memcg: fix deadlock between lock_page_cgroup and mapping tree_lock
[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/limits.h>
31 #include <linux/mutex.h>
32 #include <linux/slab.h>
33 #include <linux/swap.h>
34 #include <linux/spinlock.h>
35 #include <linux/fs.h>
36 #include <linux/seq_file.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mm_inline.h>
39 #include <linux/page_cgroup.h>
40 #include "internal.h"
41
42 #include <asm/uaccess.h>
43
44 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
45 #define MEM_CGROUP_RECLAIM_RETRIES      5
46
47 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
48 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 0 */
49 int do_swap_account __read_mostly;
50 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
51 #else
52 #define do_swap_account         (0)
53 #endif
54
55 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
56
57 /*
58  * Statistics for memory cgroup.
59  */
60 enum mem_cgroup_stat_index {
61         /*
62          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
63          */
64         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
65         MEM_CGROUP_STAT_RSS,       /* # of pages charged as rss */
66         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
67         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
68
69         MEM_CGROUP_STAT_NSTATS,
70 };
71
72 struct mem_cgroup_stat_cpu {
73         s64 count[MEM_CGROUP_STAT_NSTATS];
74 } ____cacheline_aligned_in_smp;
75
76 struct mem_cgroup_stat {
77         struct mem_cgroup_stat_cpu cpustat[0];
78 };
79
80 /*
81  * For accounting under irq disable, no need for increment preempt count.
82  */
83 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
84                 enum mem_cgroup_stat_index idx, int val)
85 {
86         stat->count[idx] += val;
87 }
88
89 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
90                 enum mem_cgroup_stat_index idx)
91 {
92         int cpu;
93         s64 ret = 0;
94         for_each_possible_cpu(cpu)
95                 ret += stat->cpustat[cpu].count[idx];
96         return ret;
97 }
98
99 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
100 {
101         s64 ret;
102
103         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
104         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
105         return ret;
106 }
107
108 /*
109  * per-zone information in memory controller.
110  */
111 struct mem_cgroup_per_zone {
112         /*
113          * spin_lock to protect the per cgroup LRU
114          */
115         struct list_head        lists[NR_LRU_LISTS];
116         unsigned long           count[NR_LRU_LISTS];
117
118         struct zone_reclaim_stat reclaim_stat;
119 };
120 /* Macro for accessing counter */
121 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
122
123 struct mem_cgroup_per_node {
124         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
125 };
126
127 struct mem_cgroup_lru_info {
128         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
129 };
130
131 /*
132  * The memory controller data structure. The memory controller controls both
133  * page cache and RSS per cgroup. We would eventually like to provide
134  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
135  * to help the administrator determine what knobs to tune.
136  *
137  * TODO: Add a water mark for the memory controller. Reclaim will begin when
138  * we hit the water mark. May be even add a low water mark, such that
139  * no reclaim occurs from a cgroup at it's low water mark, this is
140  * a feature that will be implemented much later in the future.
141  */
142 struct mem_cgroup {
143         struct cgroup_subsys_state css;
144         /*
145          * the counter to account for memory usage
146          */
147         struct res_counter res;
148         /*
149          * the counter to account for mem+swap usage.
150          */
151         struct res_counter memsw;
152         /*
153          * Per cgroup active and inactive list, similar to the
154          * per zone LRU lists.
155          */
156         struct mem_cgroup_lru_info info;
157
158         /*
159           protect against reclaim related member.
160         */
161         spinlock_t reclaim_param_lock;
162
163         int     prev_priority;  /* for recording reclaim priority */
164
165         /*
166          * While reclaiming in a hiearchy, we cache the last child we
167          * reclaimed from.
168          */
169         int last_scanned_child;
170         /*
171          * Should the accounting and control be hierarchical, per subtree?
172          */
173         bool use_hierarchy;
174         unsigned long   last_oom_jiffies;
175         atomic_t        refcnt;
176
177         unsigned int    swappiness;
178
179         /*
180          * statistics. This must be placed at the end of memcg.
181          */
182         struct mem_cgroup_stat stat;
183 };
184
185 enum charge_type {
186         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
187         MEM_CGROUP_CHARGE_TYPE_MAPPED,
188         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
189         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
190         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
191         NR_CHARGE_TYPE,
192 };
193
194 /* only for here (for easy reading.) */
195 #define PCGF_CACHE      (1UL << PCG_CACHE)
196 #define PCGF_USED       (1UL << PCG_USED)
197 #define PCGF_LOCK       (1UL << PCG_LOCK)
198 static const unsigned long
199 pcg_default_flags[NR_CHARGE_TYPE] = {
200         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */
201         PCGF_USED | PCGF_LOCK, /* Anon */
202         PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */
203         0, /* FORCE */
204 };
205
206 /* for encoding cft->private value on file */
207 #define _MEM                    (0)
208 #define _MEMSWAP                (1)
209 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
210 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
211 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
212
213 static void mem_cgroup_get(struct mem_cgroup *mem);
214 static void mem_cgroup_put(struct mem_cgroup *mem);
215 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
216
217 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
218                                          struct page_cgroup *pc,
219                                          bool charge)
220 {
221         int val = (charge)? 1 : -1;
222         struct mem_cgroup_stat *stat = &mem->stat;
223         struct mem_cgroup_stat_cpu *cpustat;
224         int cpu = get_cpu();
225
226         cpustat = &stat->cpustat[cpu];
227         if (PageCgroupCache(pc))
228                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
229         else
230                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
231
232         if (charge)
233                 __mem_cgroup_stat_add_safe(cpustat,
234                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
235         else
236                 __mem_cgroup_stat_add_safe(cpustat,
237                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
238         put_cpu();
239 }
240
241 static struct mem_cgroup_per_zone *
242 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
243 {
244         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
245 }
246
247 static struct mem_cgroup_per_zone *
248 page_cgroup_zoneinfo(struct page_cgroup *pc)
249 {
250         struct mem_cgroup *mem = pc->mem_cgroup;
251         int nid = page_cgroup_nid(pc);
252         int zid = page_cgroup_zid(pc);
253
254         if (!mem)
255                 return NULL;
256
257         return mem_cgroup_zoneinfo(mem, nid, zid);
258 }
259
260 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
261                                         enum lru_list idx)
262 {
263         int nid, zid;
264         struct mem_cgroup_per_zone *mz;
265         u64 total = 0;
266
267         for_each_online_node(nid)
268                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
269                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
270                         total += MEM_CGROUP_ZSTAT(mz, idx);
271                 }
272         return total;
273 }
274
275 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
276 {
277         return container_of(cgroup_subsys_state(cont,
278                                 mem_cgroup_subsys_id), struct mem_cgroup,
279                                 css);
280 }
281
282 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
283 {
284         /*
285          * mm_update_next_owner() may clear mm->owner to NULL
286          * if it races with swapoff, page migration, etc.
287          * So this can be called with p == NULL.
288          */
289         if (unlikely(!p))
290                 return NULL;
291
292         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
293                                 struct mem_cgroup, css);
294 }
295
296 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
297 {
298         struct mem_cgroup *mem = NULL;
299
300         if (!mm)
301                 return NULL;
302         /*
303          * Because we have no locks, mm->owner's may be being moved to other
304          * cgroup. We use css_tryget() here even if this looks
305          * pessimistic (rather than adding locks here).
306          */
307         rcu_read_lock();
308         do {
309                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
310                 if (unlikely(!mem))
311                         break;
312         } while (!css_tryget(&mem->css));
313         rcu_read_unlock();
314         return mem;
315 }
316
317 static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem)
318 {
319         if (!mem)
320                 return true;
321         return css_is_removed(&mem->css);
322 }
323
324
325 /*
326  * Call callback function against all cgroup under hierarchy tree.
327  */
328 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
329                           int (*func)(struct mem_cgroup *, void *))
330 {
331         int found, ret, nextid;
332         struct cgroup_subsys_state *css;
333         struct mem_cgroup *mem;
334
335         if (!root->use_hierarchy)
336                 return (*func)(root, data);
337
338         nextid = 1;
339         do {
340                 ret = 0;
341                 mem = NULL;
342
343                 rcu_read_lock();
344                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
345                                    &found);
346                 if (css && css_tryget(css))
347                         mem = container_of(css, struct mem_cgroup, css);
348                 rcu_read_unlock();
349
350                 if (mem) {
351                         ret = (*func)(mem, data);
352                         css_put(&mem->css);
353                 }
354                 nextid = found + 1;
355         } while (!ret && css);
356
357         return ret;
358 }
359
360 /*
361  * Following LRU functions are allowed to be used without PCG_LOCK.
362  * Operations are called by routine of global LRU independently from memcg.
363  * What we have to take care of here is validness of pc->mem_cgroup.
364  *
365  * Changes to pc->mem_cgroup happens when
366  * 1. charge
367  * 2. moving account
368  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
369  * It is added to LRU before charge.
