memcg: fix memory.memsw.usage_in_bytes for root cgroup
[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         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
74
75         MEM_CGROUP_STAT_NSTATS,
76 };
77
78 struct mem_cgroup_stat_cpu {
79         s64 count[MEM_CGROUP_STAT_NSTATS];
80 } ____cacheline_aligned_in_smp;
81
82 struct mem_cgroup_stat {
83         struct mem_cgroup_stat_cpu cpustat[0];
84 };
85
86 static inline void
87 __mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
88                                 enum mem_cgroup_stat_index idx)
89 {
90         stat->count[idx] = 0;
91 }
92
93 static inline s64
94 __mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
95                                 enum mem_cgroup_stat_index idx)
96 {
97         return stat->count[idx];
98 }
99
100 /*
101  * For accounting under irq disable, no need for increment preempt count.
102  */
103 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
104                 enum mem_cgroup_stat_index idx, int val)
105 {
106         stat->count[idx] += val;
107 }
108
109 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
110                 enum mem_cgroup_stat_index idx)
111 {
112         int cpu;
113         s64 ret = 0;
114         for_each_possible_cpu(cpu)
115                 ret += stat->cpustat[cpu].count[idx];
116         return ret;
117 }
118
119 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
120 {
121         s64 ret;
122
123         ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
124         ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
125         return ret;
126 }
127
128 /*
129  * per-zone information in memory controller.
130  */
131 struct mem_cgroup_per_zone {
132         /*
133          * spin_lock to protect the per cgroup LRU
134          */
135         struct list_head        lists[NR_LRU_LISTS];
136         unsigned long           count[NR_LRU_LISTS];
137
138         struct zone_reclaim_stat reclaim_stat;
139         struct rb_node          tree_node;      /* RB tree node */
140         unsigned long long      usage_in_excess;/* Set to the value by which */
141                                                 /* the soft limit is exceeded*/
142         bool                    on_tree;
143         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
144                                                 /* use container_of        */
145 };
146 /* Macro for accessing counter */
147 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
148
149 struct mem_cgroup_per_node {
150         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
151 };
152
153 struct mem_cgroup_lru_info {
154         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
155 };
156
157 /*
158  * Cgroups above their limits are maintained in a RB-Tree, independent of
159  * their hierarchy representation
160  */
161
162 struct mem_cgroup_tree_per_zone {
163         struct rb_root rb_root;
164         spinlock_t lock;
165 };
166
167 struct mem_cgroup_tree_per_node {
168         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
169 };
170
171 struct mem_cgroup_tree {
172         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
173 };
174
175 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176
177 /*
178  * The memory controller data structure. The memory controller controls both
179  * page cache and RSS per cgroup. We would eventually like to provide
180  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
181  * to help the administrator determine what knobs to tune.
182  *
183  * TODO: Add a water mark for the memory controller. Reclaim will begin when
184  * we hit the water mark. May be even add a low water mark, such that
185  * no reclaim occurs from a cgroup at it's low water mark, this is
186  * a feature that will be implemented much later in the future.
187  */
188 struct mem_cgroup {
189         struct cgroup_subsys_state css;
190         /*
191          * the counter to account for memory usage
192          */
193         struct res_counter res;
194         /*
195          * the counter to account for mem+swap usage.
196          */
197         struct res_counter memsw;
198         /*
199          * Per cgroup active and inactive list, similar to the
200          * per zone LRU lists.
201          */
202         struct mem_cgroup_lru_info info;
203
204         /*
205           protect against reclaim related member.
206         */
207         spinlock_t reclaim_param_lock;
208
209         int     prev_priority;  /* for recording reclaim priority */
210
211         /*
212          * While reclaiming in a hierarchy, we cache the last child we
213          * reclaimed from.
214          */
215         int last_scanned_child;
216         /*
217          * Should the accounting and control be hierarchical, per subtree?
218          */
219         bool use_hierarchy;
220         unsigned long   last_oom_jiffies;
221         atomic_t        refcnt;
222
223         unsigned int    swappiness;
224
225         /* set when res.limit == memsw.limit */
226         bool            memsw_is_minimum;
227
228         /*
229          * statistics. This must be placed at the end of memcg.
230          */
231         struct mem_cgroup_stat stat;
232 };
233
234 /*
235  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
236  * limit reclaim to prevent infinite loops, if they ever occur.
237  */
238 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
239 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
240
241 enum charge_type {
242         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
243         MEM_CGROUP_CHARGE_TYPE_MAPPED,
244         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
245         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
246         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
247         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
248         NR_CHARGE_TYPE,
249 };
250
251 /* only for here (for easy reading.) */
252 #define PCGF_CACHE      (1UL << PCG_CACHE)
253 #define PCGF_USED       (1UL << PCG_USED)
254 #define PCGF_LOCK       (1UL << PCG_LOCK)
255 /* Not used, but added here for completeness */
256 #define PCGF_ACCT       (1UL << PCG_ACCT)
257
258 /* for encoding cft->private value on file */
259 #define _MEM                    (0)
260 #define _MEMSWAP                (1)
261 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
262 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
263 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
264
265 /*
266  * Reclaim flags for mem_cgroup_hierarchical_reclaim
267  */
268 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
269 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
270 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
271 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
272 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
273 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
274
275 static void mem_cgroup_get(struct mem_cgroup *mem);
276 static void mem_cgroup_put(struct mem_cgroup *mem);
277 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
278
279 static struct mem_cgroup_per_zone *
280 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
281 {
282         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
283 }
284
285 static struct mem_cgroup_per_zone *
286 page_cgroup_zoneinfo(struct page_cgroup *pc)
287 {
288         struct mem_cgroup *mem = pc->mem_cgroup;
289         int nid = page_cgroup_nid(pc);
290         int zid = page_cgroup_zid(pc);
291
292         if (!mem)
293                 return NULL;
294
295         return mem_cgroup_zoneinfo(mem, nid, zid);
296 }
297
298 static struct mem_cgroup_tree_per_zone *
299 soft_limit_tree_node_zone(int nid, int zid)
300 {
301         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
302 }
303
304 static struct mem_cgroup_tree_per_zone *
305 soft_limit_tree_from_page(struct page *page)
306 {
307         int nid = page_to_nid(page);
308         int zid = page_zonenum(page);
309
310         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
311 }
312
313 static void
314 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
315                                 struct mem_cgroup_per_zone *mz,
316                                 struct mem_cgroup_tree_per_zone *mctz,
317                                 unsigned long long new_usage_in_excess)
318 {
319         struct rb_node **p = &mctz->rb_root.rb_node;
320         struct rb_node *parent = NULL;
321         struct mem_cgroup_per_zone *mz_node;
322
323         if (mz->on_tree)
324                 return;
325
326         mz->usage_in_excess = new_usage_in_excess;
327         if (!mz->usage_in_excess)
328                 return;
329         while (*p) {
330                 parent = *p;
331                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
332                                         tree_node);
333                 if (mz->usage_in_excess < mz_node->usage_in_excess)
334                         p = &(*p)->rb_left;
335                 /*
336                  * We can't avoid mem cgroups that are over their soft
337                  * limit by the same amount
338                  */
339                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
340                         p = &(*p)->rb_right;
341         }
342         rb_link_node(&mz->tree_node, parent, p);
343         rb_insert_color(&mz->tree_node, &mctz->rb_root);
344         mz->on_tree = true;
345 }
346
347 static void
348 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
349                                 struct mem_cgroup_per_zone *mz,
350                                 struct mem_cgroup_tree_per_zone *mctz)
351 {
352         if (!mz->on_tree)
353                 return;
354         rb_erase(&mz->tree_node, &mctz->rb_root);
355         mz->on_tree = false;
356 }
357
358 static void
359 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
360                                 struct mem_cgroup_per_zone *mz,
361                                 struct mem_cgroup_tree_per_zone *mctz)
362 {
363         spin_lock(&mctz->lock);
364         __mem_cgroup_remove_exceeded(mem, mz, mctz);
365         spin_unlock(&mctz->lock);
366 }
367
368 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
369 {
370         bool ret = false;
371         int cpu;
372         s64 val;
373         struct mem_cgroup_stat_cpu *cpustat;
374
375         cpu = get_cpu();
376         cpustat = &mem->stat.cpustat[cpu];
377         val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
378         if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
379                 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
380                 ret = true;
381         }
382         put_cpu();
383         return ret;
384 }
385
386 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
387 {
388         unsigned long long excess;
389         struct mem_cgroup_per_zone *mz;
390         struct mem_cgroup_tree_per_zone *mctz;
391         int nid = page_to_nid(page);
392         int zid = page_zonenum(page);
393         mctz = soft_limit_tree_from_page(page);
394
395         /*
396          * Necessary to update all ancestors when hierarchy is used.
397          * because their event counter is not touched.
398          */
399         for (; mem; mem = parent_mem_cgroup(mem)) {
400                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
401                 excess = res_counter_soft_limit_excess(&mem->res);
402                 /*
403                  * We have to update the tree if mz is on RB-tree or
404                  * mem is over its softlimit.
405                  */
406                 if (excess || mz->on_tree) {
407                         spin_lock(&mctz->lock);
408                         /* if on-tree, remove it */
409                         if (mz->on_tree)
410                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
411                         /*
412                          * Insert again. mz->usage_in_excess will be updated.
413                          * If excess is 0, no tree ops.
414                          */
415                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
416                         spin_unlock(&mctz->lock);
417                 }
418         }
419 }
420
421 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
422 {
423         int node, zone;
424         struct mem_cgroup_per_zone *mz;
425         struct mem_cgroup_tree_per_zone *mctz;
426
427         for_each_node_state(node, N_POSSIBLE) {
428                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
429                         mz = mem_cgroup_zoneinfo(mem, node, zone);
430                         mctz = soft_limit_tree_node_zone(node, zone);
431                         mem_cgroup_remove_exceeded(mem, mz, mctz);
432                 }
433         }
434 }
435
436 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
437 {
438         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
439 }
440
441 static struct mem_cgroup_per_zone *
442 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
443 {
444         struct rb_node *rightmost = NULL;
445         struct mem_cgroup_per_zone *mz;
446
447 retry:
448         mz = NULL;
449         rightmost = rb_last(&mctz->rb_root);
450         if (!rightmost)
451                 goto done;              /* Nothing to reclaim from */
452
453         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
454         /*
455          * Remove the node now but someone else can add it back,
456          * we will to add it back at the end of reclaim to its correct
457          * position in the tree.
