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