memory controller: soft limit organize cgroups
[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/rbtree.h>
33 #include <linux/slab.h>
34 #include <linux/swap.h>
35 #include <linux/spinlock.h>
36 #include <linux/fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mm_inline.h>
40 #include <linux/page_cgroup.h>
41 #include "internal.h"
42
43 #include <asm/uaccess.h>
44
45 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
46 #define MEM_CGROUP_RECLAIM_RETRIES      5
47 struct mem_cgroup *root_mem_cgroup __read_mostly;
48
49 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
50 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
51 int do_swap_account __read_mostly;
52 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
53 #else
54 #define do_swap_account         (0)
55 #endif
56
57 static DEFINE_MUTEX(memcg_tasklist);    /* can be hold under cgroup_mutex */
58 #define SOFTLIMIT_EVENTS_THRESH (1000)
59
60 /*
61  * Statistics for memory cgroup.
62  */
63 enum mem_cgroup_stat_index {
64         /*
65          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
66          */
67         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
68         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
69         MEM_CGROUP_STAT_MAPPED_FILE,  /* # of pages charged as file rss */
70         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
71         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
72         MEM_CGROUP_STAT_EVENTS, /* sum of pagein + pageout for internal use */
73
74         MEM_CGROUP_STAT_NSTATS,
75 };
76
77 struct mem_cgroup_stat_cpu {
78         s64 count[MEM_CGROUP_STAT_NSTATS];
79 } ____cacheline_aligned_in_smp;
80
81 struct mem_cgroup_stat {
82         struct mem_cgroup_stat_cpu cpustat[0];
83 };
84
85 static inline void
86 __mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
87                                 enum mem_cgroup_stat_index idx)
88 {
89         stat->count[idx] = 0;
90 }
91
92 static inline s64
93 __mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
94                                 enum mem_cgroup_stat_index idx)
95 {
96         return stat->count[idx];
97 }
98
99 /*
100  * For accounting under irq disable, no need for increment preempt count.
101  */
102 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
103                 enum mem_cgroup_stat_index idx, int val)
104 {
105         stat->count[idx] += val;
106 }
107
108 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
109                 enum mem_cgroup_stat_index idx)
110 {
111         int cpu;
112         s64 ret = 0;
113         for_each_possible_cpu(cpu)
114                 ret += stat->cpustat[cpu].count[idx];
115         return ret;
116 }
117
118 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
119 {
120         s64 ret;
121
122         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
123         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
124         return ret;
125 }
126
127 /*
128  * per-zone information in memory controller.
129  */
130 struct mem_cgroup_per_zone {
131         /*
132          * spin_lock to protect the per cgroup LRU
133          */
134         struct list_head        lists[NR_LRU_LISTS];
135         unsigned long           count[NR_LRU_LISTS];
136
137         struct zone_reclaim_stat reclaim_stat;
138         struct rb_node          tree_node;      /* RB tree node */
139         unsigned long long      usage_in_excess;/* Set to the value by which */
140                                                 /* the soft limit is exceeded*/
141         bool                    on_tree;
142 };
143 /* Macro for accessing counter */
144 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
145
146 struct mem_cgroup_per_node {
147         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
148 };
149
150 struct mem_cgroup_lru_info {
151         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
152 };
153
154 /*
155  * Cgroups above their limits are maintained in a RB-Tree, independent of
156  * their hierarchy representation
157  */
158
159 struct mem_cgroup_tree_per_zone {
160         struct rb_root rb_root;
161         spinlock_t lock;
162 };
163
164 struct mem_cgroup_tree_per_node {
165         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
166 };
167
168 struct mem_cgroup_tree {
169         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
170 };
171
172 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173
174 /*
175  * The memory controller data structure. The memory controller controls both
176  * page cache and RSS per cgroup. We would eventually like to provide
177  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
178  * to help the administrator determine what knobs to tune.
179  *
180  * TODO: Add a water mark for the memory controller. Reclaim will begin when
181  * we hit the water mark. May be even add a low water mark, such that
182  * no reclaim occurs from a cgroup at it's low water mark, this is
183  * a feature that will be implemented much later in the future.
184  */
185 struct mem_cgroup {
186         struct cgroup_subsys_state css;
187         /*
188          * the counter to account for memory usage
189          */
190         struct res_counter res;
191         /*
192          * the counter to account for mem+swap usage.
193          */
194         struct res_counter memsw;
195         /*
196          * Per cgroup active and inactive list, similar to the
197          * per zone LRU lists.
198          */
199         struct mem_cgroup_lru_info info;
200
201         /*
202           protect against reclaim related member.
203         */
204         spinlock_t reclaim_param_lock;
205
206         int     prev_priority;  /* for recording reclaim priority */
207
208         /*
209          * While reclaiming in a hiearchy, we cache the last child we
210          * reclaimed from.
211          */
212         int last_scanned_child;
213         /*
214          * Should the accounting and control be hierarchical, per subtree?
215          */
216         bool use_hierarchy;
217         unsigned long   last_oom_jiffies;
218         atomic_t        refcnt;
219
220         unsigned int    swappiness;
221
222         /* set when res.limit == memsw.limit */
223         bool            memsw_is_minimum;
224
225         /*
226          * statistics. This must be placed at the end of memcg.
227          */
228         struct mem_cgroup_stat stat;
229 };
230
231 enum charge_type {
232         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
233         MEM_CGROUP_CHARGE_TYPE_MAPPED,
234         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
235         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
236         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
237         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
238         NR_CHARGE_TYPE,
239 };
240
241 /* only for here (for easy reading.) */
242 #define PCGF_CACHE      (1UL << PCG_CACHE)
243 #define PCGF_USED       (1UL << PCG_USED)
244 #define PCGF_LOCK       (1UL << PCG_LOCK)
245 /* Not used, but added here for completeness */
246 #define PCGF_ACCT       (1UL << PCG_ACCT)
247
248 /* for encoding cft->private value on file */
249 #define _MEM                    (0)
250 #define _MEMSWAP                (1)
251 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
252 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
253 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
254
255 static void mem_cgroup_get(struct mem_cgroup *mem);
256 static void mem_cgroup_put(struct mem_cgroup *mem);
257 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
258
259 static struct mem_cgroup_per_zone *
260 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
261 {
262         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
263 }
264
265 static struct mem_cgroup_per_zone *
266 page_cgroup_zoneinfo(struct page_cgroup *pc)
267 {
268         struct mem_cgroup *mem = pc->mem_cgroup;
269         int nid = page_cgroup_nid(pc);
270         int zid = page_cgroup_zid(pc);
271
272         if (!mem)
273                 return NULL;
274
275         return mem_cgroup_zoneinfo(mem, nid, zid);
276 }
277
278 static struct mem_cgroup_tree_per_zone *
279 soft_limit_tree_node_zone(int nid, int zid)
280 {
281         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
282 }
283
284 static struct mem_cgroup_tree_per_zone *
285 soft_limit_tree_from_page(struct page *page)
286 {
287         int nid = page_to_nid(page);
288         int zid = page_zonenum(page);
289
290         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
291 }
292
293 static void
294 mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
295                                 struct mem_cgroup_per_zone *mz,
296                                 struct mem_cgroup_tree_per_zone *mctz)
297 {
298         struct rb_node **p = &mctz->rb_root.rb_node;
299         struct rb_node *parent = NULL;
300         struct mem_cgroup_per_zone *mz_node;
301
302         if (mz->on_tree)
303                 return;
304
305         mz->usage_in_excess = res_counter_soft_limit_excess(&mem->res);
306         spin_lock(&mctz->lock);
307         while (*p) {
308                 parent = *p;
309                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
310                                         tree_node);
311                 if (mz->usage_in_excess < mz_node->usage_in_excess)
312                         p = &(*p)->rb_left;
313                 /*
314                  * We can't avoid mem cgroups that are over their soft
315                  * limit by the same amount
316                  */
317                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
318                         p = &(*p)->rb_right;
319         }
320         rb_link_node(&mz->tree_node, parent, p);
321         rb_insert_color(&mz->tree_node, &mctz->rb_root);
322         mz->on_tree = true;
323         spin_unlock(&mctz->lock);
324 }
325
326 static void
327 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
328                                 struct mem_cgroup_per_zone *mz,
329                                 struct mem_cgroup_tree_per_zone *mctz)
330 {
331         spin_lock(&mctz->lock);
332         rb_erase(&mz->tree_node, &mctz->rb_root);
333         mz->on_tree = false;
334         spin_unlock(&mctz->lock);
335 }
336
337 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
338 {
339         bool ret = false;
340         int cpu;
341         s64 val;
342         struct mem_cgroup_stat_cpu *cpustat;
343
344         cpu = get_cpu();
345         cpustat = &mem->stat.cpustat[cpu];
346         val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
347         if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
348                 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
349                 ret = true;
350         }
351         put_cpu();
352         return ret;
353 }
354
355 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
356 {
357         unsigned long long prev_usage_in_excess, new_usage_in_excess;
358         bool updated_tree = false;
359         struct mem_cgroup_per_zone *mz;
360         struct mem_cgroup_tree_per_zone *mctz;
361
362         mz = mem_cgroup_zoneinfo(mem, page_to_nid(page), page_zonenum(page));
363         mctz = soft_limit_tree_from_page(page);
364
365         /*
366          * We do updates in lazy mode, mem's are removed
367          * lazily from the per-zone, per-node rb tree
368          */
369         prev_usage_in_excess = mz->usage_in_excess;
370
371         new_usage_in_excess = res_counter_soft_limit_excess(&mem->res);
372         if (prev_usage_in_excess) {
373                 mem_cgroup_remove_exceeded(mem, mz, mctz);
374                 updated_tree = true;
375         }
376         if (!new_usage_in_excess)
377                 goto done;
378         mem_cgroup_insert_exceeded(mem, mz, mctz);
379
380 done:
381         if (updated_tree) {
382                 spin_lock(&mctz->lock);
383                 mz->usage_in_excess = new_usage_in_excess;
384                 spin_unlock(&mctz->lock);
385         }
386 }
387
388 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
389 {
390         int node, zone;
391         struct mem_cgroup_per_zone *mz;
392         struct mem_cgroup_tree_per_zone *mctz;
393
394         for_each_node_state(node, N_POSSIBLE) {
395                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
396                         mz = mem_cgroup_zoneinfo(mem, node, zone);
397                         mctz = soft_limit_tree_node_zone(node, zone);
398                         mem_cgroup_remove_exceeded(mem, mz, mctz);
399                 }
400         }
401 }
402
403 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
404                                          struct page_cgroup *pc,
405                                          bool charge)
406 {
407         int val = (charge)? 1 : -1;
408         struct mem_cgroup_stat *stat = &mem->stat;
409         struct mem_cgroup_stat_cpu *cpustat;
410         int cpu = get_cpu();
411
412         cpustat = &stat->cpustat[cpu];
413         if (PageCgroupCache(pc))
414                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
415         else
416                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
417
418         if (charge)
419                 __mem_cgroup_stat_add_safe(cpustat,
420                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
421         else
422                 __mem_cgroup_stat_add_safe(cpustat,
423                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
424         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1);
425         put_cpu();
426 }
427
428 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
429                                         enum lru_list idx)
430 {
431         int nid, zid;
432         struct mem_cgroup_per_zone *mz;
433         u64 total = 0;
434
435         for_each_online_node(nid)
436                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
437                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
438                         total += MEM_CGROUP_ZSTAT(mz, idx);
439                 }
440         return total;
441 }
442
443 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
444 {
445         return container_of(cgroup_subsys_state(cont,
446                                 mem_cgroup_subsys_id), struct mem_cgroup,
447                                 css);
448 }
449
450 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
451 {
452         /*
453          * mm_update_next_owner() may clear mm->owner to NULL
454          * if it races with swapoff, page migration, etc.