458          */
459         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
460         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
461                 !css_tryget(&mz->mem->css))
462                 goto retry;
463 done:
464         return mz;
465 }
466
467 static struct mem_cgroup_per_zone *
468 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
469 {
470         struct mem_cgroup_per_zone *mz;
471
472         spin_lock(&mctz->lock);
473         mz = __mem_cgroup_largest_soft_limit_node(mctz);
474         spin_unlock(&mctz->lock);
475         return mz;
476 }
477
478 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
479                                          bool charge)
480 {
481         int val = (charge) ? 1 : -1;
482         struct mem_cgroup_stat *stat = &mem->stat;
483         struct mem_cgroup_stat_cpu *cpustat;
484         int cpu = get_cpu();
485
486         cpustat = &stat->cpustat[cpu];
487         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
488         put_cpu();
489 }
490
491 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
492                                          struct page_cgroup *pc,
493                                          bool charge)
494 {
495         int val = (charge) ? 1 : -1;
496         struct mem_cgroup_stat *stat = &mem->stat;
497         struct mem_cgroup_stat_cpu *cpustat;
498         int cpu = get_cpu();
499
500         cpustat = &stat->cpustat[cpu];
501         if (PageCgroupCache(pc))
502                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
503         else
504                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
505
506         if (charge)
507                 __mem_cgroup_stat_add_safe(cpustat,
508                                 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
509         else
510                 __mem_cgroup_stat_add_safe(cpustat,
511                                 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
512         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1);
513         put_cpu();
514 }
515
516 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
517                                         enum lru_list idx)
518 {
519         int nid, zid;
520         struct mem_cgroup_per_zone *mz;
521         u64 total = 0;
522
523         for_each_online_node(nid)
524                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
525                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
526                         total += MEM_CGROUP_ZSTAT(mz, idx);
527                 }
528         return total;
529 }
530
531 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
532 {
533         return container_of(cgroup_subsys_state(cont,
534                                 mem_cgroup_subsys_id), struct mem_cgroup,
535                                 css);
536 }
537
538 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
539 {
540         /*
541          * mm_update_next_owner() may clear mm->owner to NULL
542          * if it races with swapoff, page migration, etc.
543          * So this can be called with p == NULL.
544          */
545         if (unlikely(!p))
546                 return NULL;
547
548         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
549                                 struct mem_cgroup, css);
550 }
551
552 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
553 {
554         struct mem_cgroup *mem = NULL;
555
556         if (!mm)
557                 return NULL;
558         /*
559          * Because we have no locks, mm->owner's may be being moved to other
560          * cgroup. We use css_tryget() here even if this looks
561          * pessimistic (rather than adding locks here).
562          */
563         rcu_read_lock();
564         do {
565                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
566                 if (unlikely(!mem))
567                         break;
568         } while (!css_tryget(&mem->css));
569         rcu_read_unlock();
570         return mem;
571 }
572
573 /*
574  * Call callback function against all cgroup under hierarchy tree.
575  */
576 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
577                           int (*func)(struct mem_cgroup *, void *))
578 {
579         int found, ret, nextid;
580         struct cgroup_subsys_state *css;
581         struct mem_cgroup *mem;
582
583         if (!root->use_hierarchy)
584                 return (*func)(root, data);
585
586         nextid = 1;
587         do {
588                 ret = 0;
589                 mem = NULL;
590
591                 rcu_read_lock();
592                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
593                                    &found);
594                 if (css && css_tryget(css))
595                         mem = container_of(css, struct mem_cgroup, css);
596                 rcu_read_unlock();
597
598                 if (mem) {
599                         ret = (*func)(mem, data);
600                         css_put(&mem->css);
601                 }
602                 nextid = found + 1;
603         } while (!ret && css);
604
605         return ret;
606 }
607
608 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
609 {
610         return (mem == root_mem_cgroup);
611 }
612
613 /*
614  * Following LRU functions are allowed to be used without PCG_LOCK.
615  * Operations are called by routine of global LRU independently from memcg.
616  * What we have to take care of here is validness of pc->mem_cgroup.
617  *
618  * Changes to pc->mem_cgroup happens when
619  * 1. charge
620  * 2. moving account
621  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
622  * It is added to LRU before charge.
623  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
624  * When moving account, the page is not on LRU. It's isolated.
625  */
626
627 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
628 {
629         struct page_cgroup *pc;
630         struct mem_cgroup_per_zone *mz;
631
632         if (mem_cgroup_disabled())
633                 return;
634         pc = lookup_page_cgroup(page);
635         /* can happen while we handle swapcache. */
636         if (!TestClearPageCgroupAcctLRU(pc))
637                 return;
638         VM_BUG_ON(!pc->mem_cgroup);
639         /*
640          * We don't check PCG_USED bit. It's cleared when the "page" is finally
641          * removed from global LRU.
642          */
643         mz = page_cgroup_zoneinfo(pc);
644         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
645         if (mem_cgroup_is_root(pc->mem_cgroup))
646                 return;
647         VM_BUG_ON(list_empty(&pc->lru));
648         list_del_init(&pc->lru);
649         return;
650 }
651
652 void mem_cgroup_del_lru(struct page *page)
653 {
654         mem_cgroup_del_lru_list(page, page_lru(page));
655 }
656
657 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
658 {
659         struct mem_cgroup_per_zone *mz;
660         struct page_cgroup *pc;
661
662         if (mem_cgroup_disabled())
663                 return;
664
665         pc = lookup_page_cgroup(page);
666         /*
667          * Used bit is set without atomic ops but after smp_wmb().
668          * For making pc->mem_cgroup visible, insert smp_rmb() here.
669          */
670         smp_rmb();
671         /* unused or root page is not rotated. */
672         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
673                 return;
674         mz = page_cgroup_zoneinfo(pc);
675         list_move(&pc->lru, &mz->lists[lru]);
676 }
677
678 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
679 {
680         struct page_cgroup *pc;
681         struct mem_cgroup_per_zone *mz;
682
683         if (mem_cgroup_disabled())
684                 return;
685         pc = lookup_page_cgroup(page);
686         VM_BUG_ON(PageCgroupAcctLRU(pc));
687         /*
688          * Used bit is set without atomic ops but after smp_wmb().
689          * For making pc->mem_cgroup visible, insert smp_rmb() here.
690          */
691         smp_rmb();
692         if (!PageCgroupUsed(pc))
693                 return;
694
695         mz = page_cgroup_zoneinfo(pc);
696         MEM_CGROUP_ZSTAT(mz, lru) += 1;
697         SetPageCgroupAcctLRU(pc);
698         if (mem_cgroup_is_root(pc->mem_cgroup))
699                 return;
700         list_add(&pc->lru, &mz->lists[lru]);
701 }
702
703 /*
704  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
705  * lru because the page may.be reused after it's fully uncharged (because of
706  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
707  * it again. This function is only used to charge SwapCache. It's done under
708  * lock_page and expected that zone->lru_lock is never held.
709  */
710 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
711 {
712         unsigned long flags;
713         struct zone *zone = page_zone(page);
714         struct page_cgroup *pc = lookup_page_cgroup(page);
715
716         spin_lock_irqsave(&zone->lru_lock, flags);
717         /*
718          * Forget old LRU when this page_cgroup is *not* used. This Used bit
719          * is guarded by lock_page() because the page is SwapCache.
720          */
721         if (!PageCgroupUsed(pc))
722                 mem_cgroup_del_lru_list(page, page_lru(page));
723         spin_unlock_irqrestore(&zone->lru_lock, flags);
724 }
725
726 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
727 {
728         unsigned long flags;
729         struct zone *zone = page_zone(page);
730         struct page_cgroup *pc = lookup_page_cgroup(page);
731
732         spin_lock_irqsave(&zone->lru_lock, flags);
733         /* link when the page is linked to LRU but page_cgroup isn't */
734         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
735                 mem_cgroup_add_lru_list(page, page_lru(page));
736         spin_unlock_irqrestore(&zone->lru_lock, flags);
737 }
738
739
740 void mem_cgroup_move_lists(struct page *page,
741                            enum lru_list from, enum lru_list to)
742 {
743         if (mem_cgroup_disabled())
744                 return;
745         mem_cgroup_del_lru_list(page, from);
746         mem_cgroup_add_lru_list(page, to);
747 }
748
749 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
750 {
751         int ret;
752         struct mem_cgroup *curr = NULL;
753
754         task_lock(task);
755         rcu_read_lock();
756         curr = try_get_mem_cgroup_from_mm(task->mm);
757         rcu_read_unlock();
758         task_unlock(task);
759         if (!curr)
760                 return 0;
761         if (curr->use_hierarchy)
762                 ret = css_is_ancestor(&curr->css, &mem->css);
763         else
764                 ret = (curr == mem);
765         css_put(&curr->css);
766         return ret;
767 }
768
769 /*
770  * prev_priority control...this will be used in memory reclaim path.
771  */
772 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
773 {
774         int prev_priority;
775
776         spin_lock(&mem->reclaim_param_lock);
777         prev_priority = mem->prev_priority;
778         spin_unlock(&mem->reclaim_param_lock);
779
780         return prev_priority;
781 }
782
783 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
784 {
785         spin_lock(&mem->reclaim_param_lock);
786         if (priority < mem->prev_priority)
787                 mem->prev_priority = priority;
788         spin_unlock(&mem->reclaim_param_lock);
789 }
790
791 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
792 {
793         spin_lock(&mem->reclaim_param_lock);
794         mem->prev_priority = priority;
795         spin_unlock(&mem->reclaim_param_lock);
796 }
797
798 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
799 {
800         unsigned long active;
801         unsigned long inactive;
802         unsigned long gb;
803         unsigned long inactive_ratio;
804
805         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
806         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
807
808         gb = (inactive + active) >> (30 - PAGE_SHIFT);
809         if (gb)
810                 inactive_ratio = int_sqrt(10 * gb);
811         else
812                 inactive_ratio = 1;
813
814         if (present_pages) {
815                 present_pages[0] = inactive;
816                 present_pages[1] = active;
817         }
818
819         return inactive_ratio;
820 }
821
822 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
823 {
824         unsigned long active;
825         unsigned long inactive;
826         unsigned long present_pages[2];
827         unsigned long inactive_ratio;
828
829         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
830
831         inactive = present_pages[0];
832         active = present_pages[1];
833
834         if (inactive * inactive_ratio < active)
835                 return 1;
836
837         return 0;
838 }
839
840 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
841 {
842         unsigned long active;
843         unsigned long inactive;
844
845         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
846         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
847
848         return (active > inactive);
849 }
850
851 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
852                                        struct zone *zone,
853                                        enum lru_list lru)
854 {
855         int nid = zone->zone_pgdat->node_id;
856         int zid = zone_idx(zone);
857         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
858
859         return MEM_CGROUP_ZSTAT(mz, lru);
860 }
861
862 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
863                                                       struct zone *zone)
864 {
865         int nid = zone->zone_pgdat->node_id;
866         int zid = zone_idx(zone);
867         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
868
869         return &mz->reclaim_stat;
870 }
871
872 struct zone_reclaim_stat *
873 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
874 {
875         struct page_cgroup *pc;
876         struct mem_cgroup_per_zone *mz;
877
878         if (mem_cgroup_disabled())
879                 return NULL;
880
881         pc = lookup_page_cgroup(page);
882         /*
883          * Used bit is set without atomic ops but after smp_wmb().
884          * For making pc->mem_cgroup visible, insert smp_rmb() here.