370  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
371  * When moving account, the page is not on LRU. It's isolated.
372  */
373
374 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
375 {
376         struct page_cgroup *pc;
377         struct mem_cgroup *mem;
378         struct mem_cgroup_per_zone *mz;
379
380         if (mem_cgroup_disabled())
381                 return;
382         pc = lookup_page_cgroup(page);
383         /* can happen while we handle swapcache. */
384         if (list_empty(&pc->lru) || !pc->mem_cgroup)
385                 return;
386         /*
387          * We don't check PCG_USED bit. It's cleared when the "page" is finally
388          * removed from global LRU.
389          */
390         mz = page_cgroup_zoneinfo(pc);
391         mem = pc->mem_cgroup;
392         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
393         list_del_init(&pc->lru);
394         return;
395 }
396
397 void mem_cgroup_del_lru(struct page *page)
398 {
399         mem_cgroup_del_lru_list(page, page_lru(page));
400 }
401
402 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
403 {
404         struct mem_cgroup_per_zone *mz;
405         struct page_cgroup *pc;
406
407         if (mem_cgroup_disabled())
408                 return;
409
410         pc = lookup_page_cgroup(page);
411         /*
412          * Used bit is set without atomic ops but after smp_wmb().
413          * For making pc->mem_cgroup visible, insert smp_rmb() here.
414          */
415         smp_rmb();
416         /* unused page is not rotated. */
417         if (!PageCgroupUsed(pc))
418                 return;
419         mz = page_cgroup_zoneinfo(pc);
420         list_move(&pc->lru, &mz->lists[lru]);
421 }
422
423 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
424 {
425         struct page_cgroup *pc;
426         struct mem_cgroup_per_zone *mz;
427
428         if (mem_cgroup_disabled())
429                 return;
430         pc = lookup_page_cgroup(page);
431         /*
432          * Used bit is set without atomic ops but after smp_wmb().
433          * For making pc->mem_cgroup visible, insert smp_rmb() here.
434          */
435         smp_rmb();
436         if (!PageCgroupUsed(pc))
437                 return;
438
439         mz = page_cgroup_zoneinfo(pc);
440         MEM_CGROUP_ZSTAT(mz, lru) += 1;
441         list_add(&pc->lru, &mz->lists[lru]);
442 }
443
444 /*
445  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
446  * lru because the page may.be reused after it's fully uncharged (because of
447  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
448  * it again. This function is only used to charge SwapCache. It's done under
449  * lock_page and expected that zone->lru_lock is never held.
450  */
451 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
452 {
453         unsigned long flags;
454         struct zone *zone = page_zone(page);
455         struct page_cgroup *pc = lookup_page_cgroup(page);
456
457         spin_lock_irqsave(&zone->lru_lock, flags);
458         /*
459          * Forget old LRU when this page_cgroup is *not* used. This Used bit
460          * is guarded by lock_page() because the page is SwapCache.
461          */
462         if (!PageCgroupUsed(pc))
463                 mem_cgroup_del_lru_list(page, page_lru(page));
464         spin_unlock_irqrestore(&zone->lru_lock, flags);
465 }
466
467 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
468 {
469         unsigned long flags;
470         struct zone *zone = page_zone(page);
471         struct page_cgroup *pc = lookup_page_cgroup(page);
472
473         spin_lock_irqsave(&zone->lru_lock, flags);
474         /* link when the page is linked to LRU but page_cgroup isn't */
475         if (PageLRU(page) && list_empty(&pc->lru))
476                 mem_cgroup_add_lru_list(page, page_lru(page));
477         spin_unlock_irqrestore(&zone->lru_lock, flags);
478 }
479
480
481 void mem_cgroup_move_lists(struct page *page,
482                            enum lru_list from, enum lru_list to)
483 {
484         if (mem_cgroup_disabled())
485                 return;
486         mem_cgroup_del_lru_list(page, from);
487         mem_cgroup_add_lru_list(page, to);
488 }
489
490 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
491 {
492         int ret;
493         struct mem_cgroup *curr = NULL;
494
495         task_lock(task);
496         rcu_read_lock();
497         curr = try_get_mem_cgroup_from_mm(task->mm);
498         rcu_read_unlock();
499         task_unlock(task);
500         if (!curr)
501                 return 0;
502         if (curr->use_hierarchy)
503                 ret = css_is_ancestor(&curr->css, &mem->css);
504         else
505                 ret = (curr == mem);
506         css_put(&curr->css);
507         return ret;
508 }
509
510 /*
511  * prev_priority control...this will be used in memory reclaim path.
512  */
513 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
514 {
515         int prev_priority;
516
517         spin_lock(&mem->reclaim_param_lock);
518         prev_priority = mem->prev_priority;
519         spin_unlock(&mem->reclaim_param_lock);
520
521         return prev_priority;
522 }
523
524 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
525 {
526         spin_lock(&mem->reclaim_param_lock);
527         if (priority < mem->prev_priority)
528                 mem->prev_priority = priority;
529         spin_unlock(&mem->reclaim_param_lock);
530 }
531
532 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
533 {
534         spin_lock(&mem->reclaim_param_lock);
535         mem->prev_priority = priority;
536         spin_unlock(&mem->reclaim_param_lock);
537 }
538
539 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
540 {
541         unsigned long active;
542         unsigned long inactive;
543         unsigned long gb;
544         unsigned long inactive_ratio;
545
546         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
547         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
548
549         gb = (inactive + active) >> (30 - PAGE_SHIFT);
550         if (gb)
551                 inactive_ratio = int_sqrt(10 * gb);
552         else
553                 inactive_ratio = 1;
554
555         if (present_pages) {
556                 present_pages[0] = inactive;
557                 present_pages[1] = active;
558         }
559
560         return inactive_ratio;
561 }
562
563 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
564 {
565         unsigned long active;
566         unsigned long inactive;
567         unsigned long present_pages[2];
568         unsigned long inactive_ratio;
569
570         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
571
572         inactive = present_pages[0];
573         active = present_pages[1];
574
575         if (inactive * inactive_ratio < active)
576                 return 1;
577
578         return 0;
579 }
580
581 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
582                                        struct zone *zone,
583                                        enum lru_list lru)
584 {
585         int nid = zone->zone_pgdat->node_id;
586         int zid = zone_idx(zone);
587         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
588
589         return MEM_CGROUP_ZSTAT(mz, lru);
590 }
591
592 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
593                                                       struct zone *zone)
594 {
595         int nid = zone->zone_pgdat->node_id;
596         int zid = zone_idx(zone);
597         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
598
599         return &mz->reclaim_stat;
600 }
601
602 struct zone_reclaim_stat *
603 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
604 {
605         struct page_cgroup *pc;
606         struct mem_cgroup_per_zone *mz;
607
608         if (mem_cgroup_disabled())
609                 return NULL;
610
611         pc = lookup_page_cgroup(page);
612         /*
613          * Used bit is set without atomic ops but after smp_wmb().
614          * For making pc->mem_cgroup visible, insert smp_rmb() here.
615          */
616         smp_rmb();
617         if (!PageCgroupUsed(pc))
618                 return NULL;
619
620         mz = page_cgroup_zoneinfo(pc);
621         if (!mz)
622                 return NULL;
623
624         return &mz->reclaim_stat;
625 }
626
627 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
628                                         struct list_head *dst,
629                                         unsigned long *scanned, int order,
630                                         int mode, struct zone *z,
631                                         struct mem_cgroup *mem_cont,
632                                         int active, int file)
633 {
634         unsigned long nr_taken = 0;
635         struct page *page;
636         unsigned long scan;
637         LIST_HEAD(pc_list);
638         struct list_head *src;
639         struct page_cgroup *pc, *tmp;
640         int nid = z->zone_pgdat->node_id;
641         int zid = zone_idx(z);
642         struct mem_cgroup_per_zone *mz;
643         int lru = LRU_FILE * !!file + !!active;
644
645         BUG_ON(!mem_cont);
646         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
647         src = &mz->lists[lru];
648
649         scan = 0;
650         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
651                 if (scan >= nr_to_scan)
652                         break;
653
654                 page = pc->page;
655                 if (unlikely(!PageCgroupUsed(pc)))
656                         continue;
657                 if (unlikely(!PageLRU(page)))
658                         continue;
659
660                 scan++;
661                 if (__isolate_lru_page(page, mode, file) == 0) {
662                         list_move(&page->lru, dst);
663                         nr_taken++;
664                 }
665         }
666
667         *scanned = scan;
668         return nr_taken;
669 }
670
671 #define mem_cgroup_from_res_counter(counter, member)    \
672         container_of(counter, struct mem_cgroup, member)
673
674 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
675 {
676         if (do_swap_account) {
677                 if (res_counter_check_under_limit(&mem->res) &&
678                         res_counter_check_under_limit(&mem->memsw))
679                         return true;
680         } else
681                 if (res_counter_check_under_limit(&mem->res))
682                         return true;
683         return false;
684 }
685
686 static unsigned int get_swappiness(struct mem_cgroup *memcg)
687 {
688         struct cgroup *cgrp = memcg->css.cgroup;
689         unsigned int swappiness;
690
691         /* root ? */
692         if (cgrp->parent == NULL)
693                 return vm_swappiness;
694
695         spin_lock(&memcg->reclaim_param_lock);
696         swappiness = memcg->swappiness;
697         spin_unlock(&memcg->reclaim_param_lock);
698
699         return swappiness;
700 }
701
702 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
703 {
704         int *val = data;
705         (*val)++;
706         return 0;
707 }
708
709 /**
710  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
711  * @memcg: The memory cgroup that went over limit
712  * @p: Task that is going to be killed
713  *
714  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
715  * enabled
716  */
717 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
718 {
719         struct cgroup *task_cgrp;
720         struct cgroup *mem_cgrp;
721         /*
722          * Need a buffer in BSS, can't rely on allocations. The code relies
723          * on the assumption that OOM is serialized for memory controller.