455          * So this can be called with p == NULL.
456          */
457         if (unlikely(!p))
458                 return NULL;
459
460         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
461                                 struct mem_cgroup, css);
462 }
463
464 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
465 {
466         struct mem_cgroup *mem = NULL;
467
468         if (!mm)
469                 return NULL;
470         /*
471          * Because we have no locks, mm->owner's may be being moved to other
472          * cgroup. We use css_tryget() here even if this looks
473          * pessimistic (rather than adding locks here).
474          */
475         rcu_read_lock();
476         do {
477                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
478                 if (unlikely(!mem))
479                         break;
480         } while (!css_tryget(&mem->css));
481         rcu_read_unlock();
482         return mem;
483 }
484
485 /*
486  * Call callback function against all cgroup under hierarchy tree.
487  */
488 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
489                           int (*func)(struct mem_cgroup *, void *))
490 {
491         int found, ret, nextid;
492         struct cgroup_subsys_state *css;
493         struct mem_cgroup *mem;
494
495         if (!root->use_hierarchy)
496                 return (*func)(root, data);
497
498         nextid = 1;
499         do {
500                 ret = 0;
501                 mem = NULL;
502
503                 rcu_read_lock();
504                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
505                                    &found);
506                 if (css && css_tryget(css))
507                         mem = container_of(css, struct mem_cgroup, css);
508                 rcu_read_unlock();
509
510                 if (mem) {
511                         ret = (*func)(mem, data);
512                         css_put(&mem->css);
513                 }
514                 nextid = found + 1;
515         } while (!ret && css);
516
517         return ret;
518 }
519
520 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
521 {
522         return (mem == root_mem_cgroup);
523 }
524
525 /*
526  * Following LRU functions are allowed to be used without PCG_LOCK.
527  * Operations are called by routine of global LRU independently from memcg.
528  * What we have to take care of here is validness of pc->mem_cgroup.
529  *
530  * Changes to pc->mem_cgroup happens when
531  * 1. charge
532  * 2. moving account
533  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
534  * It is added to LRU before charge.
535  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
536  * When moving account, the page is not on LRU. It's isolated.
537  */
538
539 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
540 {
541         struct page_cgroup *pc;
542         struct mem_cgroup_per_zone *mz;
543
544         if (mem_cgroup_disabled())
545                 return;
546         pc = lookup_page_cgroup(page);
547         /* can happen while we handle swapcache. */
548         if (!TestClearPageCgroupAcctLRU(pc))
549                 return;
550         VM_BUG_ON(!pc->mem_cgroup);
551         /*
552          * We don't check PCG_USED bit. It's cleared when the "page" is finally
553          * removed from global LRU.
554          */
555         mz = page_cgroup_zoneinfo(pc);
556         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
557         if (mem_cgroup_is_root(pc->mem_cgroup))
558                 return;
559         VM_BUG_ON(list_empty(&pc->lru));
560         list_del_init(&pc->lru);
561         return;
562 }
563
564 void mem_cgroup_del_lru(struct page *page)
565 {
566         mem_cgroup_del_lru_list(page, page_lru(page));
567 }
568
569 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
570 {
571         struct mem_cgroup_per_zone *mz;
572         struct page_cgroup *pc;
573
574         if (mem_cgroup_disabled())
575                 return;
576
577         pc = lookup_page_cgroup(page);
578         /*
579          * Used bit is set without atomic ops but after smp_wmb().
580          * For making pc->mem_cgroup visible, insert smp_rmb() here.
581          */
582         smp_rmb();
583         /* unused or root page is not rotated. */
584         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
585                 return;
586         mz = page_cgroup_zoneinfo(pc);
587         list_move(&pc->lru, &mz->lists[lru]);
588 }
589
590 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
591 {
592         struct page_cgroup *pc;
593         struct mem_cgroup_per_zone *mz;
594
595         if (mem_cgroup_disabled())
596                 return;
597         pc = lookup_page_cgroup(page);
598         VM_BUG_ON(PageCgroupAcctLRU(pc));
599         /*
600          * Used bit is set without atomic ops but after smp_wmb().
601          * For making pc->mem_cgroup visible, insert smp_rmb() here.
602          */
603         smp_rmb();
604         if (!PageCgroupUsed(pc))
605                 return;
606
607         mz = page_cgroup_zoneinfo(pc);
608         MEM_CGROUP_ZSTAT(mz, lru) += 1;
609         SetPageCgroupAcctLRU(pc);
610         if (mem_cgroup_is_root(pc->mem_cgroup))
611                 return;
612         list_add(&pc->lru, &mz->lists[lru]);
613 }
614
615 /*
616  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
617  * lru because the page may.be reused after it's fully uncharged (because of
618  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
619  * it again. This function is only used to charge SwapCache. It's done under
620  * lock_page and expected that zone->lru_lock is never held.
621  */
622 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
623 {
624         unsigned long flags;
625         struct zone *zone = page_zone(page);
626         struct page_cgroup *pc = lookup_page_cgroup(page);
627
628         spin_lock_irqsave(&zone->lru_lock, flags);
629         /*
630          * Forget old LRU when this page_cgroup is *not* used. This Used bit
631          * is guarded by lock_page() because the page is SwapCache.
632          */
633         if (!PageCgroupUsed(pc))
634                 mem_cgroup_del_lru_list(page, page_lru(page));
635         spin_unlock_irqrestore(&zone->lru_lock, flags);
636 }
637
638 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
639 {
640         unsigned long flags;
641         struct zone *zone = page_zone(page);
642         struct page_cgroup *pc = lookup_page_cgroup(page);
643
644         spin_lock_irqsave(&zone->lru_lock, flags);
645         /* link when the page is linked to LRU but page_cgroup isn't */
646         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
647                 mem_cgroup_add_lru_list(page, page_lru(page));
648         spin_unlock_irqrestore(&zone->lru_lock, flags);
649 }
650
651
652 void mem_cgroup_move_lists(struct page *page,
653                            enum lru_list from, enum lru_list to)
654 {
655         if (mem_cgroup_disabled())
656                 return;
657         mem_cgroup_del_lru_list(page, from);
658         mem_cgroup_add_lru_list(page, to);
659 }
660
661 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
662 {
663         int ret;
664         struct mem_cgroup *curr = NULL;
665
666         task_lock(task);
667         rcu_read_lock();
668         curr = try_get_mem_cgroup_from_mm(task->mm);
669         rcu_read_unlock();
670         task_unlock(task);
671         if (!curr)
672                 return 0;
673         if (curr->use_hierarchy)
674                 ret = css_is_ancestor(&curr->css, &mem->css);
675         else
676                 ret = (curr == mem);
677         css_put(&curr->css);
678         return ret;
679 }
680
681 /*
682  * prev_priority control...this will be used in memory reclaim path.
683  */
684 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
685 {
686         int prev_priority;
687
688         spin_lock(&mem->reclaim_param_lock);
689         prev_priority = mem->prev_priority;
690         spin_unlock(&mem->reclaim_param_lock);
691
692         return prev_priority;
693 }
694
695 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
696 {
697         spin_lock(&mem->reclaim_param_lock);
698         if (priority < mem->prev_priority)
699                 mem->prev_priority = priority;
700         spin_unlock(&mem->reclaim_param_lock);
701 }
702
703 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
704 {
705         spin_lock(&mem->reclaim_param_lock);
706         mem->prev_priority = priority;
707         spin_unlock(&mem->reclaim_param_lock);
708 }
709
710 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
711 {
712         unsigned long active;
713         unsigned long inactive;
714         unsigned long gb;
715         unsigned long inactive_ratio;
716
717         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
718         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
719
720         gb = (inactive + active) >> (30 - PAGE_SHIFT);
721         if (gb)
722                 inactive_ratio = int_sqrt(10 * gb);
723         else
724                 inactive_ratio = 1;
725
726         if (present_pages) {
727                 present_pages[0] = inactive;
728                 present_pages[1] = active;
729         }
730
731         return inactive_ratio;
732 }
733
734 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
735 {
736         unsigned long active;
737         unsigned long inactive;
738         unsigned long present_pages[2];
739         unsigned long inactive_ratio;
740
741         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
742
743         inactive = present_pages[0];
744         active = present_pages[1];
745
746         if (inactive * inactive_ratio < active)
747                 return 1;
748
749         return 0;
750 }
751
752 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
753 {
754         unsigned long active;
755         unsigned long inactive;
756
757         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
758         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
759
760         return (active > inactive);
761 }
762
763 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
764                                        struct zone *zone,
765                                        enum lru_list lru)
766 {
767         int nid = zone->zone_pgdat->node_id;
768         int zid = zone_idx(zone);
769         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
770
771         return MEM_CGROUP_ZSTAT(mz, lru);
772 }
773
774 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
775                                                       struct zone *zone)
776 {
777         int nid = zone->zone_pgdat->node_id;
778         int zid = zone_idx(zone);
779         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
780
781         return &mz->reclaim_stat;
782 }
783
784 struct zone_reclaim_stat *
785 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
786 {
787         struct page_cgroup *pc;
788         struct mem_cgroup_per_zone *mz;
789
790         if (mem_cgroup_disabled())
791                 return NULL;
792
793         pc = lookup_page_cgroup(page);
794         /*
795          * Used bit is set without atomic ops but after smp_wmb().