885          */
886         smp_rmb();
887         if (!PageCgroupUsed(pc))
888                 return NULL;
889
890         mz = page_cgroup_zoneinfo(pc);
891         if (!mz)
892                 return NULL;
893
894         return &mz->reclaim_stat;
895 }
896
897 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
898                                         struct list_head *dst,
899                                         unsigned long *scanned, int order,
900                                         int mode, struct zone *z,
901                                         struct mem_cgroup *mem_cont,
902                                         int active, int file)
903 {
904         unsigned long nr_taken = 0;
905         struct page *page;
906         unsigned long scan;
907         LIST_HEAD(pc_list);
908         struct list_head *src;
909         struct page_cgroup *pc, *tmp;
910         int nid = z->zone_pgdat->node_id;
911         int zid = zone_idx(z);
912         struct mem_cgroup_per_zone *mz;
913         int lru = LRU_FILE * file + active;
914         int ret;
915
916         BUG_ON(!mem_cont);
917         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
918         src = &mz->lists[lru];
919
920         scan = 0;
921         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
922                 if (scan >= nr_to_scan)
923                         break;
924
925                 page = pc->page;
926                 if (unlikely(!PageCgroupUsed(pc)))
927                         continue;
928                 if (unlikely(!PageLRU(page)))
929                         continue;
930
931                 scan++;
932                 ret = __isolate_lru_page(page, mode, file);
933                 switch (ret) {
934                 case 0:
935                         list_move(&page->lru, dst);
936                         mem_cgroup_del_lru(page);
937                         nr_taken++;
938                         break;
939                 case -EBUSY:
940                         /* we don't affect global LRU but rotate in our LRU */
941                         mem_cgroup_rotate_lru_list(page, page_lru(page));
942                         break;
943                 default:
944                         break;
945                 }
946         }
947
948         *scanned = scan;
949         return nr_taken;
950 }
951
952 #define mem_cgroup_from_res_counter(counter, member)    \
953         container_of(counter, struct mem_cgroup, member)
954
955 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
956 {
957         if (do_swap_account) {
958                 if (res_counter_check_under_limit(&mem->res) &&
959                         res_counter_check_under_limit(&mem->memsw))
960                         return true;
961         } else
962                 if (res_counter_check_under_limit(&mem->res))
963                         return true;
964         return false;
965 }
966
967 static unsigned int get_swappiness(struct mem_cgroup *memcg)
968 {
969         struct cgroup *cgrp = memcg->css.cgroup;
970         unsigned int swappiness;
971
972         /* root ? */
973         if (cgrp->parent == NULL)
974                 return vm_swappiness;
975
976         spin_lock(&memcg->reclaim_param_lock);
977         swappiness = memcg->swappiness;
978         spin_unlock(&memcg->reclaim_param_lock);
979
980         return swappiness;
981 }
982
983 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
984 {
985         int *val = data;
986         (*val)++;
987         return 0;
988 }
989
990 /**
991  * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
992  * @memcg: The memory cgroup that went over limit
993  * @p: Task that is going to be killed
994  *
995  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
996  * enabled
997  */
998 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
999 {
1000         struct cgroup *task_cgrp;
1001         struct cgroup *mem_cgrp;
1002         /*
1003          * Need a buffer in BSS, can't rely on allocations. The code relies
1004          * on the assumption that OOM is serialized for memory controller.
1005          * If this assumption is broken, revisit this code.
1006          */
1007         static char memcg_name[PATH_MAX];
1008         int ret;
1009
1010         if (!memcg)
1011                 return;
1012
1013
1014         rcu_read_lock();
1015
1016         mem_cgrp = memcg->css.cgroup;
1017         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1018
1019         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1020         if (ret < 0) {
1021                 /*
1022                  * Unfortunately, we are unable to convert to a useful name
1023                  * But we'll still print out the usage information
1024                  */
1025                 rcu_read_unlock();
1026                 goto done;
1027         }
1028         rcu_read_unlock();
1029
1030         printk(KERN_INFO "Task in %s killed", memcg_name);
1031
1032         rcu_read_lock();
1033         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1034         if (ret < 0) {
1035                 rcu_read_unlock();
1036                 goto done;
1037         }
1038         rcu_read_unlock();
1039
1040         /*
1041          * Continues from above, so we don't need an KERN_ level
1042          */
1043         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1044 done:
1045
1046         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1047                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1048                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1049                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1050         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1051                 "failcnt %llu\n",
1052                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1053                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1054                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1055 }
1056
1057 /*
1058  * This function returns the number of memcg under hierarchy tree. Returns
1059  * 1(self count) if no children.
1060  */
1061 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1062 {
1063         int num = 0;
1064         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1065         return num;
1066 }
1067
1068 /*
1069  * Visit the first child (need not be the first child as per the ordering
1070  * of the cgroup list, since we track last_scanned_child) of @mem and use
1071  * that to reclaim free pages from.
1072  */
1073 static struct mem_cgroup *
1074 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1075 {
1076         struct mem_cgroup *ret = NULL;
1077         struct cgroup_subsys_state *css;
1078         int nextid, found;
1079
1080         if (!root_mem->use_hierarchy) {
1081                 css_get(&root_mem->css);
1082                 ret = root_mem;
1083         }
1084
1085         while (!ret) {
1086                 rcu_read_lock();
1087                 nextid = root_mem->last_scanned_child + 1;
1088                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1089                                    &found);
1090                 if (css && css_tryget(css))
1091                         ret = container_of(css, struct mem_cgroup, css);
1092
1093                 rcu_read_unlock();
1094                 /* Updates scanning parameter */
1095                 spin_lock(&root_mem->reclaim_param_lock);
1096                 if (!css) {
1097                         /* this means start scan from ID:1 */
1098                         root_mem->last_scanned_child = 0;
1099                 } else
1100                         root_mem->last_scanned_child = found;
1101                 spin_unlock(&root_mem->reclaim_param_lock);
1102         }
1103
1104         return ret;
1105 }
1106
1107 /*
1108  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1109  * we reclaimed from, so that we don't end up penalizing one child extensively
1110  * based on its position in the children list.
1111  *
1112  * root_mem is the original ancestor that we've been reclaim from.
1113  *
1114  * We give up and return to the caller when we visit root_mem twice.
1115  * (other groups can be removed while we're walking....)
1116  *
1117  * If shrink==true, for avoiding to free too much, this returns immedieately.
1118  */
1119 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1120                                                 struct zone *zone,
1121                                                 gfp_t gfp_mask,
1122                                                 unsigned long reclaim_options)
1123 {
1124         struct mem_cgroup *victim;
1125         int ret, total = 0;
1126         int loop = 0;
1127         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1128         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1129         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1130         unsigned long excess = mem_cgroup_get_excess(root_mem);
1131
1132         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1133         if (root_mem->memsw_is_minimum)
1134                 noswap = true;
1135
1136         while (1) {
1137                 victim = mem_cgroup_select_victim(root_mem);
1138                 if (victim == root_mem) {
1139                         loop++;
1140                         if (loop >= 2) {
1141                                 /*
1142                                  * If we have not been able to reclaim
1143                                  * anything, it might because there are
1144                                  * no reclaimable pages under this hierarchy
1145                                  */
1146                                 if (!check_soft || !total) {
1147                                         css_put(&victim->css);
1148                                         break;
1149                                 }
1150                                 /*
1151                                  * We want to do more targetted reclaim.
1152                                  * excess >> 2 is not to excessive so as to
1153                                  * reclaim too much, nor too less that we keep
1154                                  * coming back to reclaim from this cgroup
1155                                  */
1156                                 if (total >= (excess >> 2) ||
1157                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1158                                         css_put(&victim->css);
1159                                         break;
1160                                 }
1161                         }
1162                 }
1163                 if (!mem_cgroup_local_usage(&victim->stat)) {
1164                         /* this cgroup's local usage == 0 */
1165                         css_put(&victim->css);
1166                         continue;
1167                 }
1168                 /* we use swappiness of local cgroup */
1169                 if (check_soft)
1170                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1171                                 noswap, get_swappiness(victim), zone,
1172                                 zone->zone_pgdat->node_id);
1173                 else
1174                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1175                                                 noswap, get_swappiness(victim));
1176                 css_put(&victim->css);
1177                 /*
1178                  * At shrinking usage, we can't check we should stop here or
1179                  * reclaim more. It's depends on callers. last_scanned_child
1180                  * will work enough for keeping fairness under tree.
1181                  */
1182                 if (shrink)
1183                         return ret;
1184                 total += ret;
1185                 if (check_soft) {
1186                         if (res_counter_check_under_soft_limit(&root_mem->res))
1187                                 return total;
1188                 } else if (mem_cgroup_check_under_limit(root_mem))
1189                         return 1 + total;
1190         }
1191         return total;
1192 }
1193
1194 bool mem_cgroup_oom_called(struct task_struct *task)
1195 {
1196         bool ret = false;
1197         struct mem_cgroup *mem;
1198         struct mm_struct *mm;
1199
1200         rcu_read_lock();
1201         mm = task->mm;
1202         if (!mm)
1203                 mm = &init_mm;
1204         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1205         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1206                 ret = true;
1207         rcu_read_unlock();
1208         return ret;
1209 }
1210
1211 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
1212 {
1213         mem->last_oom_jiffies = jiffies;
1214         return 0;
1215 }
1216
1217 static void record_last_oom(struct mem_cgroup *mem)
1218 {
1219         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
1220 }
1221
1222 /*
1223  * Currently used to update mapped file statistics, but the routine can be
1224  * generalized to update other statistics as well.
1225  */
1226 void mem_cgroup_update_mapped_file_stat(struct page *page, int val)
1227 {
1228         struct mem_cgroup *mem;
1229         struct mem_cgroup_stat *stat;
1230         struct mem_cgroup_stat_cpu *cpustat;
1231         int cpu;
1232         struct page_cgroup *pc;
1233
1234         if (!page_is_file_cache(page))
1235                 return;
1236
1237         pc = lookup_page_cgroup(page);
1238         if (unlikely(!pc))
1239                 return;
1240
1241         lock_page_cgroup(pc);
1242         mem = pc->mem_cgroup;
1243         if (!mem)
1244                 goto done;
1245
1246         if (!PageCgroupUsed(pc))
1247                 goto done;
1248
1249         /*
1250          * Preemption is already disabled, we don't need get_cpu()
1251          */
1252         cpu = smp_processor_id();
1253         stat = &mem->stat;
1254         cpustat = &stat->cpustat[cpu];
1255
1256         __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val);
1257 done:
1258         unlock_page_cgroup(pc);
1259 }
1260
1261 /*
1262  * Unlike exported interface, "oom" parameter is added. if oom==true,
1263  * oom-killer can be invoked.
1264  */
1265 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1266                         gfp_t gfp_mask, struct mem_cgroup **memcg,
1267                         bool oom, struct page *page)
1268 {
1269         struct mem_cgroup *mem, *mem_over_limit;
1270         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1271         struct res_counter *fail_res;
1272
1273         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1274                 /* Don't account this! */
1275                 *memcg = NULL;
1276                 return 0;
1277         }
1278
1279         /*
1280          * We always charge the cgroup the mm_struct belongs to.
1281          * The mm_struct's mem_cgroup changes on task migration if the
1282          * thread group leader migrates. It's possible that mm is not
1283          * set, if so charge the init_mm (happens for pagecache usage).
1284          */
1285         mem = *memcg;
1286         if (likely(!mem)) {
1287                 mem = try_get_mem_cgroup_from_mm(mm);
1288                 *memcg = mem;
1289         } else {
1290                 css_get(&mem->css);
1291         }
1292         if (unlikely(!mem))
1293                 return 0;
1294
1295         VM_BUG_ON(css_is_removed(&mem->css));
1296
1297         while (1) {
1298                 int ret = 0;
1299                 unsigned long flags = 0;
1300
1301                 if (mem_cgroup_is_root(mem))
1302                         goto done;
1303                 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
1304                 if (likely(!ret)) {
1305                         if (!do_swap_account)
1306                                 break;
1307                         ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
1308                                                         &fail_res);
1309                         if (likely(!ret))
1310                                 break;
1311                         /* mem+swap counter fails */
1312                         res_counter_uncharge(&mem->res, PAGE_SIZE);
1313                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1314                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1315                                                                         memsw);
1316                 } else
1317                         /* mem counter fails */
1318                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1319                                                                         res);
1320
1321                 if (!(gfp_mask & __GFP_WAIT))
1322                         goto nomem;
1323
1324                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1325                                                 gfp_mask, flags);
1326                 if (ret)
1327                         continue;
1328
1329                 /*
1330                  * try_to_free_mem_cgroup_pages() might not give us a full
1331                  * picture of reclaim. Some pages are reclaimed and might be
1332                  * moved to swap cache or just unmapped from the cgroup.
