SIZE(memsw_cgroup_files));
1830 };
1831 #else
1832 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
1833 {
1834         return 0;
1835 }
1836 #endif
1837
1838 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1839 {
1840         struct mem_cgroup_per_node *pn;
1841         struct mem_cgroup_per_zone *mz;
1842         enum lru_list l;
1843         int zone, tmp = node;
1844         /*
1845          * This routine is called against possible nodes.
1846          * But it's BUG to call kmalloc() against offline node.
1847          *
1848          * TODO: this routine can waste much memory for nodes which will
1849          *       never be onlined. It's better to use memory hotplug callback
1850          *       function.
1851          */
1852         if (!node_state(node, N_NORMAL_MEMORY))
1853                 tmp = -1;
1854         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
1855         if (!pn)
1856                 return 1;
1857
1858         mem->info.nodeinfo[node] = pn;
1859         memset(pn, 0, sizeof(*pn));
1860
1861         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
1862                 mz = &pn->zoneinfo[zone];
1863                 for_each_lru(l)
1864                         INIT_LIST_HEAD(&mz->lists[l]);
1865         }
1866         return 0;
1867 }
1868
1869 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1870 {
1871         kfree(mem->info.nodeinfo[node]);
1872 }
1873
1874 static int mem_cgroup_size(void)
1875 {
1876         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
1877         return sizeof(struct mem_cgroup) + cpustat_size;
1878 }
1879
1880 static struct mem_cgroup *mem_cgroup_alloc(void)
1881 {
1882         struct mem_cgroup *mem;
1883         int size = mem_cgroup_size();
1884
1885         if (size < PAGE_SIZE)
1886                 mem = kmalloc(size, GFP_KERNEL);
1887         else
1888                 mem = vmalloc(size);
1889
1890         if (mem)
1891                 memset(mem, 0, size);
1892         return mem;
1893 }
1894
1895 /*
1896  * At destroying mem_cgroup, references from swap_cgroup can remain.
1897  * (scanning all at force_empty is too costly...)
1898  *
1899  * Instead of clearing all references at force_empty, we remember
1900  * the number of reference from swap_cgroup and free mem_cgroup when
1901  * it goes down to 0.
1902  *
1903  * When mem_cgroup is destroyed, mem->obsolete will be set to 0 and
1904  * entry which points to this memcg will be ignore at swapin.
1905  *
1906  * Removal of cgroup itself succeeds regardless of refs from swap.
1907  */
1908
1909 static void mem_cgroup_free(struct mem_cgroup *mem)
1910 {
1911         int node;
1912
1913         if (atomic_read(&mem->refcnt) > 0)
1914                 return;
1915
1916
1917         for_each_node_state(node, N_POSSIBLE)
1918                 free_mem_cgroup_per_zone_info(mem, node);
1919
1920         if (mem_cgroup_size() < PAGE_SIZE)
1921                 kfree(mem);
1922         else
1923                 vfree(mem);
1924 }
1925
1926 static void mem_cgroup_get(struct mem_cgroup *mem)
1927 {
1928         atomic_inc(&mem->refcnt);
1929 }
1930
1931 static void mem_cgroup_put(struct mem_cgroup *mem)
1932 {
1933         if (atomic_dec_and_test(&mem->refcnt)) {
1934                 if (!mem->obsolete)
1935                         return;
1936                 mem_cgroup_free(mem);
1937         }
1938 }
1939
1940
1941 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1942 static void __init enable_swap_cgroup(void)
1943 {
1944         if (!mem_cgroup_disabled() && really_do_swap_account)
1945                 do_swap_account = 1;
1946 }
1947 #else
1948 static void __init enable_swap_cgroup(void)
1949 {
1950 }
1951 #endif
1952
1953 static struct cgroup_subsys_state *
1954 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1955 {
1956         struct mem_cgroup *mem, *parent;
1957         int node;
1958
1959         mem = mem_cgroup_alloc();
1960         if (!mem)
1961                 return ERR_PTR(-ENOMEM);
1962
1963         for_each_node_state(node, N_POSSIBLE)
1964                 if (alloc_mem_cgroup_per_zone_info(mem, node))
1965                         goto free_out;
1966         /* root ? */
1967         if (cont->parent == NULL) {
1968                 enable_swap_cgroup();
1969                 parent = NULL;
1970         } else {
1971                 parent = mem_cgroup_from_cont(cont->parent);
1972                 mem->use_hierarchy = parent->use_hierarchy;
1973         }
1974
1975         if (parent && parent->use_hierarchy) {
1976                 res_counter_init(&mem->res, &parent->res);
1977                 res_counter_init(&mem->memsw, &parent->memsw);
1978         } else {
1979                 res_counter_init(&mem->res, NULL);
1980                 res_counter_init(&mem->memsw, NULL);
1981         }
1982
1983         mem->last_scanned_child = NULL;
1984
1985         return &mem->css;
1986 free_out:
1987         for_each_node_state(node, N_POSSIBLE)
1988                 free_mem_cgroup_per_zone_info(mem, node);
1989         mem_cgroup_free(mem);
1990         return ERR_PTR(-ENOMEM);
1991 }
1992
1993 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1994                                         struct cgroup *cont)
1995 {
1996         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1997         mem->obsolete = 1;
1998         mem_cgroup_force_empty(mem, false);
1999 }
2000
2001 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2002                                 struct cgroup *cont)
2003 {
2004         mem_cgroup_free(mem_cgroup_from_cont(cont));
2005 }
2006
2007 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2008                                 struct cgroup *cont)
2009 {
2010         int ret;
2011
2012         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2013                                 ARRAY_SIZE(mem_cgroup_files));
2014
2015         if (!ret)
2016                 ret = register_memsw_files(cont, ss);
2017         return ret;
2018 }
2019
2020 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2021                                 struct cgroup *cont,
2022                                 struct cgroup *old_cont,
2023                                 struct task_struct *p)
2024 {
2025         /*
2026          * FIXME: It's better to move charges of this process from old
2027          * memcg to new memcg. But it's just on TODO-List now.
2028          */
2029 }
2030
2031 struct cgroup_subsys mem_cgroup_subsys = {
2032         .name = "memory",
2033         .subsys_id = mem_cgroup_subsys_id,
2034         .create = mem_cgroup_create,
2035         .pre_destroy = mem_cgroup_pre_destroy,
2036         .destroy = mem_cgroup_destroy,
2037         .populate = mem_cgroup_populate,
2038         .attach = mem_cgroup_move_task,
2039         .early_init = 0,
2040 };
2041
2042 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2043
2044 static int __init disable_swap_account(char *s)
2045 {
2046         really_do_swap_account = 0;
2047         return 1;
2048 }
2049 __setup("noswapaccount", disable_swap_account);
2050 #endif