796          * For making pc->mem_cgroup visible, insert smp_rmb() here.
797          */
798         smp_rmb();
799         if (!PageCgroupUsed(pc))
800                 return NULL;
801
802         mz = page_cgroup_zoneinfo(pc);
803         if (!mz)
804                 return NULL;
805
806         return &mz->reclaim_stat;
807 }
808
809 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
810                                         struct list_head *dst,
811                                         unsigned long *scanned, int order,
812                                         int mode, struct zone *z,
813                                         struct mem_cgroup *mem_cont,
814                                         int active, int file)
815 {
816         unsigned long nr_taken = 0;
817         struct page *page;
818         unsigned long scan;
819         LIST_HEAD(pc_list);
820         struct list_head *src;
821         struct page_cgroup *pc, *tmp;
822         int nid = z->zone_pgdat->node_id;
823         int zid = zone_idx(z);
824         struct mem_cgroup_per_zone *mz;
825         int lru = LRU_FILE * file + active;
826         int ret;
827
828         BUG_ON(!mem_cont);
829         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
830         src = &mz->lists[lru];
831
832         scan = 0;
833         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
834                 if (scan >= nr_to_scan)
835                         break;
836
837                 page = pc->page;
838                 if (unlikely(!PageCgroupUsed(pc)))
839                         continue;
840                 if (unlikely(!PageLRU(page)))
841                         continue;
842
843                 scan++;
844                 ret = __isolate_lru_page(page, mode, file);
845                 switch (ret) {
846                 case 0:
847                         list_move(&page->lru, dst);
848                         mem_cgroup_del_lru(page);
849                         nr_taken++;
850                         break;
851                 case -EBUSY:
852                         /* we don't affect global LRU but rotate in our LRU */
853                         mem_cgroup_rotate_lru_list(page, page_lru(page));
854                         break;
855                 default:
856                         break;
857                 }
858         }
859
860         *scanned = scan;
861         return nr_taken;
862 }
863
864 #define mem_cgroup_from_res_counter(counter, member)    \
865         container_of(counter, struct mem_cgroup, member)
866
867 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
868 {
869         if (do_swap_account) {
870                 if (res_counter_check_under_limit(&mem->res) &&
871                         res_counter_check_under_limit(&mem->memsw))
872                         return true;
873         } else
874                 if (res_counter_check_under_limit(&mem->res))
875                         return true;
876         return false;
877 }
878
879 static unsigned int get_swappiness(struct mem_cgroup *memcg)
880 {
881         struct cgroup *cgrp = memcg->css.cgroup;
882         unsigned int swappiness;
883
884         /* root ? */
885         if (cgrp->parent == NULL)
886                 return vm_swappiness;
887
888         spin_lock(&memcg->reclaim_param_lock);
889         swappiness = memcg->swappiness;
890         spin_unlock(&memcg->reclaim_param_lock);
891
892         return swappiness;
893 }
894
895 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
896 {
897         int *val = data;
898         (*val)++;
899         return 0;
900 }
901
902 /**
903  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
904  * @memcg: The memory cgroup that went over limit
905  * @p: Task that is going to be killed
906  *
907  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
908  * enabled
909  */
910 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
911 {
912         struct cgroup *task_cgrp;
913         struct cgroup *mem_cgrp;
914         /*
915          * Need a buffer in BSS, can't rely on allocations. The code relies
916          * on the assumption that OOM is serialized for memory controller.
917          * If this assumption is broken, revisit this code.
918          */
919         static char memcg_name[PATH_MAX];
920         int ret;
921
922         if (!memcg)
923                 return;
924
925
926         rcu_read_lock();
927
928         mem_cgrp = memcg->css.cgroup;
929         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
930
931         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
932         if (ret < 0) {
933                 /*
934                  * Unfortunately, we are unable to convert to a useful name
935                  * But we'll still print out the usage information
936                  */
937                 rcu_read_unlock();
938                 goto done;
939         }
940         rcu_read_unlock();
941
942         printk(KERN_INFO "Task in %s killed", memcg_name);
943
944         rcu_read_lock();
945         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
946         if (ret < 0) {
947                 rcu_read_unlock();
948                 goto done;
949         }
950         rcu_read_unlock();
951
952         /*
953          * Continues from above, so we don't need an KERN_ level
954          */
955         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
956 done:
957
958         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
959                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
960                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
961                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
962         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
963                 "failcnt %llu\n",
964                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
965                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
966                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
967 }
968
969 /*
970  * This function returns the number of memcg under hierarchy tree. Returns
971  * 1(self count) if no children.
972  */
973 static int mem_cgroup_count_children(struct mem_cgroup *mem)
974 {
975         int num = 0;
976         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
977         return num;
978 }
979
980 /*
981  * Visit the first child (need not be the first child as per the ordering
982  * of the cgroup list, since we track last_scanned_child) of @mem and use
983  * that to reclaim free pages from.
984  */
985 static struct mem_cgroup *
986 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
987 {
988         struct mem_cgroup *ret = NULL;
989         struct cgroup_subsys_state *css;
990         int nextid, found;
991
992         if (!root_mem->use_hierarchy) {
993                 css_get(&root_mem->css);
994                 ret = root_mem;
995         }
996
997         while (!ret) {
998                 rcu_read_lock();
999                 nextid = root_mem->last_scanned_child + 1;
1000                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1001                                    &found);
1002                 if (css && css_tryget(css))
1003                         ret = container_of(css, struct mem_cgroup, css);
1004
1005                 rcu_read_unlock();
1006                 /* Updates scanning parameter */
1007                 spin_lock(&root_mem->reclaim_param_lock);
1008                 if (!css) {
1009                         /* this means start scan from ID:1 */
1010                         root_mem->last_scanned_child = 0;
1011                 } else
1012                         root_mem->last_scanned_child = found;
1013                 spin_unlock(&root_mem->reclaim_param_lock);
1014         }
1015
1016         return ret;
1017 }
1018
1019 /*
1020  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1021  * we reclaimed from, so that we don't end up penalizing one child extensively
1022  * based on its position in the children list.
1023  *
1024  * root_mem is the original ancestor that we've been reclaim from.
1025  *
1026  * We give up and return to the caller when we visit root_mem twice.
1027  * (other groups can be removed while we're walking....)
1028  *
1029  * If shrink==true, for avoiding to free too much, this returns immedieately.
1030  */
1031 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1032                                    gfp_t gfp_mask, bool noswap, bool shrink)
1033 {
1034         struct mem_cgroup *victim;
1035         int ret, total = 0;
1036         int loop = 0;
1037
1038         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1039         if (root_mem->memsw_is_minimum)
1040                 noswap = true;
1041
1042         while (loop < 2) {
1043                 victim = mem_cgroup_select_victim(root_mem);
1044                 if (victim == root_mem)
1045                         loop++;
1046                 if (!mem_cgroup_local_usage(&victim->stat)) {
1047                         /* this cgroup's local usage == 0 */
1048                         css_put(&victim->css);
1049                         continue;
1050                 }
1051                 /* we use swappiness of local cgroup */
1052                 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap,
1053                                                    get_swappiness(victim));
1054                 css_put(&victim->css);
1055                 /*
1056                  * At shrinking usage, we can't check we should stop here or
1057                  * reclaim more. It's depends on callers. last_scanned_child
1058                  * will work enough for keeping fairness under tree.
1059                  */
1060                 if (shrink)
1061                         return ret;
1062                 total += ret;
1063                 if (mem_cgroup_check_under_limit(root_mem))
1064                         return 1 + total;
1065         }
1066         return total;
1067 }
1068
1069 bool mem_cgroup_oom_called(struct task_struct *task)
1070 {
1071         bool ret = false;
1072         struct mem_cgroup *mem;
1073         struct mm_struct *mm;
1074
1075         rcu_read_lock();
1076         mm = task->mm;
1077         if (!mm)
1078                 mm = &init_mm;
1079         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1080         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1081                 ret = true;
1082         rcu_read_unlock();
1083         return ret;
1084 }
1085
1086 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
1087 {
1088         mem->last_oom_jiffies = jiffies;
1089         return 0;
1090 }
1091
1092 static void record_last_oom(struct mem_cgroup *mem)
1093 {
1094         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
1095 }
1096
1097 /*
1098  * Currently used to update mapped file statistics, but the routine can be
1099  * generalized to update other statistics as well.
1100  */
1101 void mem_cgroup_update_mapped_file_stat(struct page *page, int val)
1102 {
1103         struct mem_cgroup *mem;
1104         struct mem_cgroup_stat *stat;
1105         struct mem_cgroup_stat_cpu *cpustat;
1106         int cpu;
1107         struct page_cgroup *pc;
1108
1109         if (!page_is_file_cache(page))
1110                 return;
1111
1112         pc = lookup_page_cgroup(page);
1113         if (unlikely(!pc))
1114                 return;
1115
1116         lock_page_cgroup(pc);
1117         mem = pc->mem_cgroup;
1118         if (!mem)
1119                 goto done;
1120
1121         if (!PageCgroupUsed(pc))
1122                 goto done;
1123
1124         /*
1125          * Preemption is already disabled, we don't need get_cpu()
1126          */
1127         cpu = smp_processor_id();
1128         stat = &mem->stat;
1129         cpustat = &stat->cpustat[cpu];
1130
1131         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val);
1132 done:
1133         unlock_page_cgroup(pc);
1134 }
1135
1136 /*
1137  * Unlike exported interface, "oom" parameter is added. if oom==true,
1138  * oom-killer can be invoked.
1139  */
1140 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1141                         gfp_t gfp_mask, struct mem_cgroup **memcg,
1142                         bool oom, struct page *page)
1143 {
1144         struct mem_cgroup *mem, *mem_over_limit, *mem_over_soft_limit;
1145         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1146         struct res_counter *fail_res, *soft_fail_res = NULL;
1147
1148         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1149                 /* Don't account this! */
1150                 *memcg = NULL;
1151                 return 0;
1152         }
1153
1154         /*
1155          * We always charge the cgroup the mm_struct belongs to.
1156          * The mm_struct's mem_cgroup changes on task migration if the
1157          * thread group leader migrates. It's possible that mm is not
1158          * set, if so charge the init_mm (happens for pagecache usage).