1333                  * Check the limit again to see if the reclaim reduced the
1334                  * current usage of the cgroup before giving up
1335                  *
1336                  */
1337                 if (mem_cgroup_check_under_limit(mem_over_limit))
1338                         continue;
1339
1340                 if (!nr_retries--) {
1341                         if (oom) {
1342                                 mutex_lock(&memcg_tasklist);
1343                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1344                                 mutex_unlock(&memcg_tasklist);
1345                                 record_last_oom(mem_over_limit);
1346                         }
1347                         goto nomem;
1348                 }
1349         }
1350         /*
1351          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1352          * if they exceeds softlimit.
1353          */
1354         if (mem_cgroup_soft_limit_check(mem))
1355                 mem_cgroup_update_tree(mem, page);
1356 done:
1357         return 0;
1358 nomem:
1359         css_put(&mem->css);
1360         return -ENOMEM;
1361 }
1362
1363 /*
1364  * A helper function to get mem_cgroup from ID. must be called under
1365  * rcu_read_lock(). The caller must check css_is_removed() or some if
1366  * it's concern. (dropping refcnt from swap can be called against removed
1367  * memcg.)
1368  */
1369 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1370 {
1371         struct cgroup_subsys_state *css;
1372
1373         /* ID 0 is unused ID */
1374         if (!id)
1375                 return NULL;
1376         css = css_lookup(&mem_cgroup_subsys, id);
1377         if (!css)
1378                 return NULL;
1379         return container_of(css, struct mem_cgroup, css);
1380 }
1381
1382 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1383 {
1384         struct mem_cgroup *mem;
1385         struct page_cgroup *pc;
1386         unsigned short id;
1387         swp_entry_t ent;
1388
1389         VM_BUG_ON(!PageLocked(page));
1390
1391         if (!PageSwapCache(page))
1392                 return NULL;
1393
1394         pc = lookup_page_cgroup(page);
1395         lock_page_cgroup(pc);
1396         if (PageCgroupUsed(pc)) {
1397                 mem = pc->mem_cgroup;
1398                 if (mem && !css_tryget(&mem->css))
1399                         mem = NULL;
1400         } else {
1401                 ent.val = page_private(page);
1402                 id = lookup_swap_cgroup(ent);
1403                 rcu_read_lock();
1404                 mem = mem_cgroup_lookup(id);
1405                 if (mem && !css_tryget(&mem->css))
1406                         mem = NULL;
1407                 rcu_read_unlock();
1408         }
1409         unlock_page_cgroup(pc);
1410         return mem;
1411 }
1412
1413 /*
1414  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1415  * USED state. If already USED, uncharge and return.
1416  */
1417
1418 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1419                                      struct page_cgroup *pc,
1420                                      enum charge_type ctype)
1421 {
1422         /* try_charge() can return NULL to *memcg, taking care of it. */
1423         if (!mem)
1424                 return;
1425
1426         lock_page_cgroup(pc);
1427         if (unlikely(PageCgroupUsed(pc))) {
1428                 unlock_page_cgroup(pc);
1429                 if (!mem_cgroup_is_root(mem)) {
1430                         res_counter_uncharge(&mem->res, PAGE_SIZE);
1431                         if (do_swap_account)
1432                                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1433                 }
1434                 css_put(&mem->css);
1435                 return;
1436         }
1437
1438         pc->mem_cgroup = mem;
1439         /*
1440          * We access a page_cgroup asynchronously without lock_page_cgroup().
1441          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1442          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1443          * before USED bit, we need memory barrier here.
1444          * See mem_cgroup_add_lru_list(), etc.
1445          */
1446         smp_wmb();
1447         switch (ctype) {
1448         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1449         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1450                 SetPageCgroupCache(pc);
1451                 SetPageCgroupUsed(pc);
1452                 break;
1453         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1454                 ClearPageCgroupCache(pc);
1455                 SetPageCgroupUsed(pc);
1456                 break;
1457         default:
1458                 break;
1459         }
1460
1461         mem_cgroup_charge_statistics(mem, pc, true);
1462
1463         unlock_page_cgroup(pc);
1464 }
1465
1466 /**
1467  * mem_cgroup_move_account - move account of the page
1468  * @pc: page_cgroup of the page.
1469  * @from: mem_cgroup which the page is moved from.
1470  * @to: mem_cgroup which the page is moved to. @from != @to.
1471  *
1472  * The caller must confirm following.
1473  * - page is not on LRU (isolate_page() is useful.)
1474  *
1475  * returns 0 at success,
1476  * returns -EBUSY when lock is busy or "pc" is unstable.
1477  *
1478  * This function does "uncharge" from old cgroup but doesn't do "charge" to
1479  * new cgroup. It should be done by a caller.
1480  */
1481
1482 static int mem_cgroup_move_account(struct page_cgroup *pc,
1483         struct mem_cgroup *from, struct mem_cgroup *to)
1484 {
1485         struct mem_cgroup_per_zone *from_mz, *to_mz;
1486         int nid, zid;
1487         int ret = -EBUSY;
1488         struct page *page;
1489         int cpu;
1490         struct mem_cgroup_stat *stat;
1491         struct mem_cgroup_stat_cpu *cpustat;
1492
1493         VM_BUG_ON(from == to);
1494         VM_BUG_ON(PageLRU(pc->page));
1495
1496         nid = page_cgroup_nid(pc);
1497         zid = page_cgroup_zid(pc);
1498         from_mz =  mem_cgroup_zoneinfo(from, nid, zid);
1499         to_mz =  mem_cgroup_zoneinfo(to, nid, zid);
1500
1501         if (!trylock_page_cgroup(pc))
1502                 return ret;
1503
1504         if (!PageCgroupUsed(pc))
1505                 goto out;
1506
1507         if (pc->mem_cgroup != from)
1508                 goto out;
1509
1510         if (!mem_cgroup_is_root(from))
1511                 res_counter_uncharge(&from->res, PAGE_SIZE);
1512         mem_cgroup_charge_statistics(from, pc, false);
1513
1514         page = pc->page;
1515         if (page_is_file_cache(page) && page_mapped(page)) {
1516                 cpu = smp_processor_id();
1517                 /* Update mapped_file data for mem_cgroup "from" */
1518                 stat = &from->stat;
1519                 cpustat = &stat->cpustat[cpu];
1520                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1521                                                 -1);
1522
1523                 /* Update mapped_file data for mem_cgroup "to" */
1524                 stat = &to->stat;
1525                 cpustat = &stat->cpustat[cpu];
1526                 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1527                                                 1);
1528         }
1529
1530         if (do_swap_account && !mem_cgroup_is_root(from))
1531                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1532         css_put(&from->css);
1533
1534         css_get(&to->css);
1535         pc->mem_cgroup = to;
1536         mem_cgroup_charge_statistics(to, pc, true);
1537         ret = 0;
1538 out:
1539         unlock_page_cgroup(pc);
1540         /*
1541          * We charges against "to" which may not have any tasks. Then, "to"
1542          * can be under rmdir(). But in current implementation, caller of
1543          * this function is just force_empty() and it's garanteed that
1544          * "to" is never removed. So, we don't check rmdir status here.
1545          */
1546         return ret;
1547 }
1548
1549 /*
1550  * move charges to its parent.
1551  */
1552
1553 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1554                                   struct mem_cgroup *child,
1555                                   gfp_t gfp_mask)
1556 {
1557         struct page *page = pc->page;
1558         struct cgroup *cg = child->css.cgroup;
1559         struct cgroup *pcg = cg->parent;
1560         struct mem_cgroup *parent;
1561         int ret;
1562
1563         /* Is ROOT ? */
1564         if (!pcg)
1565                 return -EINVAL;
1566
1567
1568         parent = mem_cgroup_from_cont(pcg);
1569
1570
1571         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page);
1572         if (ret || !parent)
1573                 return ret;
1574
1575         if (!get_page_unless_zero(page)) {
1576                 ret = -EBUSY;
1577                 goto uncharge;
1578         }
1579
1580         ret = isolate_lru_page(page);
1581
1582         if (ret)
1583                 goto cancel;
1584
1585         ret = mem_cgroup_move_account(pc, child, parent);
1586
1587         putback_lru_page(page);
1588         if (!ret) {
1589                 put_page(page);
1590                 /* drop extra refcnt by try_charge() */
1591                 css_put(&parent->css);
1592                 return 0;
1593         }
1594
1595 cancel:
1596         put_page(page);
1597 uncharge:
1598         /* drop extra refcnt by try_charge() */
1599         css_put(&parent->css);
1600         /* uncharge if move fails */
1601         if (!mem_cgroup_is_root(parent)) {
1602                 res_counter_uncharge(&parent->res, PAGE_SIZE);
1603                 if (do_swap_account)
1604                         res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1605         }
1606         return ret;
1607 }
1608
1609 /*
1610  * Charge the memory controller for page usage.
1611  * Return
1612  * 0 if the charge was successful
1613  * < 0 if the cgroup is over its limit
1614  */
1615 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1616                                 gfp_t gfp_mask, enum charge_type ctype,
1617                                 struct mem_cgroup *memcg)
1618 {
1619         struct mem_cgroup *mem;
1620         struct page_cgroup *pc;
1621         int ret;
1622
1623         pc = lookup_page_cgroup(page);
1624         /* can happen at boot */
1625         if (unlikely(!pc))
1626                 return 0;
1627         prefetchw(pc);
1628
1629         mem = memcg;
1630         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page);
1631         if (ret || !mem)
1632                 return ret;
1633
1634         __mem_cgroup_commit_charge(mem, pc, ctype);
1635         return 0;
1636 }
1637
1638 int mem_cgroup_newpage_charge(struct page *page,
1639                               struct mm_struct *mm, gfp_t gfp_mask)
1640 {
1641         if (mem_cgroup_disabled())
1642                 return 0;
1643         if (PageCompound(page))
1644                 return 0;
1645         /*
1646          * If already mapped, we don't have to account.
1647          * If page cache, page->mapping has address_space.
1648          * But page->mapping may have out-of-use anon_vma pointer,
1649          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1650          * is NULL.
1651          */
1652         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1653                 return 0;
1654         if (unlikely(!mm))
1655                 mm = &init_mm;
1656         return mem_cgroup_charge_common(page, mm, gfp_mask,
1657                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1658 }
1659
1660 static void
1661 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1662                                         enum charge_type ctype);
1663
1664 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1665                                 gfp_t gfp_mask)
1666 {
1667         struct mem_cgroup *mem = NULL;
1668         int ret;
1669
1670         if (mem_cgroup_disabled())
1671                 return 0;
1672         if (PageCompound(page))
1673                 return 0;
1674         /*
1675          * Corner case handling. This is called from add_to_page_cache()
1676          * in usual. But some FS (shmem) precharges this page before calling it
1677          * and call add_to_page_cache() with GFP_NOWAIT.
1678          *
1679          * For GFP_NOWAIT case, the page may be pre-charged before calling
1680          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1681          * charge twice. (It works but has to pay a bit larger cost.)