724          * If this assumption is broken, revisit this code.
725          */
726         static char memcg_name[PATH_MAX];
727         int ret;
728
729         if (!memcg)
730                 return;
731
732
733         rcu_read_lock();
734
735         mem_cgrp = memcg->css.cgroup;
736         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
737
738         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
739         if (ret < 0) {
740                 /*
741                  * Unfortunately, we are unable to convert to a useful name
742                  * But we'll still print out the usage information
743                  */
744                 rcu_read_unlock();
745                 goto done;
746         }
747         rcu_read_unlock();
748
749         printk(KERN_INFO "Task in %s killed", memcg_name);
750
751         rcu_read_lock();
752         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
753         if (ret < 0) {
754                 rcu_read_unlock();
755                 goto done;
756         }
757         rcu_read_unlock();
758
759         /*
760          * Continues from above, so we don't need an KERN_ level
761          */
762         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
763 done:
764
765         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
766                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
767                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
768                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
769         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
770                 "failcnt %llu\n",
771                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
772                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
773                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
774 }
775
776 /*
777  * This function returns the number of memcg under hierarchy tree. Returns
778  * 1(self count) if no children.
779  */
780 static int mem_cgroup_count_children(struct mem_cgroup *mem)
781 {
782         int num = 0;
783         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
784         return num;
785 }
786
787 /*
788  * Visit the first child (need not be the first child as per the ordering
789  * of the cgroup list, since we track last_scanned_child) of @mem and use
790  * that to reclaim free pages from.
791  */
792 static struct mem_cgroup *
793 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
794 {
795         struct mem_cgroup *ret = NULL;
796         struct cgroup_subsys_state *css;
797         int nextid, found;
798
799         if (!root_mem->use_hierarchy) {
800                 css_get(&root_mem->css);
801                 ret = root_mem;
802         }
803
804         while (!ret) {
805                 rcu_read_lock();
806                 nextid = root_mem->last_scanned_child + 1;
807                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
808                                    &found);
809                 if (css && css_tryget(css))
810                         ret = container_of(css, struct mem_cgroup, css);
811
812                 rcu_read_unlock();
813                 /* Updates scanning parameter */
814                 spin_lock(&root_mem->reclaim_param_lock);
815                 if (!css) {
816                         /* this means start scan from ID:1 */
817                         root_mem->last_scanned_child = 0;
818                 } else
819                         root_mem->last_scanned_child = found;
820                 spin_unlock(&root_mem->reclaim_param_lock);
821         }
822
823         return ret;
824 }
825
826 /*
827  * Scan the hierarchy if needed to reclaim memory. We remember the last child
828  * we reclaimed from, so that we don't end up penalizing one child extensively
829  * based on its position in the children list.
830  *
831  * root_mem is the original ancestor that we've been reclaim from.
832  *
833  * We give up and return to the caller when we visit root_mem twice.
834  * (other groups can be removed while we're walking....)
835  *
836  * If shrink==true, for avoiding to free too much, this returns immedieately.
837  */
838 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
839                                    gfp_t gfp_mask, bool noswap, bool shrink)
840 {
841         struct mem_cgroup *victim;
842         int ret, total = 0;
843         int loop = 0;
844
845         while (loop < 2) {
846                 victim = mem_cgroup_select_victim(root_mem);
847                 if (victim == root_mem)
848                         loop++;
849                 if (!mem_cgroup_local_usage(&victim->stat)) {
850                         /* this cgroup's local usage == 0 */
851                         css_put(&victim->css);
852                         continue;
853                 }
854                 /* we use swappiness of local cgroup */
855                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
856                                                    get_swappiness(victim));
857                 css_put(&victim->css);
858                 /*
859                  * At shrinking usage, we can't check we should stop here or
860                  * reclaim more. It's depends on callers. last_scanned_child
861                  * will work enough for keeping fairness under tree.
862                  */
863                 if (shrink)
864                         return ret;
865                 total += ret;
866                 if (mem_cgroup_check_under_limit(root_mem))
867                         return 1 + total;
868         }
869         return total;
870 }
871
872 bool mem_cgroup_oom_called(struct task_struct *task)
873 {
874         bool ret = false;
875         struct mem_cgroup *mem;
876         struct mm_struct *mm;
877
878         rcu_read_lock();
879         mm = task->mm;
880         if (!mm)
881                 mm = &init_mm;
882         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
883         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
884                 ret = true;
885         rcu_read_unlock();
886         return ret;
887 }
888
889 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
890 {
891         mem->last_oom_jiffies = jiffies;
892         return 0;
893 }
894
895 static void record_last_oom(struct mem_cgroup *mem)
896 {
897         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
898 }
899
900
901 /*
902  * Unlike exported interface, "oom" parameter is added. if oom==true,
903  * oom-killer can be invoked.
904  */
905 static int __mem_cgroup_try_charge(struct mm_struct *mm,
906                         gfp_t gfp_mask, struct mem_cgroup **memcg,
907                         bool oom)
908 {
909         struct mem_cgroup *mem, *mem_over_limit;
910         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
911         struct res_counter *fail_res;
912
913         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
914                 /* Don't account this! */
915                 *memcg = NULL;
916                 return 0;
917         }
918
919         /*
920          * We always charge the cgroup the mm_struct belongs to.
921          * The mm_struct's mem_cgroup changes on task migration if the
922          * thread group leader migrates. It's possible that mm is not
923          * set, if so charge the init_mm (happens for pagecache usage).
924          */
925         mem = *memcg;
926         if (likely(!mem)) {
927                 mem = try_get_mem_cgroup_from_mm(mm);
928                 *memcg = mem;
929         } else {
930                 css_get(&mem->css);
931         }
932         if (unlikely(!mem))
933                 return 0;
934
935         VM_BUG_ON(!mem || mem_cgroup_is_obsolete(mem));
936
937         while (1) {
938                 int ret;
939                 bool noswap = false;
940
941                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
942                 if (likely(!ret)) {
943                         if (!do_swap_account)
944                                 break;
945                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
946                                                         &fail_res);
947                         if (likely(!ret))
948                                 break;
949                         /* mem+swap counter fails */
950                         res_counter_uncharge(&mem->res, PAGE_SIZE);
951                         noswap = true;
952                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
953                                                                         memsw);
954                 } else
955                         /* mem counter fails */
956                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
957                                                                         res);
958
959                 if (!(gfp_mask & __GFP_WAIT))
960                         goto nomem;
961
962                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
963                                                         noswap, false);
964                 if (ret)
965                         continue;
966
967                 /*
968                  * try_to_free_mem_cgroup_pages() might not give us a full
969                  * picture of reclaim. Some pages are reclaimed and might be
970                  * moved to swap cache or just unmapped from the cgroup.
971                  * Check the limit again to see if the reclaim reduced the
972                  * current usage of the cgroup before giving up
973                  *
974                  */
975                 if (mem_cgroup_check_under_limit(mem_over_limit))
976                         continue;
977
978                 if (!nr_retries--) {
979                         if (oom) {
980                                 mutex_lock(&memcg_tasklist);
981                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
982                                 mutex_unlock(&memcg_tasklist);
983                                 record_last_oom(mem_over_limit);
984                         }
985                         goto nomem;
986                 }
987         }
988         return 0;
989 nomem:
990         css_put(&mem->css);
991         return -ENOMEM;
992 }
993
994
995 /*
996  * A helper function to get mem_cgroup from ID. must be called under
997  * rcu_read_lock(). The caller must check css_is_removed() or some if
998  * it's concern. (dropping refcnt from swap can be called against removed
999  * memcg.)