1159          */
1160         mem = *memcg;
1161         if (likely(!mem)) {
1162                 mem = try_get_mem_cgroup_from_mm(mm);
1163                 *memcg = mem;
1164         } else {
1165                 css_get(&mem->css);
1166         }
1167         if (unlikely(!mem))
1168                 return 0;
1169
1170         VM_BUG_ON(css_is_removed(&mem->css));
1171
1172         while (1) {
1173                 int ret;
1174                 bool noswap = false;
1175
1176                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res,
1177                                                 &soft_fail_res);
1178                 if (likely(!ret)) {
1179                         if (!do_swap_account)
1180                                 break;
1181                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
1182                                                         &fail_res, NULL);
1183                         if (likely(!ret))
1184                                 break;
1185                         /* mem+swap counter fails */
1186                         res_counter_uncharge(&mem->res, PAGE_SIZE, NULL);
1187                         noswap = true;
1188                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1189                                                                         memsw);
1190                 } else
1191                         /* mem counter fails */
1192                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1193                                                                         res);
1194
1195                 if (!(gfp_mask & __GFP_WAIT))
1196                         goto nomem;
1197
1198                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask,
1199                                                         noswap, false);
1200                 if (ret)
1201                         continue;
1202
1203                 /*
1204                  * try_to_free_mem_cgroup_pages() might not give us a full
1205                  * picture of reclaim. Some pages are reclaimed and might be
1206                  * moved to swap cache or just unmapped from the cgroup.
1207                  * Check the limit again to see if the reclaim reduced the
1208                  * current usage of the cgroup before giving up
1209                  *
1210                  */
1211                 if (mem_cgroup_check_under_limit(mem_over_limit))
1212                         continue;
1213
1214                 if (!nr_retries--) {
1215                         if (oom) {
1216                                 mutex_lock(&memcg_tasklist);
1217                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1218                                 mutex_unlock(&memcg_tasklist);
1219                                 record_last_oom(mem_over_limit);
1220                         }
1221                         goto nomem;
1222                 }
1223         }
1224         /*
1225          * Insert just the ancestor, we should trickle down to the correct
1226          * cgroup for reclaim, since the other nodes will be below their
1227          * soft limit
1228          */
1229         if (soft_fail_res) {
1230                 mem_over_soft_limit =
1231                         mem_cgroup_from_res_counter(soft_fail_res, res);
1232                 if (mem_cgroup_soft_limit_check(mem_over_soft_limit))
1233                         mem_cgroup_update_tree(mem_over_soft_limit, page);
1234         }
1235         return 0;
1236 nomem:
1237         css_put(&mem->css);
1238         return -ENOMEM;
1239 }
1240
1241 /*
1242  * A helper function to get mem_cgroup from ID. must be called under
1243  * rcu_read_lock(). The caller must check css_is_removed() or some if
1244  * it's concern. (dropping refcnt from swap can be called against removed
1245  * memcg.)
1246  */
1247 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1248 {
1249         struct cgroup_subsys_state *css;
1250
1251         /* ID 0 is unused ID */
1252         if (!id)
1253                 return NULL;
1254         css = css_lookup(&mem_cgroup_subsys, id);
1255         if (!css)
1256                 return NULL;
1257         return container_of(css, struct mem_cgroup, css);
1258 }
1259
1260 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1261 {
1262         struct mem_cgroup *mem;
1263         struct page_cgroup *pc;
1264         unsigned short id;
1265         swp_entry_t ent;
1266
1267         VM_BUG_ON(!PageLocked(page));
1268
1269         if (!PageSwapCache(page))
1270                 return NULL;
1271
1272         pc = lookup_page_cgroup(page);
1273         lock_page_cgroup(pc);
1274         if (PageCgroupUsed(pc)) {
1275                 mem = pc->mem_cgroup;
1276                 if (mem && !css_tryget(&mem->css))
1277                         mem = NULL;
1278         } else {
1279                 ent.val = page_private(page);
1280                 id = lookup_swap_cgroup(ent);
1281                 rcu_read_lock();
1282                 mem = mem_cgroup_lookup(id);
1283                 if (mem && !css_tryget(&mem->css))
1284                         mem = NULL;
1285                 rcu_read_unlock();
1286         }
1287         unlock_page_cgroup(pc);
1288         return mem;
1289 }
1290
1291 /*
1292  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1293  * USED state. If already USED, uncharge and return.
1294  */
1295
1296 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1297                                      struct page_cgroup *pc,
1298                                      enum charge_type ctype)
1299 {
1300         /* try_charge() can return NULL to *memcg, taking care of it. */
1301         if (!mem)
1302                 return;
1303
1304         lock_page_cgroup(pc);
1305         if (unlikely(PageCgroupUsed(pc))) {
1306                 unlock_page_cgroup(pc);
1307                 res_counter_uncharge(&mem->res, PAGE_SIZE, NULL);
1308                 if (do_swap_account)
1309                         res_counter_uncharge(&mem->memsw, PAGE_SIZE, NULL);
1310                 css_put(&mem->css);
1311                 return;
1312         }
1313
1314         pc->mem_cgroup = mem;
1315         /*
1316          * We access a page_cgroup asynchronously without lock_page_cgroup().
1317          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1318          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1319          * before USED bit, we need memory barrier here.
1320          * See mem_cgroup_add_lru_list(), etc.
1321          */
1322         smp_wmb();
1323         switch (ctype) {
1324         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1325         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1326                 SetPageCgroupCache(pc);
1327                 SetPageCgroupUsed(pc);
1328                 break;
1329         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1330                 ClearPageCgroupCache(pc);
1331                 SetPageCgroupUsed(pc);
1332                 break;
1333         default:
1334                 break;
1335         }
1336
1337         mem_cgroup_charge_statistics(mem, pc, true);
1338
1339         unlock_page_cgroup(pc);
1340 }
1341
1342 /**
1343  * mem_cgroup_move_account - move account of the page
1344  * @pc: page_cgroup of the page.
1345  * @from: mem_cgroup which the page is moved from.
1346  * @to: mem_cgroup which the page is moved to. @from != @to.
1347  *
1348  * The caller must confirm following.
1349  * - page is not on LRU (isolate_page() is useful.)
1350  *
1351  * returns 0 at success,
1352  * returns -EBUSY when lock is busy or "pc" is unstable.
1353  *
1354  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1355  * new cgroup. It should be done by a caller.
1356  */
1357
1358 static int mem_cgroup_move_account(struct page_cgroup *pc,
1359         struct mem_cgroup *from, struct mem_cgroup *to)
1360 {
1361         struct mem_cgroup_per_zone *from_mz, *to_mz;
1362         int nid, zid;
1363         int ret = -EBUSY;
1364         struct page *page;
1365         int cpu;
1366         struct mem_cgroup_stat *stat;
1367         struct mem_cgroup_stat_cpu *cpustat;
1368
1369         VM_BUG_ON(from == to);
1370         VM_BUG_ON(PageLRU(pc->page));
1371
1372         nid = page_cgroup_nid(pc);
1373         zid = page_cgroup_zid(pc);
1374         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1375         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1376
1377         if (!trylock_page_cgroup(pc))
1378                 return ret;
1379
1380         if (!PageCgroupUsed(pc))
1381                 goto out;
1382
1383         if (pc->mem_cgroup != from)
1384                 goto out;
1385
1386         res_counter_uncharge(&from->res, PAGE_SIZE, NULL);
1387         mem_cgroup_charge_statistics(from, pc, false);
1388
1389         page = pc->page;
1390         if (page_is_file_cache(page) && page_mapped(page)) {
1391                 cpu = smp_processor_id();
1392                 /* Update mapped_file data for mem_cgroup "from" */
1393                 stat = &from->stat;
1394                 cpustat = &stat->cpustat[cpu];
1395                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1396                                                 -1);
1397
1398                 /* Update mapped_file data for mem_cgroup "to" */
1399                 stat = &to->stat;
1400                 cpustat = &stat->cpustat[cpu];
1401                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1402                                                 1);
1403         }
1404
1405         if (do_swap_account)
1406                 res_counter_uncharge(&from->memsw, PAGE_SIZE, NULL);
1407         css_put(&from->css);
1408
1409         css_get(&to->css);
1410         pc->mem_cgroup = to;
1411         mem_cgroup_charge_statistics(to, pc, true);
1412         ret = 0;
1413 out:
1414         unlock_page_cgroup(pc);
1415         /*
1416          * We charges against "to" which may not have any tasks. Then, "to"
1417          * can be under rmdir(). But in current implementation, caller of
1418          * this function is just force_empty() and it's garanteed that
1419          * "to" is never removed. So, we don't check rmdir status here.
1420          */
1421         return ret;
1422 }
1423
1424 /*
1425  * move charges to its parent.
1426  */
1427
1428 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1429                                   struct mem_cgroup *child,
1430                                   gfp_t gfp_mask)
1431 {
1432         struct page *page = pc->page;
1433         struct cgroup *cg = child->css.cgroup;
1434         struct cgroup *pcg = cg->parent;
1435         struct mem_cgroup *parent;
1436         int ret;
1437
1438         /* Is ROOT ? */
1439         if (!pcg)
1440                 return -EINVAL;
1441
1442
1443         parent = mem_cgroup_from_cont(pcg);
1444
1445
1446         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page);
1447         if (ret || !parent)
1448                 return ret;
1449
1450         if (!get_page_unless_zero(page)) {
1451                 ret = -EBUSY;
1452                 goto uncharge;
1453         }
1454
1455         ret = isolate_lru_page(page);
1456
1457         if (ret)
1458                 goto cancel;
1459
1460         ret = mem_cgroup_move_account(pc, child, parent);
1461
1462         putback_lru_page(page);
1463         if (!ret) {
1464                 put_page(page);
1465                 /* drop extra refcnt by try_charge() */
1466                 css_put(&parent->css);
1467                 return 0;
1468         }
1469
1470 cancel:
1471         put_page(page);
1472 uncharge:
1473         /* drop extra refcnt by try_charge() */
1474         css_put(&parent->css);
1475         /* uncharge if move fails */
1476         res_counter_uncharge(&parent->res, PAGE_SIZE, NULL);
1477         if (do_swap_account)
1478                 res_counter_uncharge(&parent->memsw, PAGE_SIZE, NULL);
1479         return ret;
1480 }
1481
1482 /*
1483  * Charge the memory controller for page usage.