1682          * And when the page is SwapCache, it should take swap information
1683          * into account. This is under lock_page() now.
1684          */
1685         if (!(gfp_mask & __GFP_WAIT)) {
1686                 struct page_cgroup *pc;
1687
1688
1689                 pc = lookup_page_cgroup(page);
1690                 if (!pc)
1691                         return 0;
1692                 lock_page_cgroup(pc);
1693                 if (PageCgroupUsed(pc)) {
1694                         unlock_page_cgroup(pc);
1695                         return 0;
1696                 }
1697                 unlock_page_cgroup(pc);
1698         }
1699
1700         if (unlikely(!mm && !mem))
1701                 mm = &init_mm;
1702
1703         if (page_is_file_cache(page))
1704                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1705                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1706
1707         /* shmem */
1708         if (PageSwapCache(page)) {
1709                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1710                 if (!ret)
1711                         __mem_cgroup_commit_charge_swapin(page, mem,
1712                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1713         } else
1714                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1715                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1716
1717         return ret;
1718 }
1719
1720 /*
1721  * While swap-in, try_charge -> commit or cancel, the page is locked.
1722  * And when try_charge() successfully returns, one refcnt to memcg without
1723  * struct page_cgroup is acquired. This refcnt will be consumed by
1724  * "commit()" or removed by "cancel()"
1725  */
1726 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1727                                  struct page *page,
1728                                  gfp_t mask, struct mem_cgroup **ptr)
1729 {
1730         struct mem_cgroup *mem;
1731         int ret;
1732
1733         if (mem_cgroup_disabled())
1734                 return 0;
1735
1736         if (!do_swap_account)
1737                 goto charge_cur_mm;
1738         /*
1739          * A racing thread's fault, or swapoff, may have already updated
1740          * the pte, and even removed page from swap cache: in those cases
1741          * do_swap_page()'s pte_same() test will fail; but there's also a
1742          * KSM case which does need to charge the page.
1743          */
1744         if (!PageSwapCache(page))
1745                 goto charge_cur_mm;
1746         mem = try_get_mem_cgroup_from_swapcache(page);
1747         if (!mem)
1748                 goto charge_cur_mm;
1749         *ptr = mem;
1750         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page);
1751         /* drop extra refcnt from tryget */
1752         css_put(&mem->css);
1753         return ret;
1754 charge_cur_mm:
1755         if (unlikely(!mm))
1756                 mm = &init_mm;
1757         return __mem_cgroup_try_charge(mm, mask, ptr, true, page);
1758 }
1759
1760 static void
1761 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1762                                         enum charge_type ctype)
1763 {
1764         struct page_cgroup *pc;
1765
1766         if (mem_cgroup_disabled())
1767                 return;
1768         if (!ptr)
1769                 return;
1770         cgroup_exclude_rmdir(&ptr->css);
1771         pc = lookup_page_cgroup(page);
1772         mem_cgroup_lru_del_before_commit_swapcache(page);
1773         __mem_cgroup_commit_charge(ptr, pc, ctype);
1774         mem_cgroup_lru_add_after_commit_swapcache(page);
1775         /*
1776          * Now swap is on-memory. This means this page may be
1777          * counted both as mem and swap....double count.
1778          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1779          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1780          * may call delete_from_swap_cache() before reach here.
1781          */
1782         if (do_swap_account && PageSwapCache(page)) {
1783                 swp_entry_t ent = {.val = page_private(page)};
1784                 unsigned short id;
1785                 struct mem_cgroup *memcg;
1786
1787                 id = swap_cgroup_record(ent, 0);
1788                 rcu_read_lock();
1789                 memcg = mem_cgroup_lookup(id);
1790                 if (memcg) {
1791                         /*
1792                          * This recorded memcg can be obsolete one. So, avoid
1793                          * calling css_tryget
1794                          */
1795                         if (!mem_cgroup_is_root(memcg))
1796                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1797                         mem_cgroup_swap_statistics(memcg, false);
1798                         mem_cgroup_put(memcg);
1799                 }
1800                 rcu_read_unlock();
1801         }
1802         /*
1803          * At swapin, we may charge account against cgroup which has no tasks.
1804          * So, rmdir()->pre_destroy() can be called while we do this charge.
1805          * In that case, we need to call pre_destroy() again. check it here.
1806          */
1807         cgroup_release_and_wakeup_rmdir(&ptr->css);
1808 }
1809
1810 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1811 {
1812         __mem_cgroup_commit_charge_swapin(page, ptr,
1813                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
1814 }
1815
1816 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1817 {
1818         if (mem_cgroup_disabled())
1819                 return;
1820         if (!mem)
1821                 return;
1822         if (!mem_cgroup_is_root(mem)) {
1823                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1824                 if (do_swap_account)
1825                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1826         }
1827         css_put(&mem->css);
1828 }
1829
1830
1831 /*
1832  * uncharge if !page_mapped(page)
1833  */
1834 static struct mem_cgroup *
1835 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1836 {
1837         struct page_cgroup *pc;
1838         struct mem_cgroup *mem = NULL;
1839         struct mem_cgroup_per_zone *mz;
1840
1841         if (mem_cgroup_disabled())
1842                 return NULL;
1843
1844         if (PageSwapCache(page))
1845                 return NULL;
1846
1847         /*
1848          * Check if our page_cgroup is valid
1849          */
1850         pc = lookup_page_cgroup(page);
1851         if (unlikely(!pc || !PageCgroupUsed(pc)))
1852                 return NULL;
1853
1854         lock_page_cgroup(pc);
1855
1856         mem = pc->mem_cgroup;
1857
1858         if (!PageCgroupUsed(pc))
1859                 goto unlock_out;
1860
1861         switch (ctype) {
1862         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1863         case MEM_CGROUP_CHARGE_TYPE_DROP:
1864                 if (page_mapped(page))
1865                         goto unlock_out;
1866                 break;
1867         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1868                 if (!PageAnon(page)) {  /* Shared memory */
1869                         if (page->mapping && !page_is_file_cache(page))
1870                                 goto unlock_out;
1871                 } else if (page_mapped(page)) /* Anon */
1872                                 goto unlock_out;
1873                 break;
1874         default:
1875                 break;
1876         }
1877
1878         if (!mem_cgroup_is_root(mem)) {
1879                 res_counter_uncharge(&mem->res, PAGE_SIZE);
1880                 if (do_swap_account &&
1881                                 (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1882                         res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1883         }
1884         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1885                 mem_cgroup_swap_statistics(mem, true);
1886         mem_cgroup_charge_statistics(mem, pc, false);
1887
1888         ClearPageCgroupUsed(pc);
1889         /*
1890          * pc->mem_cgroup is not cleared here. It will be accessed when it's
1891          * freed from LRU. This is safe because uncharged page is expected not
1892          * to be reused (freed soon). Exception is SwapCache, it's handled by
1893          * special functions.
1894          */
1895
1896         mz = page_cgroup_zoneinfo(pc);
1897         unlock_page_cgroup(pc);
1898
1899         if (mem_cgroup_soft_limit_check(mem))
1900                 mem_cgroup_update_tree(mem, page);
1901         /* at swapout, this memcg will be accessed to record to swap */
1902         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1903                 css_put(&mem->css);
1904
1905         return mem;
1906
1907 unlock_out:
1908         unlock_page_cgroup(pc);
1909         return NULL;
1910 }
1911
1912 void mem_cgroup_uncharge_page(struct page *page)
1913 {
1914         /* early check. */
1915         if (page_mapped(page))
1916                 return;
1917         if (page->mapping && !PageAnon(page))
1918                 return;
1919         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1920 }
1921
1922 void mem_cgroup_uncharge_cache_page(struct page *page)
1923 {
1924         VM_BUG_ON(page_mapped(page));
1925         VM_BUG_ON(page->mapping);
1926         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1927 }
1928
1929 #ifdef CONFIG_SWAP
1930 /*
1931  * called after __delete_from_swap_cache() and drop "page" account.
1932  * memcg information is recorded to swap_cgroup of "ent"
1933  */
1934 void
1935 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
1936 {
1937         struct mem_cgroup *memcg;
1938         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
1939
1940         if (!swapout) /* this was a swap cache but the swap is unused ! */
1941                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
1942
1943         memcg = __mem_cgroup_uncharge_common(page, ctype);
1944
1945         /* record memcg information */
1946         if (do_swap_account && swapout && memcg) {
1947                 swap_cgroup_record(ent, css_id(&memcg->css));
1948                 mem_cgroup_get(memcg);
1949         }
1950         if (swapout && memcg)
1951                 css_put(&memcg->css);
1952 }
1953 #endif
1954
1955 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1956 /*
1957  * called from swap_entry_free(). remove record in swap_cgroup and
1958  * uncharge "memsw" account.
1959  */
1960 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1961 {
1962         struct mem_cgroup *memcg;
1963         unsigned short id;
1964
1965         if (!do_swap_account)
1966                 return;
1967
1968         id = swap_cgroup_record(ent, 0);
1969         rcu_read_lock();
1970         memcg = mem_cgroup_lookup(id);
1971         if (memcg) {
1972                 /*
1973                  * We uncharge this because swap is freed.
1974                  * This memcg can be obsolete one. We avoid calling css_tryget
1975                  */
1976                 if (!mem_cgroup_is_root(memcg))
1977                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1978                 mem_cgroup_swap_statistics(memcg, false);
1979                 mem_cgroup_put(memcg);
1980         }
1981         rcu_read_unlock();
1982 }
1983 #endif
1984
1985 /*
1986  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1987  * page belongs to.
1988  */
1989 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1990 {
1991         struct page_cgroup *pc;
1992         struct mem_cgroup *mem = NULL;
1993         int ret = 0;
1994
1995         if (mem_cgroup_disabled())
1996                 return 0;
1997
1998         pc = lookup_page_cgroup(page);
1999         lock_page_cgroup(pc);
2000         if (PageCgroupUsed(pc)) {
2001                 mem = pc->mem_cgroup;
2002                 css_get(&mem->css);
2003         }
2004         unlock_page_cgroup(pc);
2005
2006         if (mem) {
2007                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
2008                                                 page);
2009                 css_put(&mem->css);
2010         }
2011         *ptr = mem;
2012         return ret;
2013 }
2014
2015 /* remove redundant charge if migration failed*/
2016 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2017                 struct page *oldpage, struct page *newpage)
2018 {
2019         struct page *target, *unused;
2020         struct page_cgroup *pc;
2021         enum charge_type ctype;
2022
2023         if (!mem)
2024                 return;
2025         cgroup_exclude_rmdir(&mem->css);
2026         /* at migration success, oldpage->mapping is NULL. */
2027         if (oldpage->mapping) {
2028                 target = oldpage;
2029                 unused = NULL;
2030         } else {
2031                 target = newpage;
2032                 unused = oldpage;
2033         }
2034
2035         if (PageAnon(target))
2036                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2037         else if (page_is_file_cache(target))
2038                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2039         else
2040                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2041
2042         /* unused page is not on radix-tree now. */
2043         if (unused)
2044                 __mem_cgroup_uncharge_common(unused, ctype);
2045
2046         pc = lookup_page_cgroup(target);
2047         /*
2048          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2049          * So, double-counting is effectively avoided.
2050          */
2051         __mem_cgroup_commit_charge(mem, pc, ctype);
2052
2053         /*
2054          * Both of oldpage and newpage are still under lock_page().