1000  */
1001 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1002 {
1003         struct cgroup_subsys_state *css;
1004
1005         /* ID 0 is unused ID */
1006         if (!id)
1007                 return NULL;
1008         css = css_lookup(&mem_cgroup_subsys, id);
1009         if (!css)
1010                 return NULL;
1011         return container_of(css, struct mem_cgroup, css);
1012 }
1013
1014 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1015 {
1016         struct mem_cgroup *mem;
1017         struct page_cgroup *pc;
1018         unsigned short id;
1019         swp_entry_t ent;
1020
1021         VM_BUG_ON(!PageLocked(page));
1022
1023         if (!PageSwapCache(page))
1024                 return NULL;
1025
1026         pc = lookup_page_cgroup(page);
1027         lock_page_cgroup(pc);
1028         if (PageCgroupUsed(pc)) {
1029                 mem = pc->mem_cgroup;
1030                 if (mem && !css_tryget(&mem->css))
1031                         mem = NULL;
1032         } else {
1033                 ent.val = page_private(page);
1034                 id = lookup_swap_cgroup(ent);
1035                 rcu_read_lock();
1036                 mem = mem_cgroup_lookup(id);
1037                 if (mem && !css_tryget(&mem->css))
1038                         mem = NULL;
1039                 rcu_read_unlock();
1040         }
1041         unlock_page_cgroup(pc);
1042         return mem;
1043 }
1044
1045 /*
1046  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1047  * USED state. If already USED, uncharge and return.
1048  */
1049
1050 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1051                                      struct page_cgroup *pc,
1052                                      enum charge_type ctype)
1053 {
1054         /* try_charge() can return NULL to *memcg, taking care of it. */
1055         if (!mem)
1056                 return;
1057
1058         lock_page_cgroup(pc);
1059         if (unlikely(PageCgroupUsed(pc))) {
1060                 unlock_page_cgroup(pc);
1061                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1062                 if (do_swap_account)
1063                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1064                 css_put(&mem->css);
1065                 return;
1066         }
1067         pc->mem_cgroup = mem;
1068         smp_wmb();
1069         pc->flags = pcg_default_flags[ctype];
1070
1071         mem_cgroup_charge_statistics(mem, pc, true);
1072
1073         unlock_page_cgroup(pc);
1074 }
1075
1076 /**
1077  * mem_cgroup_move_account - move account of the page
1078  * @pc: page_cgroup of the page.
1079  * @from: mem_cgroup which the page is moved from.
1080  * @to: mem_cgroup which the page is moved to. @from != @to.
1081  *
1082  * The caller must confirm following.
1083  * - page is not on LRU (isolate_page() is useful.)
1084  *
1085  * returns 0 at success,
1086  * returns -EBUSY when lock is busy or "pc" is unstable.
1087  *
1088  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1089  * new cgroup. It should be done by a caller.
1090  */
1091
1092 static int mem_cgroup_move_account(struct page_cgroup *pc,
1093         struct mem_cgroup *from, struct mem_cgroup *to)
1094 {
1095         struct mem_cgroup_per_zone *from_mz, *to_mz;
1096         int nid, zid;
1097         int ret = -EBUSY;
1098
1099         VM_BUG_ON(from == to);
1100         VM_BUG_ON(PageLRU(pc->page));
1101
1102         nid = page_cgroup_nid(pc);
1103         zid = page_cgroup_zid(pc);
1104         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1105         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1106
1107         if (!trylock_page_cgroup(pc))
1108                 return ret;
1109
1110         if (!PageCgroupUsed(pc))
1111                 goto out;
1112
1113         if (pc->mem_cgroup != from)
1114                 goto out;
1115
1116         res_counter_uncharge(&from->res, PAGE_SIZE);
1117         mem_cgroup_charge_statistics(from, pc, false);
1118         if (do_swap_account)
1119                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1120         css_put(&from->css);
1121
1122         css_get(&to->css);
1123         pc->mem_cgroup = to;
1124         mem_cgroup_charge_statistics(to, pc, true);
1125         ret = 0;
1126 out:
1127         unlock_page_cgroup(pc);
1128         return ret;
1129 }
1130
1131 /*
1132  * move charges to its parent.
1133  */
1134
1135 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1136                                   struct mem_cgroup *child,
1137                                   gfp_t gfp_mask)
1138 {
1139         struct page *page = pc->page;
1140         struct cgroup *cg = child->css.cgroup;
1141         struct cgroup *pcg = cg->parent;
1142         struct mem_cgroup *parent;
1143         int ret;
1144
1145         /* Is ROOT ? */
1146         if (!pcg)
1147                 return -EINVAL;
1148
1149
1150         parent = mem_cgroup_from_cont(pcg);
1151
1152
1153         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1154         if (ret || !parent)
1155                 return ret;
1156
1157         if (!get_page_unless_zero(page)) {
1158                 ret = -EBUSY;
1159                 goto uncharge;
1160         }
1161
1162         ret = isolate_lru_page(page);
1163
1164         if (ret)
1165                 goto cancel;
1166
1167         ret = mem_cgroup_move_account(pc, child, parent);
1168
1169         putback_lru_page(page);
1170         if (!ret) {
1171                 put_page(page);
1172                 /* drop extra refcnt by try_charge() */
1173                 css_put(&parent->css);
1174                 return 0;
1175         }
1176
1177 cancel:
1178         put_page(page);
1179 uncharge:
1180         /* drop extra refcnt by try_charge() */
1181         css_put(&parent->css);
1182         /* uncharge if move fails */
1183         res_counter_uncharge(&parent->res, PAGE_SIZE);
1184         if (do_swap_account)
1185                 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1186         return ret;
1187 }
1188
1189 /*
1190  * Charge the memory controller for page usage.
1191  * Return
1192  * 0 if the charge was successful
1193  * < 0 if the cgroup is over its limit
1194  */
1195 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1196                                 gfp_t gfp_mask, enum charge_type ctype,
1197                                 struct mem_cgroup *memcg)
1198 {
1199         struct mem_cgroup *mem;
1200         struct page_cgroup *pc;
1201         int ret;
1202
1203         pc = lookup_page_cgroup(page);
1204         /* can happen at boot */
1205         if (unlikely(!pc))
1206                 return 0;
1207         prefetchw(pc);
1208
1209         mem = memcg;
1210         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1211         if (ret || !mem)
1212                 return ret;
1213
1214         __mem_cgroup_commit_charge(mem, pc, ctype);
1215         return 0;
1216 }
1217
1218 int mem_cgroup_newpage_charge(struct page *page,
1219                               struct mm_struct *mm, gfp_t gfp_mask)
1220 {
1221         if (mem_cgroup_disabled())
1222                 return 0;
1223         if (PageCompound(page))
1224                 return 0;
1225         /*
1226          * If already mapped, we don't have to account.
1227          * If page cache, page->mapping has address_space.
1228          * But page->mapping may have out-of-use anon_vma pointer,
1229          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1230          * is NULL.
1231          */
1232         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1233                 return 0;
1234         if (unlikely(!mm))
1235                 mm = &init_mm;
1236         return mem_cgroup_charge_common(page, mm, gfp_mask,
1237                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1238 }
1239
1240 static void
1241 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1242                                         enum charge_type ctype);
1243
1244 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1245                                 gfp_t gfp_mask)
1246 {
1247         struct mem_cgroup *mem = NULL;
1248         int ret;
1249
1250         if (mem_cgroup_disabled())
1251                 return 0;
1252         if (PageCompound(page))
1253                 return 0;
1254         /*
1255          * Corner case handling. This is called from add_to_page_cache()
1256          * in usual. But some FS (shmem) precharges this page before calling it
1257          * and call add_to_page_cache() with GFP_NOWAIT.
1258          *
1259          * For GFP_NOWAIT case, the page may be pre-charged before calling
1260          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1261          * charge twice. (It works but has to pay a bit larger cost.)
1262          * And when the page is SwapCache, it should take swap information
1263          * into account. This is under lock_page() now.
1264          */
1265         if (!(gfp_mask & __GFP_WAIT)) {
1266                 struct page_cgroup *pc;
1267
1268
1269                 pc = lookup_page_cgroup(page);
1270                 if (!pc)
1271                         return 0;
1272                 lock_page_cgroup(pc);
1273                 if (PageCgroupUsed(pc)) {
1274                         unlock_page_cgroup(pc);
1275                         return 0;
1276                 }
1277                 unlock_page_cgroup(pc);
1278         }
1279
1280         if (unlikely(!mm && !mem))
1281                 mm = &init_mm;
1282
1283         if (page_is_file_cache(page))
1284                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1285                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1286
1287         /* shmem */
1288         if (PageSwapCache(page)) {
1289                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1290                 if (!ret)
1291                         __mem_cgroup_commit_charge_swapin(page, mem,
1292                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1293         } else
1294                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1295                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1296
1297         return ret;
1298 }
1299
1300 /*
1301  * While swap-in, try_charge -> commit or cancel, the page is locked.