1484  * Return
1485  * 0 if the charge was successful
1486  * < 0 if the cgroup is over its limit
1487  */
1488 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1489                                 gfp_t gfp_mask, enum charge_type ctype,
1490                                 struct mem_cgroup *memcg)
1491 {
1492         struct mem_cgroup *mem;
1493         struct page_cgroup *pc;
1494         int ret;
1495
1496         pc = lookup_page_cgroup(page);
1497         /* can happen at boot */
1498         if (unlikely(!pc))
1499                 return 0;
1500         prefetchw(pc);
1501
1502         mem = memcg;
1503         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page);
1504         if (ret || !mem)
1505                 return ret;
1506
1507         __mem_cgroup_commit_charge(mem, pc, ctype);
1508         return 0;
1509 }
1510
1511 int mem_cgroup_newpage_charge(struct page *page,
1512                               struct mm_struct *mm, gfp_t gfp_mask)
1513 {
1514         if (mem_cgroup_disabled())
1515                 return 0;
1516         if (PageCompound(page))
1517                 return 0;
1518         /*
1519          * If already mapped, we don't have to account.
1520          * If page cache, page->mapping has address_space.
1521          * But page->mapping may have out-of-use anon_vma pointer,
1522          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1523          * is NULL.
1524          */
1525         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1526                 return 0;
1527         if (unlikely(!mm))
1528                 mm = &init_mm;
1529         return mem_cgroup_charge_common(page, mm, gfp_mask,
1530                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1531 }
1532
1533 static void
1534 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1535                                         enum charge_type ctype);
1536
1537 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1538                                 gfp_t gfp_mask)
1539 {
1540         struct mem_cgroup *mem = NULL;
1541         int ret;
1542
1543         if (mem_cgroup_disabled())
1544                 return 0;
1545         if (PageCompound(page))
1546                 return 0;
1547         /*
1548          * Corner case handling. This is called from add_to_page_cache()
1549          * in usual. But some FS (shmem) precharges this page before calling it
1550          * and call add_to_page_cache() with GFP_NOWAIT.
1551          *
1552          * For GFP_NOWAIT case, the page may be pre-charged before calling
1553          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1554          * charge twice. (It works but has to pay a bit larger cost.)
1555          * And when the page is SwapCache, it should take swap information
1556          * into account. This is under lock_page() now.
1557          */
1558         if (!(gfp_mask & __GFP_WAIT)) {
1559                 struct page_cgroup *pc;
1560
1561
1562                 pc = lookup_page_cgroup(page);
1563                 if (!pc)
1564                         return 0;
1565                 lock_page_cgroup(pc);
1566                 if (PageCgroupUsed(pc)) {
1567                         unlock_page_cgroup(pc);
1568                         return 0;
1569                 }
1570                 unlock_page_cgroup(pc);
1571         }
1572
1573         if (unlikely(!mm && !mem))
1574                 mm = &init_mm;
1575
1576         if (page_is_file_cache(page))
1577                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1578                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1579
1580         /* shmem */
1581         if (PageSwapCache(page)) {
1582                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1583                 if (!ret)
1584                         __mem_cgroup_commit_charge_swapin(page, mem,
1585                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1586         } else
1587                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1588                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1589
1590         return ret;
1591 }
1592
1593 /*
1594  * While swap-in, try_charge -> commit or cancel, the page is locked.
1595  * And when try_charge() successfully returns, one refcnt to memcg without
1596  * struct page_cgroup is aquired. This refcnt will be cumsumed by
1597  * "commit()" or removed by "cancel()"
1598  */
1599 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1600                                  struct page *page,
1601                                  gfp_t mask, struct mem_cgroup **ptr)
1602 {
1603         struct mem_cgroup *mem;
1604         int ret;
1605
1606         if (mem_cgroup_disabled())
1607                 return 0;
1608
1609         if (!do_swap_account)
1610                 goto charge_cur_mm;
1611         /*
1612          * A racing thread's fault, or swapoff, may have already updated
1613          * the pte, and even removed page from swap cache: return success
1614          * to go on to do_swap_page()'s pte_same() test, which should fail.
1615          */
1616         if (!PageSwapCache(page))
1617                 return 0;
1618         mem = try_get_mem_cgroup_from_swapcache(page);
1619         if (!mem)
1620                 goto charge_cur_mm;
1621         *ptr = mem;
1622         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page);
1623         /* drop extra refcnt from tryget */
1624         css_put(&mem->css);
1625         return ret;
1626 charge_cur_mm:
1627         if (unlikely(!mm))
1628                 mm = &init_mm;
1629         return __mem_cgroup_try_charge(mm, mask, ptr, true, page);
1630 }
1631
1632 static void
1633 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1634                                         enum charge_type ctype)
1635 {
1636         struct page_cgroup *pc;
1637
1638         if (mem_cgroup_disabled())
1639                 return;
1640         if (!ptr)
1641                 return;
1642         cgroup_exclude_rmdir(&ptr->css);
1643         pc = lookup_page_cgroup(page);
1644         mem_cgroup_lru_del_before_commit_swapcache(page);
1645         __mem_cgroup_commit_charge(ptr, pc, ctype);
1646         mem_cgroup_lru_add_after_commit_swapcache(page);
1647         /*
1648          * Now swap is on-memory. This means this page may be
1649          * counted both as mem and swap....double count.
1650          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1651          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1652          * may call delete_from_swap_cache() before reach here.
1653          */
1654         if (do_swap_account && PageSwapCache(page)) {
1655                 swp_entry_t ent = {.val = page_private(page)};
1656                 unsigned short id;
1657                 struct mem_cgroup *memcg;
1658
1659                 id = swap_cgroup_record(ent, 0);
1660                 rcu_read_lock();
1661                 memcg = mem_cgroup_lookup(id);
1662                 if (memcg) {
1663                         /*
1664                          * This recorded memcg can be obsolete one. So, avoid
1665                          * calling css_tryget
1666                          */
1667                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE, NULL);
1668                         mem_cgroup_put(memcg);
1669                 }
1670                 rcu_read_unlock();
1671         }
1672         /*
1673          * At swapin, we may charge account against cgroup which has no tasks.
1674          * So, rmdir()->pre_destroy() can be called while we do this charge.
1675          * In that case, we need to call pre_destroy() again. check it here.
1676          */
1677         cgroup_release_and_wakeup_rmdir(&ptr->css);
1678 }
1679
1680 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1681 {
1682         __mem_cgroup_commit_charge_swapin(page, ptr,
1683                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1684 }
1685
1686 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1687 {
1688         if (mem_cgroup_disabled())
1689                 return;
1690         if (!mem)
1691                 return;
1692         res_counter_uncharge(&mem->res, PAGE_SIZE, NULL);
1693         if (do_swap_account)
1694                 res_counter_uncharge(&mem->memsw, PAGE_SIZE, NULL);
1695         css_put(&mem->css);
1696 }
1697
1698
1699 /*
1700  * uncharge if !page_mapped(page)
1701  */
1702 static struct mem_cgroup *
1703 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1704 {
1705         struct page_cgroup *pc;
1706         struct mem_cgroup *mem = NULL;
1707         struct mem_cgroup_per_zone *mz;
1708         bool soft_limit_excess = false;
1709
1710         if (mem_cgroup_disabled())
1711                 return NULL;
1712
1713         if (PageSwapCache(page))
1714                 return NULL;
1715
1716         /*
1717          * Check if our page_cgroup is valid
1718          */
1719         pc = lookup_page_cgroup(page);
1720         if (unlikely(!pc || !PageCgroupUsed(pc)))
1721                 return NULL;
1722
1723         lock_page_cgroup(pc);
1724
1725         mem = pc->mem_cgroup;
1726
1727         if (!PageCgroupUsed(pc))
1728                 goto unlock_out;
1729
1730         switch (ctype) {
1731         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1732         case MEM_CGROUP_CHARGE_TYPE_DROP:
1733                 if (page_mapped(page))
1734                         goto unlock_out;
1735                 break;
1736         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1737                 if (!PageAnon(page)) {  /* Shared memory */
1738                         if (page->mapping && !page_is_file_cache(page))
1739                                 goto unlock_out;
1740                 } else if (page_mapped(page)) /* Anon */
1741                                 goto unlock_out;
1742                 break;
1743         default:
1744                 break;
1745         }
1746
1747         res_counter_uncharge(&mem->res, PAGE_SIZE, &soft_limit_excess);
1748         if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1749                 res_counter_uncharge(&mem->memsw, PAGE_SIZE, NULL);
1750         mem_cgroup_charge_statistics(mem, pc, false);
1751
1752         ClearPageCgroupUsed(pc);
1753         /*
1754          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1755          * freed from LRU. This is safe because uncharged page is expected not
1756          * to be reused (freed soon). Exception is SwapCache, it's handled by
1757          * special functions.
1758          */
1759
1760         mz = page_cgroup_zoneinfo(pc);
1761         unlock_page_cgroup(pc);
1762
1763         if (soft_limit_excess && mem_cgroup_soft_limit_check(mem))
1764                 mem_cgroup_update_tree(mem, page);
1765         /* at swapout, this memcg will be accessed to record to swap */
1766         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1767                 css_put(&mem->css);
1768
1769         return mem;
1770
1771 unlock_out:
1772         unlock_page_cgroup(pc);
1773         return NULL;
1774 }
1775
1776 void mem_cgroup_uncharge_page(struct page *page)
1777 {
1778         /* early check. */
1779         if (page_mapped(page))
1780                 return;
1781         if (page->mapping && !PageAnon(page))
1782                 return;
1783         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1784 }
1785
1786 void mem_cgroup_uncharge_cache_page(struct page *page)
1787 {
1788         VM_BUG_ON(page_mapped(page));
1789         VM_BUG_ON(page->mapping);
1790         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1791 }
1792
1793 #ifdef CONFIG_SWAP
1794 /*
1795  * called after __delete_from_swap_cache() and drop "page" account.
1796  * memcg information is recorded to swap_cgroup of "ent"
1797  */
1798 void
1799 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
1800 {
1801         struct mem_cgroup *memcg;
1802         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
1803
1804         if (!swapout) /* this was a swap cache but the swap is unused ! */
1805                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
1806
1807         memcg = __mem_cgroup_uncharge_common(page, ctype);
1808
1809         /* record memcg information */
1810         if (do_swap_account && swapout && memcg) {
1811                 swap_cgroup_record(ent, css_id(&memcg->css));
1812                 mem_cgroup_get(memcg);
1813         }
1814         if (swapout && memcg)
1815                 css_put(&memcg->css);
1816 }
1817 #endif
1818
1819 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1820 /*
1821  * called from swap_entry_free(). remove record in swap_cgroup and
1822  * uncharge "memsw" account.
1823  */
1824 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1825 {
1826         struct mem_cgroup *memcg;
1827         unsigned short id;
1828
1829         if (!do_swap_account)
1830                 return;
1831
1832         id = swap_cgroup_record(ent, 0);
1833         rcu_read_lock();
1834         memcg = mem_cgroup_lookup(id);
1835         if (memcg) {
1836                 /*
1837                  * We uncharge this because swap is freed.