2055          * Then, we don't have to care about race in radix-tree.
2056          * But we have to be careful that this page is unmapped or not.
2057          *
2058          * There is a case for !page_mapped(). At the start of
2059          * migration, oldpage was mapped. But now, it's zapped.
2060          * But we know *target* page is not freed/reused under us.
2061          * mem_cgroup_uncharge_page() does all necessary checks.
2062          */
2063         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2064                 mem_cgroup_uncharge_page(target);
2065         /*
2066          * At migration, we may charge account against cgroup which has no tasks
2067          * So, rmdir()->pre_destroy() can be called while we do this charge.
2068          * In that case, we need to call pre_destroy() again. check it here.
2069          */
2070         cgroup_release_and_wakeup_rmdir(&mem->css);
2071 }
2072
2073 /*
2074  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2075  * Calling hierarchical_reclaim is not enough because we should update
2076  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2077  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2078  * not from the memcg which this page would be charged to.
2079  * try_charge_swapin does all of these works properly.
2080  */
2081 int mem_cgroup_shmem_charge_fallback(struct page *page,
2082                             struct mm_struct *mm,
2083                             gfp_t gfp_mask)
2084 {
2085         struct mem_cgroup *mem = NULL;
2086         int ret;
2087
2088         if (mem_cgroup_disabled())
2089                 return 0;
2090
2091         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2092         if (!ret)
2093                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2094
2095         return ret;
2096 }
2097
2098 static DEFINE_MUTEX(set_limit_mutex);
2099
2100 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2101                                 unsigned long long val)
2102 {
2103         int retry_count;
2104         int progress;
2105         u64 memswlimit;
2106         int ret = 0;
2107         int children = mem_cgroup_count_children(memcg);
2108         u64 curusage, oldusage;
2109
2110         /*
2111          * For keeping hierarchical_reclaim simple, how long we should retry
2112          * is depends on callers. We set our retry-count to be function
2113          * of # of children which we should visit in this loop.
2114          */
2115         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2116
2117         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2118
2119         while (retry_count) {
2120                 if (signal_pending(current)) {
2121                         ret = -EINTR;
2122                         break;
2123                 }
2124                 /*
2125                  * Rather than hide all in some function, I do this in
2126                  * open coded manner. You see what this really does.
2127                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2128                  */
2129                 mutex_lock(&set_limit_mutex);
2130                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2131                 if (memswlimit < val) {
2132                         ret = -EINVAL;
2133                         mutex_unlock(&set_limit_mutex);
2134                         break;
2135                 }
2136                 ret = res_counter_set_limit(&memcg->res, val);
2137                 if (!ret) {
2138                         if (memswlimit == val)
2139                                 memcg->memsw_is_minimum = true;
2140                         else
2141                                 memcg->memsw_is_minimum = false;
2142                 }
2143                 mutex_unlock(&set_limit_mutex);
2144
2145                 if (!ret)
2146                         break;
2147
2148                 progress = mem_cgroup_hierarchical_reclaim(memcg, NULL,
2149                                                 GFP_KERNEL,
2150                                                 MEM_CGROUP_RECLAIM_SHRINK);
2151                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2152                 /* Usage is reduced ? */
2153                 if (curusage >= oldusage)
2154                         retry_count--;
2155                 else
2156                         oldusage = curusage;
2157         }
2158
2159         return ret;
2160 }
2161
2162 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2163                                         unsigned long long val)
2164 {
2165         int retry_count;
2166         u64 memlimit, oldusage, curusage;
2167         int children = mem_cgroup_count_children(memcg);
2168         int ret = -EBUSY;
2169
2170         /* see mem_cgroup_resize_res_limit */
2171         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2172         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2173         while (retry_count) {
2174                 if (signal_pending(current)) {
2175                         ret = -EINTR;
2176                         break;
2177                 }
2178                 /*
2179                  * Rather than hide all in some function, I do this in
2180                  * open coded manner. You see what this really does.
2181                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2182                  */
2183                 mutex_lock(&set_limit_mutex);
2184                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2185                 if (memlimit > val) {
2186                         ret = -EINVAL;
2187                         mutex_unlock(&set_limit_mutex);
2188                         break;
2189                 }
2190                 ret = res_counter_set_limit(&memcg->memsw, val);
2191                 if (!ret) {
2192                         if (memlimit == val)
2193                                 memcg->memsw_is_minimum = true;
2194                         else
2195                                 memcg->memsw_is_minimum = false;
2196                 }
2197                 mutex_unlock(&set_limit_mutex);
2198
2199                 if (!ret)
2200                         break;
2201
2202                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2203                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2204                                                 MEM_CGROUP_RECLAIM_SHRINK);
2205                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2206                 /* Usage is reduced ? */
2207                 if (curusage >= oldusage)
2208                         retry_count--;
2209                 else
2210                         oldusage = curusage;
2211         }
2212         return ret;
2213 }
2214
2215 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2216                                                 gfp_t gfp_mask, int nid,
2217                                                 int zid)
2218 {
2219         unsigned long nr_reclaimed = 0;
2220         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2221         unsigned long reclaimed;
2222         int loop = 0;
2223         struct mem_cgroup_tree_per_zone *mctz;
2224         unsigned long long excess;
2225
2226         if (order > 0)
2227                 return 0;
2228
2229         mctz = soft_limit_tree_node_zone(nid, zid);
2230         /*
2231          * This loop can run a while, specially if mem_cgroup's continuously
2232          * keep exceeding their soft limit and putting the system under
2233          * pressure
2234          */
2235         do {
2236                 if (next_mz)
2237                         mz = next_mz;
2238                 else
2239                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2240                 if (!mz)
2241                         break;
2242
2243                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2244                                                 gfp_mask,
2245                                                 MEM_CGROUP_RECLAIM_SOFT);
2246                 nr_reclaimed += reclaimed;
2247                 spin_lock(&mctz->lock);
2248
2249                 /*
2250                  * If we failed to reclaim anything from this memory cgroup
2251                  * it is time to move on to the next cgroup
2252                  */
2253                 next_mz = NULL;
2254                 if (!reclaimed) {
2255                         do {
2256                                 /*
2257                                  * Loop until we find yet another one.
2258                                  *
2259                                  * By the time we get the soft_limit lock
2260                                  * again, someone might have aded the
2261                                  * group back on the RB tree. Iterate to
2262                                  * make sure we get a different mem.
2263                                  * mem_cgroup_largest_soft_limit_node returns
2264                                  * NULL if no other cgroup is present on
2265                                  * the tree
2266                                  */
2267                                 next_mz =
2268                                 __mem_cgroup_largest_soft_limit_node(mctz);
2269                                 if (next_mz == mz) {
2270                                         css_put(&next_mz->mem->css);
2271                                         next_mz = NULL;
2272                                 } else /* next_mz == NULL or other memcg */
2273                                         break;
2274                         } while (1);
2275                 }
2276                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2277                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2278                 /*
2279                  * One school of thought says that we should not add
2280                  * back the node to the tree if reclaim returns 0.
2281                  * But our reclaim could return 0, simply because due
2282                  * to priority we are exposing a smaller subset of
2283                  * memory to reclaim from. Consider this as a longer
2284                  * term TODO.
2285                  */
2286                 /* If excess == 0, no tree ops */
2287                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2288                 spin_unlock(&mctz->lock);
2289                 css_put(&mz->mem->css);
2290                 loop++;
2291                 /*
2292                  * Could not reclaim anything and there are no more
2293                  * mem cgroups to try or we seem to be looping without
2294                  * reclaiming anything.
2295                  */
2296                 if (!nr_reclaimed &&
2297                         (next_mz == NULL ||
2298                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2299                         break;
2300         } while (!nr_reclaimed);
2301         if (next_mz)
2302                 css_put(&next_mz->mem->css);
2303         return nr_reclaimed;
2304 }
2305
2306 /*
2307  * This routine traverse page_cgroup in given list and drop them all.
2308  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2309  */
2310 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2311                                 int node, int zid, enum lru_list lru)
2312 {
2313         struct zone *zone;
2314         struct mem_cgroup_per_zone *mz;
2315         struct page_cgroup *pc, *busy;
2316         unsigned long flags, loop;
2317         struct list_head *list;
2318         int ret = 0;
2319
2320         zone = &NODE_DATA(node)->node_zones[zid];
2321         mz = mem_cgroup_zoneinfo(mem, node, zid);
2322         list = &mz->lists[lru];
2323
2324         loop = MEM_CGROUP_ZSTAT(mz, lru);
2325         /* give some margin against EBUSY etc...*/
2326         loop += 256;
2327         busy = NULL;
2328         while (loop--) {
2329                 ret = 0;
2330                 spin_lock_irqsave(&zone->lru_lock, flags);
2331                 if (list_empty(list)) {
2332                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2333                         break;
2334                 }
2335                 pc = list_entry(list->prev, struct page_cgroup, lru);
2336                 if (busy == pc) {
2337                         list_move(&pc->lru, list);
2338                         busy = 0;
2339                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2340                         continue;
2341                 }
2342                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2343
2344                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2345                 if (ret == -ENOMEM)
2346                         break;
2347
2348                 if (ret == -EBUSY || ret == -EINVAL) {
2349                         /* found lock contention or "pc" is obsolete. */
2350                         busy = pc;
2351                         cond_resched();
2352                 } else
2353                         busy = NULL;
2354         }
2355
2356         if (!ret && !list_empty(list))
2357                 return -EBUSY;
2358         return ret;
2359 }
2360
2361 /*
2362  * make mem_cgroup's charge to be 0 if there is no task.
2363  * This enables deleting this mem_cgroup.
2364  */
2365 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2366 {
2367         int ret;
2368         int node, zid, shrink;
2369         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2370         struct cgroup *cgrp = mem->css.cgroup;
2371
2372         css_get(&mem->css);
2373
2374         shrink = 0;
2375         /* should free all ? */
2376         if (free_all)
2377                 goto try_to_free;
2378 move_account:
2379         while (mem->res.usage > 0) {
2380                 ret = -EBUSY;
2381                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2382                         goto out;
2383                 ret = -EINTR;
2384                 if (signal_pending(current))
2385                         goto out;
2386                 /* This is for making all *used* pages to be on LRU. */
2387                 lru_add_drain_all();
2388                 ret = 0;
2389                 for_each_node_state(node, N_HIGH_MEMORY) {
2390                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2391                                 enum lru_list l;
2392                                 for_each_lru(l) {
2393                                         ret = mem_cgroup_force_empty_list(mem,
2394                                                         node, zid, l);
2395                                         if (ret)
2396                                                 break;
2397                                 }
2398                         }
2399                         if (ret)
2400                                 break;
2401                 }
2402                 /* it seems parent cgroup doesn't have enough mem */
2403                 if (ret == -ENOMEM)
2404                         goto try_to_free;
2405                 cond_resched();
2406         }
2407         ret = 0;
2408 out:
2409         css_put(&mem->css);
2410         return ret;
2411
2412 try_to_free:
2413         /* returns EBUSY if there is a task or if we come here twice. */
2414         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2415                 ret = -EBUSY;
2416                 goto out;
2417         }
2418         /* we call try-to-free pages for make this cgroup empty */
2419         lru_add_drain_all();
2420         /* try to free all pages in this cgroup */
2421         shrink = 1;
2422         while (nr_retries && mem->res.usage > 0) {
2423                 int progress;
2424
2425                 if (signal_pending(current)) {
2426                         ret = -EINTR;
2427                         goto out;
2428                 }
2429                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2430                                                 false, get_swappiness(mem));
2431                 if (!progress) {
2432                         nr_retries--;
2433                         /* maybe some writeback is necessary */
2434                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2435                 }
2436
2437         }
2438         lru_add_drain();
2439         /* try move_account...there may be some *locked* pages. */
2440         if (mem->res.usage)
2441                 goto move_account;
2442         ret = 0;
2443         goto out;
2444 }
2445
2446 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2447 {
2448         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2449 }
2450
2451
2452 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2453 {
2454         return mem_cgroup_from_cont(cont)->use_hierarchy;
2455 }
2456
2457 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2458                                         u64 val)
2459 {
2460         int retval = 0;
2461         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2462         struct cgroup *parent = cont->parent;
2463         struct mem_cgroup *parent_mem = NULL;
2464
2465         if (parent)
2466                 parent_mem = mem_cgroup_from_cont(parent);
2467
2468         cgroup_lock();
2469         /*
2470          * If parent's use_hierarchy is set, we can't make any modifications
2471          * in the child subtrees. If it is unset, then the change can
2472          * occur, provided the current cgroup has no children.