1302  * And when try_charge() successfully returns, one refcnt to memcg without
1303  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1304  * "commit()" or removed by "cancel()"
1305  */
1306 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1307                                  struct page *page,
1308                                  gfp_t mask, struct mem_cgroup **ptr)
1309 {
1310         struct mem_cgroup *mem;
1311         int ret;
1312
1313         if (mem_cgroup_disabled())
1314                 return 0;
1315
1316         if (!do_swap_account)
1317                 goto charge_cur_mm;
1318         /*
1319          * A racing thread's fault, or swapoff, may have already updated
1320          * the pte, and even removed page from swap cache: return success
1321          * to go on to do_swap_page()'s pte_same() test, which should fail.
1322          */
1323         if (!PageSwapCache(page))
1324                 return 0;
1325         mem = try_get_mem_cgroup_from_swapcache(page);
1326         if (!mem)
1327                 goto charge_cur_mm;
1328         *ptr = mem;
1329         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1330         /* drop extra refcnt from tryget */
1331         css_put(&mem->css);
1332         return ret;
1333 charge_cur_mm:
1334         if (unlikely(!mm))
1335                 mm = &init_mm;
1336         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1337 }
1338
1339 static void
1340 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1341                                         enum charge_type ctype)
1342 {
1343         struct page_cgroup *pc;
1344
1345         if (mem_cgroup_disabled())
1346                 return;
1347         if (!ptr)
1348                 return;
1349         pc = lookup_page_cgroup(page);
1350         mem_cgroup_lru_del_before_commit_swapcache(page);
1351         __mem_cgroup_commit_charge(ptr, pc, ctype);
1352         mem_cgroup_lru_add_after_commit_swapcache(page);
1353         /*
1354          * Now swap is on-memory. This means this page may be
1355          * counted both as mem and swap....double count.
1356          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1357          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1358          * may call delete_from_swap_cache() before reach here.
1359          */
1360         if (do_swap_account && PageSwapCache(page)) {
1361                 swp_entry_t ent = {.val = page_private(page)};
1362                 unsigned short id;
1363                 struct mem_cgroup *memcg;
1364
1365                 id = swap_cgroup_record(ent, 0);
1366                 rcu_read_lock();
1367                 memcg = mem_cgroup_lookup(id);
1368                 if (memcg) {
1369                         /*
1370                          * This recorded memcg can be obsolete one. So, avoid
1371                          * calling css_tryget
1372                          */
1373                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1374                         mem_cgroup_put(memcg);
1375                 }
1376                 rcu_read_unlock();
1377         }
1378         /* add this page(page_cgroup) to the LRU we want. */
1379
1380 }
1381
1382 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1383 {
1384         __mem_cgroup_commit_charge_swapin(page, ptr,
1385                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1386 }
1387
1388 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1389 {
1390         if (mem_cgroup_disabled())
1391                 return;
1392         if (!mem)
1393                 return;
1394         res_counter_uncharge(&mem->res, PAGE_SIZE);
1395         if (do_swap_account)
1396                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1397         css_put(&mem->css);
1398 }
1399
1400
1401 /*
1402  * uncharge if !page_mapped(page)
1403  */
1404 static struct mem_cgroup *
1405 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1406 {
1407         struct page_cgroup *pc;
1408         struct mem_cgroup *mem = NULL;
1409         struct mem_cgroup_per_zone *mz;
1410
1411         if (mem_cgroup_disabled())
1412                 return NULL;
1413
1414         if (PageSwapCache(page))
1415                 return NULL;
1416
1417         /*
1418          * Check if our page_cgroup is valid
1419          */
1420         pc = lookup_page_cgroup(page);
1421         if (unlikely(!pc || !PageCgroupUsed(pc)))
1422                 return NULL;
1423
1424         lock_page_cgroup(pc);
1425
1426         mem = pc->mem_cgroup;
1427
1428         if (!PageCgroupUsed(pc))
1429                 goto unlock_out;
1430
1431         switch (ctype) {
1432         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1433                 if (page_mapped(page))
1434                         goto unlock_out;
1435                 break;
1436         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1437                 if (!PageAnon(page)) {  /* Shared memory */
1438                         if (page->mapping && !page_is_file_cache(page))
1439                                 goto unlock_out;
1440                 } else if (page_mapped(page)) /* Anon */
1441                                 goto unlock_out;
1442                 break;
1443         default:
1444                 break;
1445         }
1446
1447         res_counter_uncharge(&mem->res, PAGE_SIZE);
1448         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1449                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1450         mem_cgroup_charge_statistics(mem, pc, false);
1451
1452         ClearPageCgroupUsed(pc);
1453         /*
1454          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1455          * freed from LRU. This is safe because uncharged page is expected not
1456          * to be reused (freed soon). Exception is SwapCache, it's handled by
1457          * special functions.
1458          */
1459
1460         mz = page_cgroup_zoneinfo(pc);
1461         unlock_page_cgroup(pc);
1462
1463         /* at swapout, this memcg will be accessed to record to swap */
1464         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1465                 css_put(&mem->css);
1466
1467         return mem;
1468
1469 unlock_out:
1470         unlock_page_cgroup(pc);
1471         return NULL;
1472 }
1473
1474 void mem_cgroup_uncharge_page(struct page *page)
1475 {
1476         /* early check. */
1477         if (page_mapped(page))
1478                 return;
1479         if (page->mapping && !PageAnon(page))
1480                 return;
1481         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1482 }
1483
1484 void mem_cgroup_uncharge_cache_page(struct page *page)
1485 {
1486         VM_BUG_ON(page_mapped(page));
1487         VM_BUG_ON(page->mapping);
1488         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1489 }
1490
1491 #ifdef CONFIG_SWAP
1492 /*
1493  * called after __delete_from_swap_cache() and drop "page" account.
1494  * memcg information is recorded to swap_cgroup of "ent"
1495  */
1496 void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent)
1497 {
1498         struct mem_cgroup *memcg;
1499
1500         memcg = __mem_cgroup_uncharge_common(page,
1501                                         MEM_CGROUP_CHARGE_TYPE_SWAPOUT);
1502         /* record memcg information */
1503         if (do_swap_account && memcg) {
1504                 swap_cgroup_record(ent, css_id(&memcg->css));
1505                 mem_cgroup_get(memcg);
1506         }
1507         if (memcg)
1508                 css_put(&memcg->css);
1509 }
1510 #endif
1511
1512 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1513 /*
1514  * called from swap_entry_free(). remove record in swap_cgroup and
1515  * uncharge "memsw" account.
1516  */
1517 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1518 {
1519         struct mem_cgroup *memcg;
1520         unsigned short id;
1521
1522         if (!do_swap_account)
1523                 return;
1524
1525         id = swap_cgroup_record(ent, 0);
1526         rcu_read_lock();
1527         memcg = mem_cgroup_lookup(id);
1528         if (memcg) {
1529                 /*
1530                  * We uncharge this because swap is freed.
1531                  * This memcg can be obsolete one. We avoid calling css_tryget
1532                  */
1533                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1534                 mem_cgroup_put(memcg);
1535         }
1536         rcu_read_unlock();
1537 }
1538 #endif
1539
1540 /*
1541  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1542  * page belongs to.
1543  */
1544 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1545 {
1546         struct page_cgroup *pc;
1547         struct mem_cgroup *mem = NULL;
1548         int ret = 0;
1549
1550         if (mem_cgroup_disabled())
1551                 return 0;
1552
1553         pc = lookup_page_cgroup(page);
1554         lock_page_cgroup(pc);
1555         if (PageCgroupUsed(pc)) {
1556                 mem = pc->mem_cgroup;
1557                 css_get(&mem->css);
1558         }
1559         unlock_page_cgroup(pc);
1560
1561         if (mem) {
1562                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
1563                 css_put(&mem->css);
1564         }
1565         *ptr = mem;
1566         return ret;
1567 }
1568
1569 /* remove redundant charge if migration failed*/
1570 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1571                 struct page *oldpage, struct page *newpage)
1572 {
1573         struct page *target, *unused;
1574         struct page_cgroup *pc;
1575         enum charge_type ctype;
1576
1577         if (!mem)
1578                 return;
1579
1580         /* at migration success, oldpage->mapping is NULL. */
1581         if (oldpage->mapping) {
1582                 target = oldpage;
1583                 unused = NULL;
1584         } else {
1585                 target = newpage;
1586                 unused = oldpage;
1587         }
1588
1589         if (PageAnon(target))
1590                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1591         else if (page_is_file_cache(target))
1592                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1593         else
1594                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1595
1596         /* unused page is not on radix-tree now. */
1597         if (unused)
1598                 __mem_cgroup_uncharge_common(unused, ctype);
1599
1600         pc = lookup_page_cgroup(target);
1601         /*
1602          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1603          * So, double-counting is effectively avoided.
1604          */
1605         __mem_cgroup_commit_charge(mem, pc, ctype);
1606
1607         /*
1608          * Both of oldpage and newpage are still under lock_page().