1838                  * This memcg can be obsolete one. We avoid calling css_tryget
1839                  */
1840                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE, NULL);
1841                 mem_cgroup_put(memcg);
1842         }
1843         rcu_read_unlock();
1844 }
1845 #endif
1846
1847 /*
1848  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1849  * page belongs to.
1850  */
1851 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1852 {
1853         struct page_cgroup *pc;
1854         struct mem_cgroup *mem = NULL;
1855         int ret = 0;
1856
1857         if (mem_cgroup_disabled())
1858                 return 0;
1859
1860         pc = lookup_page_cgroup(page);
1861         lock_page_cgroup(pc);
1862         if (PageCgroupUsed(pc)) {
1863                 mem = pc->mem_cgroup;
1864                 css_get(&mem->css);
1865         }
1866         unlock_page_cgroup(pc);
1867
1868         if (mem) {
1869                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
1870                                                 page);
1871                 css_put(&mem->css);
1872         }
1873         *ptr = mem;
1874         return ret;
1875 }
1876
1877 /* remove redundant charge if migration failed*/
1878 void mem_cgroup_end_migration(struct mem_cgroup *mem,
1879                 struct page *oldpage, struct page *newpage)
1880 {
1881         struct page *target, *unused;
1882         struct page_cgroup *pc;
1883         enum charge_type ctype;
1884
1885         if (!mem)
1886                 return;
1887         cgroup_exclude_rmdir(&mem->css);
1888         /* at migration success, oldpage->mapping is NULL. */
1889         if (oldpage->mapping) {
1890                 target = oldpage;
1891                 unused = NULL;
1892         } else {
1893                 target = newpage;
1894                 unused = oldpage;
1895         }
1896
1897         if (PageAnon(target))
1898                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
1899         else if (page_is_file_cache(target))
1900                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
1901         else
1902                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
1903
1904         /* unused page is not on radix-tree now. */
1905         if (unused)
1906                 __mem_cgroup_uncharge_common(unused, ctype);
1907
1908         pc = lookup_page_cgroup(target);
1909         /*
1910          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
1911          * So, double-counting is effectively avoided.
1912          */
1913         __mem_cgroup_commit_charge(mem, pc, ctype);
1914
1915         /*
1916          * Both of oldpage and newpage are still under lock_page().
1917          * Then, we don't have to care about race in radix-tree.
1918          * But we have to be careful that this page is unmapped or not.
1919          *
1920          * There is a case for !page_mapped(). At the start of
1921          * migration, oldpage was mapped. But now, it's zapped.
1922          * But we know *target* page is not freed/reused under us.
1923          * mem_cgroup_uncharge_page() does all necessary checks.
1924          */
1925         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
1926                 mem_cgroup_uncharge_page(target);
1927         /*
1928          * At migration, we may charge account against cgroup which has no tasks
1929          * So, rmdir()->pre_destroy() can be called while we do this charge.
1930          * In that case, we need to call pre_destroy() again. check it here.
1931          */
1932         cgroup_release_and_wakeup_rmdir(&mem->css);
1933 }
1934
1935 /*
1936  * A call to try to shrink memory usage on charge failure at shmem's swapin.
1937  * Calling hierarchical_reclaim is not enough because we should update
1938  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
1939  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
1940  * not from the memcg which this page would be charged to.
1941  * try_charge_swapin does all of these works properly.
1942  */
1943 int mem_cgroup_shmem_charge_fallback(struct page *page,
1944                             struct mm_struct *mm,
1945                             gfp_t gfp_mask)
1946 {
1947         struct mem_cgroup *mem = NULL;
1948         int ret;
1949
1950         if (mem_cgroup_disabled())
1951                 return 0;
1952
1953         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1954         if (!ret)
1955                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
1956
1957         return ret;
1958 }
1959
1960 static DEFINE_MUTEX(set_limit_mutex);
1961
1962 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
1963                                 unsigned long long val)
1964 {
1965         int retry_count;
1966         int progress;
1967         u64 memswlimit;
1968         int ret = 0;
1969         int children = mem_cgroup_count_children(memcg);
1970         u64 curusage, oldusage;
1971
1972         /*
1973          * For keeping hierarchical_reclaim simple, how long we should retry
1974          * is depends on callers. We set our retry-count to be function
1975          * of # of children which we should visit in this loop.
1976          */
1977         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
1978
1979         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
1980
1981         while (retry_count) {
1982                 if (signal_pending(current)) {
1983                         ret = -EINTR;
1984                         break;
1985                 }
1986                 /*
1987                  * Rather than hide all in some function, I do this in
1988                  * open coded manner. You see what this really does.
1989                  * We have to guarantee mem->res.limit < mem->memsw.limit.
1990                  */
1991                 mutex_lock(&set_limit_mutex);
1992                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1993                 if (memswlimit < val) {
1994                         ret = -EINVAL;
1995                         mutex_unlock(&set_limit_mutex);
1996                         break;
1997                 }
1998                 ret = res_counter_set_limit(&memcg->res, val);
1999                 if (!ret) {
2000                         if (memswlimit == val)
2001                                 memcg->memsw_is_minimum = true;
2002                         else
2003                                 memcg->memsw_is_minimum = false;
2004                 }
2005                 mutex_unlock(&set_limit_mutex);
2006
2007                 if (!ret)
2008                         break;
2009
2010                 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL,
2011                                                    false, true);
2012                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2013                 /* Usage is reduced ? */
2014                 if (curusage >= oldusage)
2015                         retry_count--;
2016                 else
2017                         oldusage = curusage;
2018         }
2019
2020         return ret;
2021 }
2022
2023 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2024                                         unsigned long long val)
2025 {
2026         int retry_count;
2027         u64 memlimit, oldusage, curusage;
2028         int children = mem_cgroup_count_children(memcg);
2029         int ret = -EBUSY;
2030
2031         /* see mem_cgroup_resize_res_limit */
2032         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2033         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2034         while (retry_count) {
2035                 if (signal_pending(current)) {
2036                         ret = -EINTR;
2037                         break;
2038                 }
2039                 /*
2040                  * Rather than hide all in some function, I do this in
2041                  * open coded manner. You see what this really does.
2042                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2043                  */
2044                 mutex_lock(&set_limit_mutex);
2045                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2046                 if (memlimit > val) {
2047                         ret = -EINVAL;
2048                         mutex_unlock(&set_limit_mutex);
2049                         break;
2050                 }
2051                 ret = res_counter_set_limit(&memcg->memsw, val);
2052                 if (!ret) {
2053                         if (memlimit == val)
2054                                 memcg->memsw_is_minimum = true;
2055                         else
2056                                 memcg->memsw_is_minimum = false;
2057                 }
2058                 mutex_unlock(&set_limit_mutex);
2059
2060                 if (!ret)
2061                         break;
2062
2063                 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true);
2064                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2065                 /* Usage is reduced ? */
2066                 if (curusage >= oldusage)
2067                         retry_count--;
2068                 else
2069                         oldusage = curusage;
2070         }
2071         return ret;
2072 }
2073
2074 /*
2075  * This routine traverse page_cgroup in given list and drop them all.
2076  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2077  */
2078 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2079                                 int node, int zid, enum lru_list lru)
2080 {
2081         struct zone *zone;
2082         struct mem_cgroup_per_zone *mz;
2083         struct page_cgroup *pc, *busy;
2084         unsigned long flags, loop;
2085         struct list_head *list;
2086         int ret = 0;
2087
2088         zone = &NODE_DATA(node)->node_zones[zid];
2089         mz = mem_cgroup_zoneinfo(mem, node, zid);
2090         list = &mz->lists[lru];
2091
2092         loop = MEM_CGROUP_ZSTAT(mz, lru);
2093         /* give some margin against EBUSY etc...*/
2094         loop += 256;
2095         busy = NULL;
2096         while (loop--) {
2097                 ret = 0;
2098                 spin_lock_irqsave(&zone->lru_lock, flags);
2099                 if (list_empty(list)) {
2100                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2101                         break;
2102                 }
2103                 pc = list_entry(list->prev, struct page_cgroup, lru);
2104                 if (busy == pc) {
2105                         list_move(&pc->lru, list);
2106                         busy = 0;
2107                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2108                         continue;
2109                 }
2110                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2111
2112                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2113                 if (ret == -ENOMEM)
2114                         break;
2115
2116                 if (ret == -EBUSY || ret == -EINVAL) {
2117                         /* found lock contention or "pc" is obsolete. */
2118                         busy = pc;
2119                         cond_resched();
2120                 } else
2121                         busy = NULL;
2122         }
2123
2124         if (!ret && !list_empty(list))
2125                 return -EBUSY;
2126         return ret;
2127 }
2128
2129 /*
2130  * make mem_cgroup's charge to be 0 if there is no task.
2131  * This enables deleting this mem_cgroup.
2132  */
2133 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2134 {
2135         int ret;
2136         int node, zid, shrink;
2137         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2138         struct cgroup *cgrp = mem->css.cgroup;
2139
2140         css_get(&mem->css);
2141
2142         shrink = 0;
2143         /* should free all ? */
2144         if (free_all)
2145                 goto try_to_free;
2146 move_account:
2147         while (mem->res.usage > 0) {
2148                 ret = -EBUSY;
2149                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2150                         goto out;
2151                 ret = -EINTR;
2152                 if (signal_pending(current))
2153                         goto out;
2154                 /* This is for making all *used* pages to be on LRU. */
2155                 lru_add_drain_all();
2156                 ret = 0;
2157                 for_each_node_state(node, N_HIGH_MEMORY) {
2158                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2159                                 enum lru_list l;
2160                                 for_each_lru(l) {
2161                                         ret = mem_cgroup_force_empty_list(mem,
2162                                                         node, zid, l);
2163                                         if (ret)
2164                                                 break;
2165                                 }
2166                         }
2167                         if (ret)
2168                                 break;
2169                 }
2170                 /* it seems parent cgroup doesn't have enough mem */
2171                 if (ret == -ENOMEM)
2172                         goto try_to_free;
2173                 cond_resched();
2174         }
2175         ret = 0;
2176 out:
2177         css_put(&mem->css);
2178         return ret;
2179
2180 try_to_free:
2181         /* returns EBUSY if there is a task or if we come here twice. */
2182         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2183                 ret = -EBUSY;
2184                 goto out;
2185         }
2186         /* we call try-to-free pages for make this cgroup empty */
2187         lru_add_drain_all();
2188         /* try to free all pages in this cgroup */
2189         shrink = 1;
2190         while (nr_retries && mem->res.usage > 0) {
2191                 int progress;
2192
2193                 if (signal_pending(current)) {
2194                         ret = -EINTR;
2195                         goto out;
2196                 }
2197                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2198                                                 false, get_swappiness(mem));
2199                 if (!progress) {
2200                         nr_retries--;
2201                         /* maybe some writeback is necessary */
2202                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2203                 }
2204
2205         }
2206         lru_add_drain();
2207         /* try move_account...there may be some *locked* pages. */
2208         if (mem->res.usage)
2209                 goto move_account;
2210         ret = 0;
2211         goto out;
2212 }
2213
2214 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2215 {
2216         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2217 }
2218
2219
2220 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2221 {
2222         return mem_cgroup_from_cont(cont)->use_hierarchy;
2223 }
2224
2225 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2226                                         u64 val)
2227 {
2228         int retval = 0;
2229         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2230         struct cgroup *parent = cont->parent;
2231         struct mem_cgroup *parent_mem = NULL;
2232
2233         if (parent)
2234                 parent_mem = mem_cgroup_from_cont(parent);
2235
2236         cgroup_lock();
2237         /*
2238          * If parent's use_hiearchy is set, we can't make any modifications
2239          * in the child subtrees. If it is unset, then the change can
2240          * occur, provided the current cgroup has no children.