2473          *
2474          * For the root cgroup, parent_mem is NULL, we allow value to be
2475          * set if there are no children.
2476          */
2477         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2478                                 (val == 1 || val == 0)) {
2479                 if (list_empty(&cont->children))
2480                         mem->use_hierarchy = val;
2481                 else
2482                         retval = -EBUSY;
2483         } else
2484                 retval = -EINVAL;
2485         cgroup_unlock();
2486
2487         return retval;
2488 }
2489
2490 struct mem_cgroup_idx_data {
2491         s64 val;
2492         enum mem_cgroup_stat_index idx;
2493 };
2494
2495 static int
2496 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2497 {
2498         struct mem_cgroup_idx_data *d = data;
2499         d->val += mem_cgroup_read_stat(&mem->stat, d->idx);
2500         return 0;
2501 }
2502
2503 static void
2504 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2505                                 enum mem_cgroup_stat_index idx, s64 *val)
2506 {
2507         struct mem_cgroup_idx_data d;
2508         d.idx = idx;
2509         d.val = 0;
2510         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2511         *val = d.val;
2512 }
2513
2514 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2515 {
2516         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2517         u64 idx_val, val;
2518         int type, name;
2519
2520         type = MEMFILE_TYPE(cft->private);
2521         name = MEMFILE_ATTR(cft->private);
2522         switch (type) {
2523         case _MEM:
2524                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2525                         mem_cgroup_get_recursive_idx_stat(mem,
2526                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2527                         val = idx_val;
2528                         mem_cgroup_get_recursive_idx_stat(mem,
2529                                 MEM_CGROUP_STAT_RSS, &idx_val);
2530                         val += idx_val;
2531                         val <<= PAGE_SHIFT;
2532                 } else
2533                         val = res_counter_read_u64(&mem->res, name);
2534                 break;
2535         case _MEMSWAP:
2536                 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2537                         mem_cgroup_get_recursive_idx_stat(mem,
2538                                 MEM_CGROUP_STAT_CACHE, &idx_val);
2539                         val = idx_val;
2540                         mem_cgroup_get_recursive_idx_stat(mem,
2541                                 MEM_CGROUP_STAT_RSS, &idx_val);
2542                         val += idx_val;
2543                         mem_cgroup_get_recursive_idx_stat(mem,
2544                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2545                         val += idx_val;
2546                         val <<= PAGE_SHIFT;
2547                 } else
2548                         val = res_counter_read_u64(&mem->memsw, name);
2549                 break;
2550         default:
2551                 BUG();
2552                 break;
2553         }
2554         return val;
2555 }
2556 /*
2557  * The user of this function is...
2558  * RES_LIMIT.
2559  */
2560 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2561                             const char *buffer)
2562 {
2563         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2564         int type, name;
2565         unsigned long long val;
2566         int ret;
2567
2568         type = MEMFILE_TYPE(cft->private);
2569         name = MEMFILE_ATTR(cft->private);
2570         switch (name) {
2571         case RES_LIMIT:
2572                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2573                         ret = -EINVAL;
2574                         break;
2575                 }
2576                 /* This function does all necessary parse...reuse it */
2577                 ret = res_counter_memparse_write_strategy(buffer, &val);
2578                 if (ret)
2579                         break;
2580                 if (type == _MEM)
2581                         ret = mem_cgroup_resize_limit(memcg, val);
2582                 else
2583                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2584                 break;
2585         case RES_SOFT_LIMIT:
2586                 ret = res_counter_memparse_write_strategy(buffer, &val);
2587                 if (ret)
2588                         break;
2589                 /*
2590                  * For memsw, soft limits are hard to implement in terms
2591                  * of semantics, for now, we support soft limits for
2592                  * control without swap
2593                  */
2594                 if (type == _MEM)
2595                         ret = res_counter_set_soft_limit(&memcg->res, val);
2596                 else
2597                         ret = -EINVAL;
2598                 break;
2599         default:
2600                 ret = -EINVAL; /* should be BUG() ? */
2601                 break;
2602         }
2603         return ret;
2604 }
2605
2606 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2607                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2608 {
2609         struct cgroup *cgroup;
2610         unsigned long long min_limit, min_memsw_limit, tmp;
2611
2612         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2613         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2614         cgroup = memcg->css.cgroup;
2615         if (!memcg->use_hierarchy)
2616                 goto out;
2617
2618         while (cgroup->parent) {
2619                 cgroup = cgroup->parent;
2620                 memcg = mem_cgroup_from_cont(cgroup);
2621                 if (!memcg->use_hierarchy)
2622                         break;
2623                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2624                 min_limit = min(min_limit, tmp);
2625                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2626                 min_memsw_limit = min(min_memsw_limit, tmp);
2627         }
2628 out:
2629         *mem_limit = min_limit;
2630         *memsw_limit = min_memsw_limit;
2631         return;
2632 }
2633
2634 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2635 {
2636         struct mem_cgroup *mem;
2637         int type, name;
2638
2639         mem = mem_cgroup_from_cont(cont);
2640         type = MEMFILE_TYPE(event);
2641         name = MEMFILE_ATTR(event);
2642         switch (name) {
2643         case RES_MAX_USAGE:
2644                 if (type == _MEM)
2645                         res_counter_reset_max(&mem->res);
2646                 else
2647                         res_counter_reset_max(&mem->memsw);
2648                 break;
2649         case RES_FAILCNT:
2650                 if (type == _MEM)
2651                         res_counter_reset_failcnt(&mem->res);
2652                 else
2653                         res_counter_reset_failcnt(&mem->memsw);
2654                 break;
2655         }
2656
2657         return 0;
2658 }
2659
2660
2661 /* For read statistics */
2662 enum {
2663         MCS_CACHE,
2664         MCS_RSS,
2665         MCS_MAPPED_FILE,
2666         MCS_PGPGIN,
2667         MCS_PGPGOUT,
2668         MCS_SWAP,
2669         MCS_INACTIVE_ANON,
2670         MCS_ACTIVE_ANON,
2671         MCS_INACTIVE_FILE,
2672         MCS_ACTIVE_FILE,
2673         MCS_UNEVICTABLE,
2674         NR_MCS_STAT,
2675 };
2676
2677 struct mcs_total_stat {
2678         s64 stat[NR_MCS_STAT];
2679 };
2680
2681 struct {
2682         char *local_name;
2683         char *total_name;
2684 } memcg_stat_strings[NR_MCS_STAT] = {
2685         {"cache", "total_cache"},
2686         {"rss", "total_rss"},
2687         {"mapped_file", "total_mapped_file"},
2688         {"pgpgin", "total_pgpgin"},
2689         {"pgpgout", "total_pgpgout"},
2690         {"swap", "total_swap"},
2691         {"inactive_anon", "total_inactive_anon"},
2692         {"active_anon", "total_active_anon"},
2693         {"inactive_file", "total_inactive_file"},
2694         {"active_file", "total_active_file"},
2695         {"unevictable", "total_unevictable"}
2696 };
2697
2698
2699 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2700 {
2701         struct mcs_total_stat *s = data;
2702         s64 val;
2703
2704         /* per cpu stat */
2705         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2706         s->stat[MCS_CACHE] += val * PAGE_SIZE;
2707         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2708         s->stat[MCS_RSS] += val * PAGE_SIZE;
2709         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE);
2710         s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE;
2711         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2712         s->stat[MCS_PGPGIN] += val;
2713         val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2714         s->stat[MCS_PGPGOUT] += val;
2715         if (do_swap_account) {
2716                 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT);
2717                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
2718         }
2719
2720         /* per zone stat */
2721         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2722         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2723         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2724         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2725         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2726         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2727         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2728         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2729         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2730         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2731         return 0;
2732 }
2733
2734 static void
2735 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2736 {
2737         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2738 }
2739
2740 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2741                                  struct cgroup_map_cb *cb)
2742 {
2743         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2744         struct mcs_total_stat mystat;
2745         int i;
2746
2747         memset(&mystat, 0, sizeof(mystat));
2748         mem_cgroup_get_local_stat(mem_cont, &mystat);
2749
2750         for (i = 0; i < NR_MCS_STAT; i++) {
2751                 if (i == MCS_SWAP && !do_swap_account)
2752                         continue;
2753                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2754         }
2755
2756         /* Hierarchical information */
2757         {
2758                 unsigned long long limit, memsw_limit;
2759                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2760                 cb->fill(cb, "hierarchical_memory_limit", limit);
2761                 if (do_swap_account)
2762                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2763         }
2764
2765         memset(&mystat, 0, sizeof(mystat));
2766         mem_cgroup_get_total_stat(mem_cont, &mystat);
2767         for (i = 0; i < NR_MCS_STAT; i++) {
2768                 if (i == MCS_SWAP && !do_swap_account)
2769                         continue;
2770                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2771         }
2772
2773 #ifdef CONFIG_DEBUG_VM
2774         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2775
2776         {
2777                 int nid, zid;
2778                 struct mem_cgroup_per_zone *mz;
2779                 unsigned long recent_rotated[2] = {0, 0};
2780                 unsigned long recent_scanned[2] = {0, 0};
2781
2782                 for_each_online_node(nid)
2783                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2784                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2785
2786                                 recent_rotated[0] +=
2787                                         mz->reclaim_stat.recent_rotated[0];
2788                                 recent_rotated[1] +=
2789                                         mz->reclaim_stat.recent_rotated[1];
2790                                 recent_scanned[0] +=
2791                                         mz->reclaim_stat.recent_scanned[0];
2792                                 recent_scanned[1] +=
2793                                         mz->reclaim_stat.recent_scanned[1];
2794                         }
2795                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2796                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2797                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2798                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2799         }
2800 #endif
2801
2802         return 0;
2803 }
2804
2805 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2806 {
2807         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2808
2809         return get_swappiness(memcg);
2810 }
2811
2812 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2813                                        u64 val)
2814 {
2815         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2816         struct mem_cgroup *parent;
2817
2818         if (val > 100)
2819                 return -EINVAL;
2820
2821         if (cgrp->parent == NULL)
2822                 return -EINVAL;
2823
2824         parent = mem_cgroup_from_cont(cgrp->parent);
2825
2826         cgroup_lock();
2827
2828         /* If under hierarchy, only empty-root can set this value */
2829         if ((parent->use_hierarchy) ||
2830             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2831                 cgroup_unlock();
2832                 return -EINVAL;
2833         }
2834
2835         spin_lock(&memcg->reclaim_param_lock);
2836         memcg->swappiness = val;
2837         spin_unlock(&memcg->reclaim_param_lock);
2838
2839         cgroup_unlock();
2840
2841         return 0;
2842 }
2843
2844
2845 static struct cftype mem_cgroup_files[] = {
2846         {
2847                 .name = "usage_in_bytes",
2848                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2849                 .read_u64 = mem_cgroup_read,
2850         },
2851         {
2852                 .name = "max_usage_in_bytes",
2853                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2854                 .trigger = mem_cgroup_reset,
2855                 .read_u64 = mem_cgroup_read,
2856         },
2857         {
2858                 .name = "limit_in_bytes",
2859                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2860                 .write_string = mem_cgroup_write,
2861                 .read_u64 = mem_cgroup_read,
2862         },
2863         {
2864                 .name = "soft_limit_in_bytes",
2865                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
2866                 .write_string = mem_cgroup_write,
2867                 .read_u64 = mem_cgroup_read,
2868         },
2869         {
2870                 .name = "failcnt",
2871                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2872                 .trigger = mem_cgroup_reset,
2873                 .read_u64 = mem_cgroup_read,
2874         },
2875         {
2876                 .name = "stat",
2877                 .read_map = mem_control_stat_show,
2878         },
2879         {
2880                 .name = "force_empty",
2881                 .trigger = mem_cgroup_force_empty_write,
2882         },
2883         {
2884                 .name = "use_hierarchy",
2885                 .write_u64 = mem_cgroup_hierarchy_write,
2886                 .read_u64 = mem_cgroup_hierarchy_read,
2887         },
2888         {
2889                 .name = "swappiness",
2890                 .read_u64 = mem_cgroup_swappiness_read,
2891                 .write_u64 = mem_cgroup_swappiness_write,
2892         },
2893 };
2894
2895 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2896 static struct cftype memsw_cgroup_files[] = {
2897         {
2898                 .name = "memsw.usage_in_bytes",
2899                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2900                 .read_u64 = mem_cgroup_read,
2901         },
2902         {
2903                 .name = "memsw.max_usage_in_bytes",
2904                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2905                 .trigger = mem_cgroup_reset,
2906                 .read_u64 = mem_cgroup_read,
2907         },
2908         {
2909                 .name = "memsw.limit_in_bytes",
2910                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2911                 .write_string = mem_cgroup_write,
2912                 .read_u64 = mem_cgroup_read,
2913         },
2914         {
2915                 .name = "memsw.failcnt",
2916                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2917                 .trigger = mem_cgroup_reset,
2918                 .read_u64 = mem_cgroup_read,
2919         },
2920 };
2921
2922 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2923 {
2924         if (!do_swap_account)
2925                 return 0;
2926         return cgroup_add_files(cont, ss, memsw_cgroup_files,
2927                                 ARRAY_SIZE(memsw_cgroup_files));
2928 };
2929 #else
2930 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2931 {
2932         return 0;
2933 }
2934 #endif
2935
2936 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2937 {
2938         struct mem_cgroup_per_node *pn;
2939         struct mem_cgroup_per_zone *mz;
2940         enum lru_list l;
2941         int zone, tmp = node;
2942         /*
2943          * This routine is called against possible nodes.