1609          * Then, we don't have to care about race in radix-tree.
1610          * But we have to be careful that this page is unmapped or not.
1611          *
1612          * There is a case for !page_mapped(). At the start of
1613          * migration, oldpage was mapped. But now, it's zapped.
1614          * But we know *target* page is not freed/reused under us.
1615          * mem_cgroup_uncharge_page() does all necessary checks.
1616          */
1617         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1618                 mem_cgroup_uncharge_page(target);
1619 }
1620
1621 /*
1622  * A call to try to shrink memory usage on charge failure at shmem's swapin.
1623  * Calling hierarchical_reclaim is not enough because we should update
1624  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
1625  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
1626  * not from the memcg which this page would be charged to.
1627  * try_charge_swapin does all of these works properly.
1628  */
1629 int mem_cgroup_shmem_charge_fallback(struct page *page,
1630                             struct mm_struct *mm,
1631                             gfp_t gfp_mask)
1632 {
1633         struct mem_cgroup *mem = NULL;
1634         int ret;
1635
1636         if (mem_cgroup_disabled())
1637                 return 0;
1638
1639         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1640         if (!ret)
1641                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
1642
1643         return ret;
1644 }
1645
1646 static DEFINE_MUTEX(set_limit_mutex);
1647
1648 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1649                                 unsigned long long val)
1650 {
1651         int retry_count;
1652         int progress;
1653         u64 memswlimit;
1654         int ret = 0;
1655         int children = mem_cgroup_count_children(memcg);
1656         u64 curusage, oldusage;
1657
1658         /*
1659          * For keeping hierarchical_reclaim simple, how long we should retry
1660          * is depends on callers. We set our retry-count to be function
1661          * of # of children which we should visit in this loop.
1662          */
1663         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
1664
1665         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1666
1667         while (retry_count) {
1668                 if (signal_pending(current)) {
1669                         ret = -EINTR;
1670                         break;
1671                 }
1672                 /*
1673                  * Rather than hide all in some function, I do this in
1674                  * open coded manner. You see what this really does.
1675                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1676                  */
1677                 mutex_lock(&set_limit_mutex);
1678                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1679                 if (memswlimit < val) {
1680                         ret = -EINVAL;
1681                         mutex_unlock(&set_limit_mutex);
1682                         break;
1683                 }
1684                 ret = res_counter_set_limit(&memcg->res, val);
1685                 mutex_unlock(&set_limit_mutex);
1686
1687                 if (!ret)
1688                         break;
1689
1690                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
1691                                                    false, true);
1692                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1693                 /* Usage is reduced ? */
1694                 if (curusage >= oldusage)
1695                         retry_count--;
1696                 else
1697                         oldusage = curusage;
1698         }
1699
1700         return ret;
1701 }
1702
1703 int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
1704                                 unsigned long long val)
1705 {
1706         int retry_count;
1707         u64 memlimit, oldusage, curusage;
1708         int children = mem_cgroup_count_children(memcg);
1709         int ret = -EBUSY;
1710
1711         if (!do_swap_account)
1712                 return -EINVAL;
1713         /* see mem_cgroup_resize_res_limit */
1714         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
1715         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1716         while (retry_count) {
1717                 if (signal_pending(current)) {
1718                         ret = -EINTR;
1719                         break;
1720                 }
1721                 /*
1722                  * Rather than hide all in some function, I do this in
1723                  * open coded manner. You see what this really does.
1724                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1725                  */
1726                 mutex_lock(&set_limit_mutex);
1727                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1728                 if (memlimit > val) {
1729                         ret = -EINVAL;
1730                         mutex_unlock(&set_limit_mutex);
1731                         break;
1732                 }
1733                 ret = res_counter_set_limit(&memcg->memsw, val);
1734                 mutex_unlock(&set_limit_mutex);
1735
1736                 if (!ret)
1737                         break;
1738
1739                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
1740                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
1741                 /* Usage is reduced ? */
1742                 if (curusage >= oldusage)
1743                         retry_count--;
1744                 else
1745                         oldusage = curusage;
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * This routine traverse page_cgroup in given list and drop them all.
1752  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
1753  */
1754 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
1755                                 int node, int zid, enum lru_list lru)
1756 {
1757         struct zone *zone;
1758         struct mem_cgroup_per_zone *mz;
1759         struct page_cgroup *pc, *busy;
1760         unsigned long flags, loop;
1761         struct list_head *list;
1762         int ret = 0;
1763
1764         zone = &NODE_DATA(node)->node_zones[zid];
1765         mz = mem_cgroup_zoneinfo(mem, node, zid);
1766         list = &mz->lists[lru];
1767
1768         loop = MEM_CGROUP_ZSTAT(mz, lru);
1769         /* give some margin against EBUSY etc...*/
1770         loop += 256;
1771         busy = NULL;
1772         while (loop--) {
1773                 ret = 0;
1774                 spin_lock_irqsave(&zone->lru_lock, flags);
1775                 if (list_empty(list)) {
1776                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1777                         break;
1778                 }
1779                 pc = list_entry(list->prev, struct page_cgroup, lru);
1780                 if (busy == pc) {
1781                         list_move(&pc->lru, list);
1782                         busy = 0;
1783                         spin_unlock_irqrestore(&zone->lru_lock, flags);
1784                         continue;
1785                 }
1786                 spin_unlock_irqrestore(&zone->lru_lock, flags);
1787
1788                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
1789                 if (ret == -ENOMEM)
1790                         break;
1791
1792                 if (ret == -EBUSY || ret == -EINVAL) {
1793                         /* found lock contention or "pc" is obsolete. */
1794                         busy = pc;
1795                         cond_resched();
1796                 } else
1797                         busy = NULL;
1798         }
1799
1800         if (!ret && !list_empty(list))
1801                 return -EBUSY;
1802         return ret;
1803 }
1804
1805 /*
1806  * make mem_cgroup's charge to be 0 if there is no task.
1807  * This enables deleting this mem_cgroup.
1808  */
1809 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
1810 {
1811         int ret;
1812         int node, zid, shrink;
1813         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1814         struct cgroup *cgrp = mem->css.cgroup;
1815
1816         css_get(&mem->css);
1817
1818         shrink = 0;
1819         /* should free all ? */
1820         if (free_all)
1821                 goto try_to_free;
1822 move_account:
1823         while (mem->res.usage > 0) {
1824                 ret = -EBUSY;
1825                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
1826                         goto out;
1827                 ret = -EINTR;
1828                 if (signal_pending(current))
1829                         goto out;
1830                 /* This is for making all *used* pages to be on LRU. */
1831                 lru_add_drain_all();
1832                 ret = 0;
1833                 for_each_node_state(node, N_HIGH_MEMORY) {
1834                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
1835                                 enum lru_list l;
1836                                 for_each_lru(l) {
1837                                         ret = mem_cgroup_force_empty_list(mem,
1838                                                         node, zid, l);
1839                                         if (ret)
1840                                                 break;
1841                                 }
1842                         }
1843                         if (ret)
1844                                 break;
1845                 }
1846                 /* it seems parent cgroup doesn't have enough mem */
1847                 if (ret == -ENOMEM)
1848                         goto try_to_free;
1849                 cond_resched();
1850         }
1851         ret = 0;
1852 out:
1853         css_put(&mem->css);
1854         return ret;
1855
1856 try_to_free:
1857         /* returns EBUSY if there is a task or if we come here twice. */
1858         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
1859                 ret = -EBUSY;
1860                 goto out;
1861         }
1862         /* we call try-to-free pages for make this cgroup empty */
1863         lru_add_drain_all();
1864         /* try to free all pages in this cgroup */
1865         shrink = 1;
1866         while (nr_retries && mem->res.usage > 0) {
1867                 int progress;
1868
1869                 if (signal_pending(current)) {
1870                         ret = -EINTR;
1871                         goto out;
1872                 }
1873                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
1874                                                 false, get_swappiness(mem));
1875                 if (!progress) {
1876                         nr_retries--;
1877                         /* maybe some writeback is necessary */
1878                         congestion_wait(WRITE, HZ/10);
1879                 }
1880
1881         }
1882         lru_add_drain();
1883         /* try move_account...there may be some *locked* pages. */
1884         if (mem->res.usage)
1885                 goto move_account;
1886         ret = 0;
1887         goto out;
1888 }
1889
1890 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
1891 {
1892         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
1893 }
1894
1895
1896 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
1897 {
1898         return mem_cgroup_from_cont(cont)->use_hierarchy;
1899 }
1900
1901 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
1902                                         u64 val)
1903 {
1904         int retval = 0;
1905         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1906         struct cgroup *parent = cont->parent;
1907         struct mem_cgroup *parent_mem = NULL;
1908
1909         if (parent)
1910                 parent_mem = mem_cgroup_from_cont(parent);
1911
1912         cgroup_lock();
1913         /*
1914          * If parent's use_hiearchy is set, we can't make any modifications
1915          * in the child subtrees. If it is unset, then the change can
1916          * occur, provided the current cgroup has no children.