2241          *
2242          * For the root cgroup, parent_mem is NULL, we allow value to be
2243          * set if there are no children.
2244          */
2245         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2246                                 (val == 1 || val == 0)) {
2247                 if (list_empty(&cont->children))
2248                         mem->use_hierarchy = val;
2249                 else
2250                         retval = -EBUSY;
2251         } else
2252                 retval = -EINVAL;
2253         cgroup_unlock();
2254
2255         return retval;
2256 }
2257
2258 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2259 {
2260         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2261         u64 val = 0;
2262         int type, name;
2263
2264         type = MEMFILE_TYPE(cft->private);
2265         name = MEMFILE_ATTR(cft->private);
2266         switch (type) {
2267         case _MEM:
2268                 val = res_counter_read_u64(&mem->res, name);
2269                 break;
2270         case _MEMSWAP:
2271                 val = res_counter_read_u64(&mem->memsw, name);
2272                 break;
2273         default:
2274                 BUG();
2275                 break;
2276         }
2277         return val;
2278 }
2279 /*
2280  * The user of this function is...
2281  * RES_LIMIT.
2282  */
2283 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2284                             const char *buffer)
2285 {
2286         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2287         int type, name;
2288         unsigned long long val;
2289         int ret;
2290
2291         type = MEMFILE_TYPE(cft->private);
2292         name = MEMFILE_ATTR(cft->private);
2293         switch (name) {
2294         case RES_LIMIT:
2295                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2296                         ret = -EINVAL;
2297                         break;
2298                 }
2299                 /* This function does all necessary parse...reuse it */
2300                 ret = res_counter_memparse_write_strategy(buffer, &val);
2301                 if (ret)
2302                         break;
2303                 if (type == _MEM)
2304                         ret = mem_cgroup_resize_limit(memcg, val);
2305                 else
2306                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2307                 break;
2308         case RES_SOFT_LIMIT:
2309                 ret = res_counter_memparse_write_strategy(buffer, &val);
2310                 if (ret)
2311                         break;
2312                 /*
2313                  * For memsw, soft limits are hard to implement in terms
2314                  * of semantics, for now, we support soft limits for
2315                  * control without swap
2316                  */
2317                 if (type == _MEM)
2318                         ret = res_counter_set_soft_limit(&memcg->res, val);
2319                 else
2320                         ret = -EINVAL;
2321                 break;
2322         default:
2323                 ret = -EINVAL; /* should be BUG() ? */
2324                 break;
2325         }
2326         return ret;
2327 }
2328
2329 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2330                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2331 {
2332         struct cgroup *cgroup;
2333         unsigned long long min_limit, min_memsw_limit, tmp;
2334
2335         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2336         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2337         cgroup = memcg->css.cgroup;
2338         if (!memcg->use_hierarchy)
2339                 goto out;
2340
2341         while (cgroup->parent) {
2342                 cgroup = cgroup->parent;
2343                 memcg = mem_cgroup_from_cont(cgroup);
2344                 if (!memcg->use_hierarchy)
2345                         break;
2346                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2347                 min_limit = min(min_limit, tmp);
2348                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2349                 min_memsw_limit = min(min_memsw_limit, tmp);
2350         }
2351 out:
2352         *mem_limit = min_limit;
2353         *memsw_limit = min_memsw_limit;
2354         return;
2355 }
2356
2357 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2358 {
2359         struct mem_cgroup *mem;
2360         int type, name;
2361
2362         mem = mem_cgroup_from_cont(cont);
2363         type = MEMFILE_TYPE(event);
2364         name = MEMFILE_ATTR(event);
2365         switch (name) {
2366         case RES_MAX_USAGE:
2367                 if (type == _MEM)
2368                         res_counter_reset_max(&mem->res);
2369                 else
2370                         res_counter_reset_max(&mem->memsw);
2371                 break;
2372         case RES_FAILCNT:
2373                 if (type == _MEM)
2374                         res_counter_reset_failcnt(&mem->res);
2375                 else
2376                         res_counter_reset_failcnt(&mem->memsw);
2377                 break;
2378         }
2379
2380         return 0;
2381 }
2382
2383
2384 /* For read statistics */
2385 enum {
2386         MCS_CACHE,
2387         MCS_RSS,
2388         MCS_MAPPED_FILE,
2389         MCS_PGPGIN,
2390         MCS_PGPGOUT,
2391         MCS_INACTIVE_ANON,
2392         MCS_ACTIVE_ANON,
2393         MCS_INACTIVE_FILE,
2394         MCS_ACTIVE_FILE,
2395         MCS_UNEVICTABLE,
2396         NR_MCS_STAT,
2397 };
2398
2399 struct mcs_total_stat {
2400         s64 stat[NR_MCS_STAT];
2401 };
2402
2403 struct {
2404         char *local_name;
2405         char *total_name;
2406 } memcg_stat_strings[NR_MCS_STAT] = {
2407         {"cache", "total_cache"},
2408         {"rss", "total_rss"},
2409         {"mapped_file", "total_mapped_file"},
2410         {"pgpgin", "total_pgpgin"},
2411         {"pgpgout", "total_pgpgout"},
2412         {"inactive_anon", "total_inactive_anon"},
2413         {"active_anon", "total_active_anon"},
2414         {"inactive_file", "total_inactive_file"},
2415         {"active_file", "total_active_file"},
2416         {"unevictable", "total_unevictable"}
2417 };
2418
2419
2420 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2421 {
2422         struct mcs_total_stat *s = data;
2423         s64 val;
2424
2425         /* per cpu stat */
2426         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2427         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2428         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2429         s->stat[MCS_RSS] += val * PAGE_SIZE;
2430         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE);
2431         s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE;
2432         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2433         s->stat[MCS_PGPGIN] += val;
2434         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2435         s->stat[MCS_PGPGOUT] += val;
2436
2437         /* per zone stat */
2438         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2439         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2440         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2441         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2442         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2443         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2444         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2445         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2446         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2447         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2448         return 0;
2449 }
2450
2451 static void
2452 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2453 {
2454         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2455 }
2456
2457 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2458                                  struct cgroup_map_cb *cb)
2459 {
2460         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2461         struct mcs_total_stat mystat;
2462         int i;
2463
2464         memset(&mystat, 0, sizeof(mystat));
2465         mem_cgroup_get_local_stat(mem_cont, &mystat);
2466
2467         for (i = 0; i < NR_MCS_STAT; i++)
2468                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2469
2470         /* Hierarchical information */
2471         {
2472                 unsigned long long limit, memsw_limit;
2473                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2474                 cb->fill(cb, "hierarchical_memory_limit", limit);
2475                 if (do_swap_account)
2476                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2477         }
2478
2479         memset(&mystat, 0, sizeof(mystat));
2480         mem_cgroup_get_total_stat(mem_cont, &mystat);
2481         for (i = 0; i < NR_MCS_STAT; i++)
2482                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2483
2484
2485 #ifdef CONFIG_DEBUG_VM
2486         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2487
2488         {
2489                 int nid, zid;
2490                 struct mem_cgroup_per_zone *mz;
2491                 unsigned long recent_rotated[2] = {0, 0};
2492                 unsigned long recent_scanned[2] = {0, 0};
2493
2494                 for_each_online_node(nid)
2495                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2496                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2497
2498                                 recent_rotated[0] +=
2499                                         mz->reclaim_stat.recent_rotated[0];
2500                                 recent_rotated[1] +=
2501                                         mz->reclaim_stat.recent_rotated[1];
2502                                 recent_scanned[0] +=
2503                                         mz->reclaim_stat.recent_scanned[0];
2504                                 recent_scanned[1] +=
2505                                         mz->reclaim_stat.recent_scanned[1];
2506                         }
2507                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2508                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2509                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2510                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2511         }
2512 #endif
2513
2514         return 0;
2515 }
2516
2517 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2518 {
2519         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2520
2521         return get_swappiness(memcg);
2522 }
2523
2524 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2525                                        u64 val)
2526 {
2527         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2528         struct mem_cgroup *parent;
2529
2530         if (val > 100)
2531                 return -EINVAL;
2532
2533         if (cgrp->parent == NULL)
2534                 return -EINVAL;
2535
2536         parent = mem_cgroup_from_cont(cgrp->parent);
2537
2538         cgroup_lock();
2539
2540         /* If under hierarchy, only empty-root can set this value */
2541         if ((parent->use_hierarchy) ||
2542             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2543                 cgroup_unlock();
2544                 return -EINVAL;
2545         }
2546
2547         spin_lock(&memcg->reclaim_param_lock);
2548         memcg->swappiness = val;
2549         spin_unlock(&memcg->reclaim_param_lock);
2550
2551         cgroup_unlock();
2552
2553         return 0;
2554 }
2555
2556
2557 static struct cftype mem_cgroup_files[] = {
2558         {
2559                 .name = "usage_in_bytes",
2560                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2561                 .read_u64 = mem_cgroup_read,
2562         },
2563         {
2564                 .name = "max_usage_in_bytes",
2565                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2566                 .trigger = mem_cgroup_reset,
2567                 .read_u64 = mem_cgroup_read,
2568         },
2569         {
2570                 .name = "limit_in_bytes",
2571                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2572                 .write_string = mem_cgroup_write,
2573                 .read_u64 = mem_cgroup_read,
2574         },
2575         {
2576                 .name = "soft_limit_in_bytes",
2577                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
2578                 .write_string = mem_cgroup_write,
2579                 .read_u64 = mem_cgroup_read,
2580         },
2581         {
2582                 .name = "failcnt",
2583                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2584                 .trigger = mem_cgroup_reset,
2585                 .read_u64 = mem_cgroup_read,
2586         },
2587         {
2588                 .name = "stat",
2589                 .read_map = mem_control_stat_show,
2590         },
2591         {
2592                 .name = "force_empty",
2593                 .trigger = mem_cgroup_force_empty_write,
2594         },
2595         {
2596                 .name = "use_hierarchy",
2597                 .write_u64 = mem_cgroup_hierarchy_write,
2598                 .read_u64 = mem_cgroup_hierarchy_read,
2599         },
2600         {
2601                 .name = "swappiness",
2602                 .read_u64 = mem_cgroup_swappiness_read,
2603                 .write_u64 = mem_cgroup_swappiness_write,
2604         },
2605 };
2606
2607 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2608 static struct cftype memsw_cgroup_files[] = {
2609         {
2610                 .name = "memsw.usage_in_bytes",
2611                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2612                 .read_u64 = mem_cgroup_read,
2613         },
2614         {
2615                 .name = "memsw.max_usage_in_bytes",
2616                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2617                 .trigger = mem_cgroup_reset,
2618                 .read_u64 = mem_cgroup_read,
2619         },
2620         {
2621                 .name = "memsw.limit_in_bytes",
2622                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2623                 .write_string = mem_cgroup_write,
2624                 .read_u64 = mem_cgroup_read,
2625         },
2626         {
2627                 .name = "memsw.failcnt",
2628                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2629                 .trigger = mem_cgroup_reset,
2630                 .read_u64 = mem_cgroup_read,
2631         },
2632 };
2633
2634 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2635 {
2636         if (!do_swap_account)
2637                 return 0;
2638         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2639                                 ARRAY_SIZE(memsw_cgroup_files));
2640 };
2641 #else
2642 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2643 {
2644         return 0;
2645 }
2646 #endif
2647
2648 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2649 {
2650         struct mem_cgroup_per_node *pn;
2651         struct mem_cgroup_per_zone *mz;
2652         enum lru_list l;
2653         int zone, tmp = node;
2654         /*
2655          * This routine is called against possible nodes.