2944          * But it's BUG to call kmalloc() against offline node.
2945          *
2946          * TODO: this routine can waste much memory for nodes which will
2947          *       never be onlined. It's better to use memory hotplug callback
2948          *       function.
2949          */
2950         if (!node_state(node, N_NORMAL_MEMORY))
2951                 tmp = -1;
2952         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2953         if (!pn)
2954                 return 1;
2955
2956         mem->info.nodeinfo[node] = pn;
2957         memset(pn, 0, sizeof(*pn));
2958
2959         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2960                 mz = &pn->zoneinfo[zone];
2961                 for_each_lru(l)
2962                         INIT_LIST_HEAD(&mz->lists[l]);
2963                 mz->usage_in_excess = 0;
2964                 mz->on_tree = false;
2965                 mz->mem = mem;
2966         }
2967         return 0;
2968 }
2969
2970 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2971 {
2972         kfree(mem->info.nodeinfo[node]);
2973 }
2974
2975 static int mem_cgroup_size(void)
2976 {
2977         int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2978         return sizeof(struct mem_cgroup) + cpustat_size;
2979 }
2980
2981 static struct mem_cgroup *mem_cgroup_alloc(void)
2982 {
2983         struct mem_cgroup *mem;
2984         int size = mem_cgroup_size();
2985
2986         if (size < PAGE_SIZE)
2987                 mem = kmalloc(size, GFP_KERNEL);
2988         else
2989                 mem = vmalloc(size);
2990
2991         if (mem)
2992                 memset(mem, 0, size);
2993         return mem;
2994 }
2995
2996 /*
2997  * At destroying mem_cgroup, references from swap_cgroup can remain.
2998  * (scanning all at force_empty is too costly...)
2999  *
3000  * Instead of clearing all references at force_empty, we remember
3001  * the number of reference from swap_cgroup and free mem_cgroup when
3002  * it goes down to 0.
3003  *
3004  * Removal of cgroup itself succeeds regardless of refs from swap.
3005  */
3006
3007 static void __mem_cgroup_free(struct mem_cgroup *mem)
3008 {
3009         int node;
3010
3011         mem_cgroup_remove_from_trees(mem);
3012         free_css_id(&mem_cgroup_subsys, &mem->css);
3013
3014         for_each_node_state(node, N_POSSIBLE)
3015                 free_mem_cgroup_per_zone_info(mem, node);
3016
3017         if (mem_cgroup_size() < PAGE_SIZE)
3018                 kfree(mem);
3019         else
3020                 vfree(mem);
3021 }
3022
3023 static void mem_cgroup_get(struct mem_cgroup *mem)
3024 {
3025         atomic_inc(&mem->refcnt);
3026 }
3027
3028 static void mem_cgroup_put(struct mem_cgroup *mem)
3029 {
3030         if (atomic_dec_and_test(&mem->refcnt)) {
3031                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3032                 __mem_cgroup_free(mem);
3033                 if (parent)
3034                         mem_cgroup_put(parent);
3035         }
3036 }
3037
3038 /*
3039  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3040  */
3041 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3042 {
3043         if (!mem->res.parent)
3044                 return NULL;
3045         return mem_cgroup_from_res_counter(mem->res.parent, res);
3046 }
3047
3048 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3049 static void __init enable_swap_cgroup(void)
3050 {
3051         if (!mem_cgroup_disabled() && really_do_swap_account)
3052                 do_swap_account = 1;
3053 }
3054 #else
3055 static void __init enable_swap_cgroup(void)
3056 {
3057 }
3058 #endif
3059
3060 static int mem_cgroup_soft_limit_tree_init(void)
3061 {
3062         struct mem_cgroup_tree_per_node *rtpn;
3063         struct mem_cgroup_tree_per_zone *rtpz;
3064         int tmp, node, zone;
3065
3066         for_each_node_state(node, N_POSSIBLE) {
3067                 tmp = node;
3068                 if (!node_state(node, N_NORMAL_MEMORY))
3069                         tmp = -1;
3070                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3071                 if (!rtpn)
3072                         return 1;
3073
3074                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3075
3076                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3077                         rtpz = &rtpn->rb_tree_per_zone[zone];
3078                         rtpz->rb_root = RB_ROOT;
3079                         spin_lock_init(&rtpz->lock);
3080                 }
3081         }
3082         return 0;
3083 }
3084
3085 static struct cgroup_subsys_state * __ref
3086 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3087 {
3088         struct mem_cgroup *mem, *parent;
3089         long error = -ENOMEM;
3090         int node;
3091
3092         mem = mem_cgroup_alloc();
3093         if (!mem)
3094                 return ERR_PTR(error);
3095
3096         for_each_node_state(node, N_POSSIBLE)
3097                 if (alloc_mem_cgroup_per_zone_info(mem, node))
3098                         goto free_out;
3099
3100         /* root ? */
3101         if (cont->parent == NULL) {
3102                 enable_swap_cgroup();
3103                 parent = NULL;
3104                 root_mem_cgroup = mem;
3105                 if (mem_cgroup_soft_limit_tree_init())
3106                         goto free_out;
3107
3108         } else {
3109                 parent = mem_cgroup_from_cont(cont->parent);
3110                 mem->use_hierarchy = parent->use_hierarchy;
3111         }
3112
3113         if (parent && parent->use_hierarchy) {
3114                 res_counter_init(&mem->res, &parent->res);
3115                 res_counter_init(&mem->memsw, &parent->memsw);
3116                 /*
3117                  * We increment refcnt of the parent to ensure that we can
3118                  * safely access it on res_counter_charge/uncharge.
3119                  * This refcnt will be decremented when freeing this
3120                  * mem_cgroup(see mem_cgroup_put).
3121                  */
3122                 mem_cgroup_get(parent);
3123         } else {
3124                 res_counter_init(&mem->res, NULL);
3125                 res_counter_init(&mem->memsw, NULL);
3126         }
3127         mem->last_scanned_child = 0;
3128         spin_lock_init(&mem->reclaim_param_lock);
3129
3130         if (parent)
3131                 mem->swappiness = get_swappiness(parent);
3132         atomic_set(&mem->refcnt, 1);
3133         return &mem->css;
3134 free_out:
3135         __mem_cgroup_free(mem);
3136         root_mem_cgroup = NULL;
3137         return ERR_PTR(error);
3138 }
3139
3140 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
3141                                         struct cgroup *cont)
3142 {
3143         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3144
3145         return mem_cgroup_force_empty(mem, false);
3146 }
3147
3148 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
3149                                 struct cgroup *cont)
3150 {
3151         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3152
3153         mem_cgroup_put(mem);
3154 }
3155
3156 static int mem_cgroup_populate(struct cgroup_subsys *ss,
3157                                 struct cgroup *cont)
3158 {
3159         int ret;
3160
3161         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
3162                                 ARRAY_SIZE(mem_cgroup_files));
3163
3164         if (!ret)
3165                 ret = register_memsw_files(cont, ss);
3166         return ret;
3167 }
3168
3169 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
3170                                 struct cgroup *cont,
3171                                 struct cgroup *old_cont,
3172                                 struct task_struct *p,
3173                                 bool threadgroup)
3174 {
3175         mutex_lock(&memcg_tasklist);
3176         /*
3177          * FIXME: It's better to move charges of this process from old
3178          * memcg to new memcg. But it's just on TODO-List now.
3179          */
3180         mutex_unlock(&memcg_tasklist);
3181 }
3182
3183 struct cgroup_subsys mem_cgroup_subsys = {
3184         .name = "memory",
3185         .subsys_id = mem_cgroup_subsys_id,
3186         .create = mem_cgroup_create,
3187         .pre_destroy = mem_cgroup_pre_destroy,
3188         .destroy = mem_cgroup_destroy,
3189         .populate = mem_cgroup_populate,
3190         .attach = mem_cgroup_move_task,
3191         .early_init = 0,
3192         .use_id = 1,
3193 };
3194
3195 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3196
3197 static int __init disable_swap_account(char *s)
3198 {
3199         really_do_swap_account = 0;
3200         return 1;
3201 }
3202 __setup("noswapaccount", disable_swap_account);
3203 #endif