1917          *
1918          * For the root cgroup, parent_mem is NULL, we allow value to be
1919          * set if there are no children.
1920          */
1921         if ((!parent_mem || !parent_mem->use_hierarchy) &&
1922                                 (val == 1 || val == 0)) {
1923                 if (list_empty(&cont->children))
1924                         mem->use_hierarchy = val;
1925                 else
1926                         retval = -EBUSY;
1927         } else
1928                 retval = -EINVAL;
1929         cgroup_unlock();
1930
1931         return retval;
1932 }
1933
1934 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
1935 {
1936         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1937         u64 val = 0;
1938         int type, name;
1939
1940         type = MEMFILE_TYPE(cft->private);
1941         name = MEMFILE_ATTR(cft->private);
1942         switch (type) {
1943         case _MEM:
1944                 val = res_counter_read_u64(&mem->res, name);
1945                 break;
1946         case _MEMSWAP:
1947                 if (do_swap_account)
1948                         val = res_counter_read_u64(&mem->memsw, name);
1949                 break;
1950         default:
1951                 BUG();
1952                 break;
1953         }
1954         return val;
1955 }
1956 /*
1957  * The user of this function is...
1958  * RES_LIMIT.
1959  */
1960 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
1961                             const char *buffer)
1962 {
1963         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
1964         int type, name;
1965         unsigned long long val;
1966         int ret;
1967
1968         type = MEMFILE_TYPE(cft->private);
1969         name = MEMFILE_ATTR(cft->private);
1970         switch (name) {
1971         case RES_LIMIT:
1972                 /* This function does all necessary parse...reuse it */
1973                 ret = res_counter_memparse_write_strategy(buffer, &val);
1974                 if (ret)
1975                         break;
1976                 if (type == _MEM)
1977                         ret = mem_cgroup_resize_limit(memcg, val);
1978                 else
1979                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
1980                 break;
1981         default:
1982                 ret = -EINVAL; /* should be BUG() ? */
1983                 break;
1984         }
1985         return ret;
1986 }
1987
1988 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
1989                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
1990 {
1991         struct cgroup *cgroup;
1992         unsigned long long min_limit, min_memsw_limit, tmp;
1993
1994         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1995         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1996         cgroup = memcg->css.cgroup;
1997         if (!memcg->use_hierarchy)
1998                 goto out;
1999
2000         while (cgroup->parent) {
2001                 cgroup = cgroup->parent;
2002                 memcg = mem_cgroup_from_cont(cgroup);
2003                 if (!memcg->use_hierarchy)
2004                         break;
2005                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2006                 min_limit = min(min_limit, tmp);
2007                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2008                 min_memsw_limit = min(min_memsw_limit, tmp);
2009         }
2010 out:
2011         *mem_limit = min_limit;
2012         *memsw_limit = min_memsw_limit;
2013         return;
2014 }
2015
2016 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2017 {
2018         struct mem_cgroup *mem;
2019         int type, name;
2020
2021         mem = mem_cgroup_from_cont(cont);
2022         type = MEMFILE_TYPE(event);
2023         name = MEMFILE_ATTR(event);
2024         switch (name) {
2025         case RES_MAX_USAGE:
2026                 if (type == _MEM)
2027                         res_counter_reset_max(&mem->res);
2028                 else
2029                         res_counter_reset_max(&mem->memsw);
2030                 break;
2031         case RES_FAILCNT:
2032                 if (type == _MEM)
2033                         res_counter_reset_failcnt(&mem->res);
2034                 else
2035                         res_counter_reset_failcnt(&mem->memsw);
2036                 break;
2037         }
2038         return 0;
2039 }
2040
2041
2042 /* For read statistics */
2043 enum {
2044         MCS_CACHE,
2045         MCS_RSS,
2046         MCS_PGPGIN,
2047         MCS_PGPGOUT,
2048         MCS_INACTIVE_ANON,
2049         MCS_ACTIVE_ANON,
2050         MCS_INACTIVE_FILE,
2051         MCS_ACTIVE_FILE,
2052         MCS_UNEVICTABLE,
2053         NR_MCS_STAT,
2054 };
2055
2056 struct mcs_total_stat {
2057         s64 stat[NR_MCS_STAT];
2058 };
2059
2060 struct {
2061         char *local_name;
2062         char *total_name;
2063 } memcg_stat_strings[NR_MCS_STAT] = {
2064         {"cache", "total_cache"},
2065         {"rss", "total_rss"},
2066         {"pgpgin", "total_pgpgin"},
2067         {"pgpgout", "total_pgpgout"},
2068         {"inactive_anon", "total_inactive_anon"},
2069         {"active_anon", "total_active_anon"},
2070         {"inactive_file", "total_inactive_file"},
2071         {"active_file", "total_active_file"},
2072         {"unevictable", "total_unevictable"}
2073 };
2074
2075
2076 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2077 {
2078         struct mcs_total_stat *s = data;
2079         s64 val;
2080
2081         /* per cpu stat */
2082         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2083         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2084         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2085         s->stat[MCS_RSS] += val * PAGE_SIZE;
2086         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2087         s->stat[MCS_PGPGIN] += val;
2088         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2089         s->stat[MCS_PGPGOUT] += val;
2090
2091         /* per zone stat */
2092         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2093         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2094         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2095         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2096         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2097         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2098         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2099         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2100         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2101         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2102         return 0;
2103 }
2104
2105 static void
2106 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2107 {
2108         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2109 }
2110
2111 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2112                                  struct cgroup_map_cb *cb)
2113 {
2114         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2115         struct mcs_total_stat mystat;
2116         int i;
2117
2118         memset(&mystat, 0, sizeof(mystat));
2119         mem_cgroup_get_local_stat(mem_cont, &mystat);
2120
2121         for (i = 0; i < NR_MCS_STAT; i++)
2122                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2123
2124         /* Hierarchical information */
2125         {
2126                 unsigned long long limit, memsw_limit;
2127                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2128                 cb->fill(cb, "hierarchical_memory_limit", limit);
2129                 if (do_swap_account)
2130                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2131         }
2132
2133         memset(&mystat, 0, sizeof(mystat));
2134         mem_cgroup_get_total_stat(mem_cont, &mystat);
2135         for (i = 0; i < NR_MCS_STAT; i++)
2136                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2137
2138
2139 #ifdef CONFIG_DEBUG_VM
2140         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2141
2142         {
2143                 int nid, zid;
2144                 struct mem_cgroup_per_zone *mz;
2145                 unsigned long recent_rotated[2] = {0, 0};
2146                 unsigned long recent_scanned[2] = {0, 0};
2147
2148                 for_each_online_node(nid)
2149                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2150                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2151
2152                                 recent_rotated[0] +=
2153                                         mz->reclaim_stat.recent_rotated[0];
2154                                 recent_rotated[1] +=
2155                                         mz->reclaim_stat.recent_rotated[1];
2156                                 recent_scanned[0] +=
2157                                         mz->reclaim_stat.recent_scanned[0];
2158                                 recent_scanned[1] +=
2159                                         mz->reclaim_stat.recent_scanned[1];
2160                         }
2161                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2162                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2163                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2164                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2165         }
2166 #endif
2167
2168         return 0;
2169 }
2170
2171 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2172 {
2173         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2174
2175         return get_swappiness(memcg);
2176 }
2177
2178 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2179                                        u64 val)
2180 {
2181         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2182         struct mem_cgroup *parent;
2183
2184         if (val > 100)
2185                 return -EINVAL;
2186
2187         if (cgrp->parent == NULL)
2188                 return -EINVAL;
2189
2190         parent = mem_cgroup_from_cont(cgrp->parent);
2191
2192         cgroup_lock();
2193
2194         /* If under hierarchy, only empty-root can set this value */
2195         if ((parent->use_hierarchy) ||
2196             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2197                 cgroup_unlock();
2198                 return -EINVAL;
2199         }
2200
2201         spin_lock(&memcg->reclaim_param_lock);
2202         memcg->swappiness = val;
2203         spin_unlock(&memcg->reclaim_param_lock);
2204
2205         cgroup_unlock();
2206
2207         return 0;
2208 }
2209
2210
2211 static struct cftype mem_cgroup_files[] = {
2212         {
2213                 .name = "usage_in_bytes",
2214                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2215                 .read_u64 = mem_cgroup_read,
2216         },
2217         {
2218                 .name = "max_usage_in_bytes",
2219                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2220                 .trigger = mem_cgroup_reset,
2221                 .read_u64 = mem_cgroup_read,
2222         },
2223         {
2224                 .name = "limit_in_bytes",
2225                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2226                 .write_string = mem_cgroup_write,
2227                 .read_u64 = mem_cgroup_read,
2228         },
2229         {
2230                 .name = "failcnt",
2231                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2232                 .trigger = mem_cgroup_reset,
2233                 .read_u64 = mem_cgroup_read,
2234         },
2235         {
2236                 .name = "stat",
2237                 .read_map = mem_control_stat_show,
2238         },
2239         {
2240                 .name = "force_empty",
2241                 .trigger = mem_cgroup_force_empty_write,
2242         },
2243         {
2244                 .name = "use_hierarchy",
2245                 .write_u64 = mem_cgroup_hierarchy_write,
2246                 .read_u64 = mem_cgroup_hierarchy_read,
2247         },
2248         {
2249                 .name = "swappiness",
2250                 .read_u64 = mem_cgroup_swappiness_read,
2251                 .write_u64 = mem_cgroup_swappiness_write,
2252         },
2253 };
2254
2255 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2256 static struct cftype memsw_cgroup_files[] = {
2257         {
2258                 .name = "memsw.usage_in_bytes",
2259                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2260                 .read_u64 = mem_cgroup_read,
2261         },
2262         {
2263                 .name = "memsw.max_usage_in_bytes",
2264                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2265                 .trigger = mem_cgroup_reset,
2266                 .read_u64 = mem_cgroup_read,
2267         },
2268         {
2269                 .name = "memsw.limit_in_bytes",
2270                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2271                 .write_string = mem_cgroup_write,
2272                 .read_u64 = mem_cgroup_read,
2273         },
2274         {
2275                 .name = "memsw.failcnt",
2276                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2277                 .trigger = mem_cgroup_reset,
2278                 .read_u64 = mem_cgroup_read,
2279         },
2280 };
2281
2282 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2283 {
2284         if (!do_swap_account)
2285                 return 0;
2286         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2287                                 ARRAY_SIZE(memsw_cgroup_files));
2288 };
2289 #else
2290 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2291 {
2292         return 0;
2293 }
2294 #endif
2295
2296 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2297 {
2298         struct mem_cgroup_per_node *pn;
2299         struct mem_cgroup_per_zone *mz;
2300         enum lru_list l;
2301         int zone, tmp = node;
2302         /*
2303          * This routine is called against possible nodes.