2656          * But it's BUG to call kmalloc() against offline node.
2657          *
2658          * TODO: this routine can waste much memory for nodes which will
2659          *       never be onlined. It's better to use memory hotplug callback
2660          *       function.
2661          */
2662         if (!node_state(node, N_NORMAL_MEMORY))
2663                 tmp = -1;
2664         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2665         if (!pn)
2666                 return 1;
2667
2668         mem->info.nodeinfo[node] = pn;
2669         memset(pn, 0, sizeof(*pn));
2670
2671         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2672                 mz = &pn->zoneinfo[zone];
2673                 for_each_lru(l)
2674                         INIT_LIST_HEAD(&mz->lists[l]);
2675                 mz->usage_in_excess = 0;
2676         }
2677         return 0;
2678 }
2679
2680 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2681 {
2682         kfree(mem->info.nodeinfo[node]);
2683 }
2684
2685 static int mem_cgroup_size(void)
2686 {
2687         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2688         return sizeof(struct mem_cgroup) + cpustat_size;
2689 }
2690
2691 static struct mem_cgroup *mem_cgroup_alloc(void)
2692 {
2693         struct mem_cgroup *mem;
2694         int size = mem_cgroup_size();
2695
2696         if (size < PAGE_SIZE)
2697                 mem = kmalloc(size, GFP_KERNEL);
2698         else
2699                 mem = vmalloc(size);
2700
2701         if (mem)
2702                 memset(mem, 0, size);
2703         return mem;
2704 }
2705
2706 /*
2707  * At destroying mem_cgroup, references from swap_cgroup can remain.
2708  * (scanning all at force_empty is too costly...)
2709  *
2710  * Instead of clearing all references at force_empty, we remember
2711  * the number of reference from swap_cgroup and free mem_cgroup when
2712  * it goes down to 0.
2713  *
2714  * Removal of cgroup itself succeeds regardless of refs from swap.
2715  */
2716
2717 static void __mem_cgroup_free(struct mem_cgroup *mem)
2718 {
2719         int node;
2720
2721         mem_cgroup_remove_from_trees(mem);
2722         free_css_id(&mem_cgroup_subsys, &mem->css);
2723
2724         for_each_node_state(node, N_POSSIBLE)
2725                 free_mem_cgroup_per_zone_info(mem, node);
2726
2727         if (mem_cgroup_size() < PAGE_SIZE)
2728                 kfree(mem);
2729         else
2730                 vfree(mem);
2731 }
2732
2733 static void mem_cgroup_get(struct mem_cgroup *mem)
2734 {
2735         atomic_inc(&mem->refcnt);
2736 }
2737
2738 static void mem_cgroup_put(struct mem_cgroup *mem)
2739 {
2740         if (atomic_dec_and_test(&mem->refcnt)) {
2741                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
2742                 __mem_cgroup_free(mem);
2743                 if (parent)
2744                         mem_cgroup_put(parent);
2745         }
2746 }
2747
2748 /*
2749  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
2750  */
2751 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
2752 {
2753         if (!mem->res.parent)
2754                 return NULL;
2755         return mem_cgroup_from_res_counter(mem->res.parent, res);
2756 }
2757
2758 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2759 static void __init enable_swap_cgroup(void)
2760 {
2761         if (!mem_cgroup_disabled() && really_do_swap_account)
2762                 do_swap_account = 1;
2763 }
2764 #else
2765 static void __init enable_swap_cgroup(void)
2766 {
2767 }
2768 #endif
2769
2770 static int mem_cgroup_soft_limit_tree_init(void)
2771 {
2772         struct mem_cgroup_tree_per_node *rtpn;
2773         struct mem_cgroup_tree_per_zone *rtpz;
2774         int tmp, node, zone;
2775
2776         for_each_node_state(node, N_POSSIBLE) {
2777                 tmp = node;
2778                 if (!node_state(node, N_NORMAL_MEMORY))
2779                         tmp = -1;
2780                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
2781                 if (!rtpn)
2782                         return 1;
2783
2784                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
2785
2786                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2787                         rtpz = &rtpn->rb_tree_per_zone[zone];
2788                         rtpz->rb_root = RB_ROOT;
2789                         spin_lock_init(&rtpz->lock);
2790                 }
2791         }
2792         return 0;
2793 }
2794
2795 static struct cgroup_subsys_state * __ref
2796 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
2797 {
2798         struct mem_cgroup *mem, *parent;
2799         long error = -ENOMEM;
2800         int node;
2801
2802         mem = mem_cgroup_alloc();
2803         if (!mem)
2804                 return ERR_PTR(error);
2805
2806         for_each_node_state(node, N_POSSIBLE)
2807                 if (alloc_mem_cgroup_per_zone_info(mem, node))
2808                         goto free_out;
2809
2810         /* root ? */
2811         if (cont->parent == NULL) {
2812                 enable_swap_cgroup();
2813                 parent = NULL;
2814                 root_mem_cgroup = mem;
2815                 if (mem_cgroup_soft_limit_tree_init())
2816                         goto free_out;
2817
2818         } else {
2819                 parent = mem_cgroup_from_cont(cont->parent);
2820                 mem->use_hierarchy = parent->use_hierarchy;
2821         }
2822
2823         if (parent && parent->use_hierarchy) {
2824                 res_counter_init(&mem->res, &parent->res);
2825                 res_counter_init(&mem->memsw, &parent->memsw);
2826                 /*
2827                  * We increment refcnt of the parent to ensure that we can
2828                  * safely access it on res_counter_charge/uncharge.
2829                  * This refcnt will be decremented when freeing this
2830                  * mem_cgroup(see mem_cgroup_put).
2831                  */
2832                 mem_cgroup_get(parent);
2833         } else {
2834                 res_counter_init(&mem->res, NULL);
2835                 res_counter_init(&mem->memsw, NULL);
2836         }
2837         mem->last_scanned_child = 0;
2838         spin_lock_init(&mem->reclaim_param_lock);
2839
2840         if (parent)
2841                 mem->swappiness = get_swappiness(parent);
2842         atomic_set(&mem->refcnt, 1);
2843         return &mem->css;
2844 free_out:
2845         __mem_cgroup_free(mem);
2846         root_mem_cgroup = NULL;
2847         return ERR_PTR(error);
2848 }
2849
2850 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
2851                                         struct cgroup *cont)
2852 {
2853         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2854
2855         return mem_cgroup_force_empty(mem, false);
2856 }
2857
2858 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
2859                                 struct cgroup *cont)
2860 {
2861         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2862
2863         mem_cgroup_put(mem);
2864 }
2865
2866 static int mem_cgroup_populate(struct cgroup_subsys *ss,
2867                                 struct cgroup *cont)
2868 {
2869         int ret;
2870
2871         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
2872                                 ARRAY_SIZE(mem_cgroup_files));
2873
2874         if (!ret)
2875                 ret = register_memsw_files(cont, ss);
2876         return ret;
2877 }
2878
2879 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
2880                                 struct cgroup *cont,
2881                                 struct cgroup *old_cont,
2882                                 struct task_struct *p,
2883                                 bool threadgroup)
2884 {
2885         mutex_lock(&memcg_tasklist);
2886         /*
2887          * FIXME: It's better to move charges of this process from old
2888          * memcg to new memcg. But it's just on TODO-List now.
2889          */
2890         mutex_unlock(&memcg_tasklist);
2891 }
2892
2893 struct cgroup_subsys mem_cgroup_subsys = {
2894         .name = "memory",
2895         .subsys_id = mem_cgroup_subsys_id,
2896         .create = mem_cgroup_create,
2897         .pre_destroy = mem_cgroup_pre_destroy,
2898         .destroy = mem_cgroup_destroy,
2899         .populate = mem_cgroup_populate,
2900         .attach = mem_cgroup_move_task,
2901         .early_init = 0,
2902         .use_id = 1,
2903 };
2904
2905 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2906
2907 static int __init disable_swap_account(char *s)
2908 {
2909         really_do_swap_account = 0;
2910         return 1;
2911 }
2912 __setup("noswapaccount", disable_swap_account);
2913 #endif