2304          * But it's BUG to call kmalloc() against offline node.
2305          *
2306          * TODO: this routine can waste much memory for nodes which will
2307          *       never be onlined. It's better to use memory hotplug callback
2308          *       function.
2309          */
2310         if (!node_state(node, N_NORMAL_MEMORY))
2311                 tmp = -1;
2312         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2313         if (!pn)
2314                 return 1;
2315
2316         mem->info.nodeinfo[node] = pn;
2317         memset(pn, 0, sizeof(*pn));
2318
2319         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2320                 mz = &pn->zoneinfo[zone];
2321                 for_each_lru(l)
2322                         INIT_LIST_HEAD(&mz->lists[l]);
2323         }
2324         return 0;
2325 }
2326
2327 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2328 {
2329         kfree(mem->info.nodeinfo[node]);
2330 }
2331
2332 static int mem_cgroup_size(void)
2333 {
2334         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2335         return sizeof(struct mem_cgroup) + cpustat_size;
2336 }
2337
2338 static struct mem_cgroup *mem_cgroup_alloc(void)
2339 {
2340         struct mem_cgroup *mem;
2341         int size = mem_cgroup_size();
2342
2343         if (size < PAGE_SIZE)
2344                 mem = kmalloc(size, GFP_KERNEL);
2345         else
2346                 mem = vmalloc(size);
2347
2348         if (mem)
2349                 memset(mem, 0, size);
2350         return mem;
2351 }
2352
2353 /*
2354  * At destroying mem_cgroup, references from swap_cgroup can remain.
2355  * (scanning all at force_empty is too costly...)
2356  *
2357  * Instead of clearing all references at force_empty, we remember
2358  * the number of reference from swap_cgroup and free mem_cgroup when
2359  * it goes down to 0.
2360  *
2361  * Removal of cgroup itself succeeds regardless of refs from swap.
2362  */
2363
2364 static void __mem_cgroup_free(struct mem_cgroup *mem)
2365 {
2366         int node;
2367
2368         free_css_id(&mem_cgroup_subsys, &mem->css);
2369
2370         for_each_node_state(node, N_POSSIBLE)
2371                 free_mem_cgroup_per_zone_info(mem, node);
2372
2373         if (mem_cgroup_size() < PAGE_SIZE)
2374                 kfree(mem);
2375         else
2376                 vfree(mem);
2377 }
2378
2379 static void mem_cgroup_get(struct mem_cgroup *mem)
2380 {
2381         atomic_inc(&mem->refcnt);
2382 }
2383
2384 static void mem_cgroup_put(struct mem_cgroup *mem)
2385 {
2386         if (atomic_dec_and_test(&mem->refcnt)) {
2387                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2388                 __mem_cgroup_free(mem);
2389                 if (parent)
2390                         mem_cgroup_put(parent);
2391         }
2392 }
2393
2394 /*
2395  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2396  */
2397 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2398 {
2399         if (!mem->res.parent)
2400                 return NULL;
2401         return mem_cgroup_from_res_counter(mem->res.parent, res);
2402 }
2403
2404 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2405 static void __init enable_swap_cgroup(void)
2406 {
2407         if (!mem_cgroup_disabled() && really_do_swap_account)
2408                 do_swap_account = 1;
2409 }
2410 #else
2411 static void __init enable_swap_cgroup(void)
2412 {
2413 }
2414 #endif
2415
2416 static struct cgroup_subsys_state * __ref
2417 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2418 {
2419         struct mem_cgroup *mem, *parent;
2420         long error = -ENOMEM;
2421         int node;
2422
2423         mem = mem_cgroup_alloc();
2424         if (!mem)
2425                 return ERR_PTR(error);
2426
2427         for_each_node_state(node, N_POSSIBLE)
2428                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2429                         goto free_out;
2430         /* root ? */
2431         if (cont->parent == NULL) {
2432                 enable_swap_cgroup();
2433                 parent = NULL;
2434         } else {
2435                 parent = mem_cgroup_from_cont(cont->parent);
2436                 mem->use_hierarchy = parent->use_hierarchy;
2437         }
2438
2439         if (parent && parent->use_hierarchy) {
2440                 res_counter_init(&mem->res, &parent->res);
2441                 res_counter_init(&mem->memsw, &parent->memsw);
2442                 /*
2443                  * We increment refcnt of the parent to ensure that we can
2444                  * safely access it on res_counter_charge/uncharge.
2445                  * This refcnt will be decremented when freeing this
2446                  * mem_cgroup(see mem_cgroup_put).
2447                  */
2448                 mem_cgroup_get(parent);
2449         } else {
2450                 res_counter_init(&mem->res, NULL);
2451                 res_counter_init(&mem->memsw, NULL);
2452         }
2453         mem->last_scanned_child = 0;
2454         spin_lock_init(&mem->reclaim_param_lock);
2455
2456         if (parent)
2457                 mem->swappiness = get_swappiness(parent);
2458         atomic_set(&mem->refcnt, 1);
2459         return &mem->css;
2460 free_out:
2461         __mem_cgroup_free(mem);
2462         return ERR_PTR(error);
2463 }
2464
2465 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2466                                         struct cgroup *cont)
2467 {
2468         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2469
2470         return mem_cgroup_force_empty(mem, false);
2471 }
2472
2473 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2474                                 struct cgroup *cont)
2475 {
2476         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2477
2478         mem_cgroup_put(mem);
2479 }
2480
2481 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2482                                 struct cgroup *cont)
2483 {
2484         int ret;
2485
2486         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2487                                 ARRAY_SIZE(mem_cgroup_files));
2488
2489         if (!ret)
2490                 ret = register_memsw_files(cont, ss);
2491         return ret;
2492 }
2493
2494 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2495                                 struct cgroup *cont,
2496                                 struct cgroup *old_cont,
2497                                 struct task_struct *p)
2498 {
2499         mutex_lock(&memcg_tasklist);
2500         /*
2501          * FIXME: It's better to move charges of this process from old
2502          * memcg to new memcg. But it's just on TODO-List now.
2503          */
2504         mutex_unlock(&memcg_tasklist);
2505 }
2506
2507 struct cgroup_subsys mem_cgroup_subsys = {
2508         .name = "memory",
2509         .subsys_id = mem_cgroup_subsys_id,
2510         .create = mem_cgroup_create,
2511         .pre_destroy = mem_cgroup_pre_destroy,
2512         .destroy = mem_cgroup_destroy,
2513         .populate = mem_cgroup_populate,
2514         .attach = mem_cgroup_move_task,
2515         .early_init = 0,
2516         .use_id = 1,
2517 };
2518
2519 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2520
2521 static int __init disable_swap_account(char *s)
2522 {
2523         really_do_swap_account = 0;
2524         return 1;
2525 }
2526 __setup("noswapaccount", disable_swap_account);
2527 #endif