006fe142d4ba48488fe67eba588c2f2151dfdba0
[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  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2 of the License, or
16  * (at your option) any later version.
17  *
18  * This program is distributed in the hope that it will be useful,
19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
21  * GNU General Public License for more details.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
44 #include <linux/fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include "internal.h"
51
52 #include <asm/uaccess.h>
53
54 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
55 #define MEM_CGROUP_RECLAIM_RETRIES      5
56 struct mem_cgroup *root_mem_cgroup __read_mostly;
57
58 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
59 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
60 int do_swap_account __read_mostly;
61 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
62 #else
63 #define do_swap_account         (0)
64 #endif
65
66 #define SOFTLIMIT_EVENTS_THRESH (1000)
67 #define THRESHOLDS_EVENTS_THRESH (100)
68
69 /*
70  * Statistics for memory cgroup.
71  */
72 enum mem_cgroup_stat_index {
73         /*
74          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
75          */
76         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
77         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
78         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
79         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
80         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
81         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
82         MEM_CGROUP_STAT_SOFTLIMIT, /* decrements on each page in/out.
83                                         used by soft limit implementation */
84         MEM_CGROUP_STAT_THRESHOLDS, /* decrements on each page in/out.
85                                         used by threshold implementation */
86
87         MEM_CGROUP_STAT_NSTATS,
88 };
89
90 struct mem_cgroup_stat_cpu {
91         s64 count[MEM_CGROUP_STAT_NSTATS];
92 };
93
94 /*
95  * per-zone information in memory controller.
96  */
97 struct mem_cgroup_per_zone {
98         /*
99          * spin_lock to protect the per cgroup LRU
100          */
101         struct list_head        lists[NR_LRU_LISTS];
102         unsigned long           count[NR_LRU_LISTS];
103
104         struct zone_reclaim_stat reclaim_stat;
105         struct rb_node          tree_node;      /* RB tree node */
106         unsigned long long      usage_in_excess;/* Set to the value by which */
107                                                 /* the soft limit is exceeded*/
108         bool                    on_tree;
109         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
110                                                 /* use container_of        */
111 };
112 /* Macro for accessing counter */
113 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
114
115 struct mem_cgroup_per_node {
116         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
117 };
118
119 struct mem_cgroup_lru_info {
120         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
121 };
122
123 /*
124  * Cgroups above their limits are maintained in a RB-Tree, independent of
125  * their hierarchy representation
126  */
127
128 struct mem_cgroup_tree_per_zone {
129         struct rb_root rb_root;
130         spinlock_t lock;
131 };
132
133 struct mem_cgroup_tree_per_node {
134         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
135 };
136
137 struct mem_cgroup_tree {
138         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
139 };
140
141 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
142
143 struct mem_cgroup_threshold {
144         struct eventfd_ctx *eventfd;
145         u64 threshold;
146 };
147
148 struct mem_cgroup_threshold_ary {
149         /* An array index points to threshold just below usage. */
150         atomic_t current_threshold;
151         /* Size of entries[] */
152         unsigned int size;
153         /* Array of thresholds */
154         struct mem_cgroup_threshold entries[0];
155 };
156
157 static bool mem_cgroup_threshold_check(struct mem_cgroup *mem);
158 static void mem_cgroup_threshold(struct mem_cgroup *mem);
159
160 /*
161  * The memory controller data structure. The memory controller controls both
162  * page cache and RSS per cgroup. We would eventually like to provide
163  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
164  * to help the administrator determine what knobs to tune.
165  *
166  * TODO: Add a water mark for the memory controller. Reclaim will begin when
167  * we hit the water mark. May be even add a low water mark, such that
168  * no reclaim occurs from a cgroup at it's low water mark, this is
169  * a feature that will be implemented much later in the future.
170  */
171 struct mem_cgroup {
172         struct cgroup_subsys_state css;
173         /*
174          * the counter to account for memory usage
175          */
176         struct res_counter res;
177         /*
178          * the counter to account for mem+swap usage.
179          */
180         struct res_counter memsw;
181         /*
182          * Per cgroup active and inactive list, similar to the
183          * per zone LRU lists.
184          */
185         struct mem_cgroup_lru_info info;
186
187         /*
188           protect against reclaim related member.
189         */
190         spinlock_t reclaim_param_lock;
191
192         int     prev_priority;  /* for recording reclaim priority */
193
194         /*
195          * While reclaiming in a hierarchy, we cache the last child we
196          * reclaimed from.
197          */
198         int last_scanned_child;
199         /*
200          * Should the accounting and control be hierarchical, per subtree?
201          */
202         bool use_hierarchy;
203         unsigned long   last_oom_jiffies;
204         atomic_t        refcnt;
205
206         unsigned int    swappiness;
207
208         /* set when res.limit == memsw.limit */
209         bool            memsw_is_minimum;
210
211         /* protect arrays of thresholds */
212         struct mutex thresholds_lock;
213
214         /* thresholds for memory usage. RCU-protected */
215         struct mem_cgroup_threshold_ary *thresholds;
216
217         /* thresholds for mem+swap usage. RCU-protected */
218         struct mem_cgroup_threshold_ary *memsw_thresholds;
219
220         /*
221          * Should we move charges of a task when a task is moved into this
222          * mem_cgroup ? And what type of charges should we move ?
223          */
224         unsigned long   move_charge_at_immigrate;
225
226         /*
227          * percpu counter.
228          */
229         struct mem_cgroup_stat_cpu *stat;
230 };
231
232 /* Stuffs for move charges at task migration. */
233 /*
234  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
235  * left-shifted bitmap of these types.
236  */
237 enum move_type {
238         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
239         NR_MOVE_TYPE,
240 };
241
242 /* "mc" and its members are protected by cgroup_mutex */
243 static struct move_charge_struct {
244         struct mem_cgroup *from;
245         struct mem_cgroup *to;
246         unsigned long precharge;
247         unsigned long moved_charge;
248         unsigned long moved_swap;
249         struct task_struct *moving_task;        /* a task moving charges */
250         wait_queue_head_t waitq;                /* a waitq for other context */
251 } mc = {
252         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
253 };
254
255 /*
256  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
257  * limit reclaim to prevent infinite loops, if they ever occur.
258  */
259 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
260 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
261
262 enum charge_type {
263         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
264         MEM_CGROUP_CHARGE_TYPE_MAPPED,
265         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
266         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
267         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
268         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
269         NR_CHARGE_TYPE,
270 };
271
272 /* only for here (for easy reading.) */
273 #define PCGF_CACHE      (1UL << PCG_CACHE)
274 #define PCGF_USED       (1UL << PCG_USED)
275 #define PCGF_LOCK       (1UL << PCG_LOCK)
276 /* Not used, but added here for completeness */
277 #define PCGF_ACCT       (1UL << PCG_ACCT)
278
279 /* for encoding cft->private value on file */
280 #define _MEM                    (0)
281 #define _MEMSWAP                (1)
282 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
283 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
284 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
285
286 /*
287  * Reclaim flags for mem_cgroup_hierarchical_reclaim
288  */
289 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
290 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
291 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
292 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
293 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
294 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
295
296 static void mem_cgroup_get(struct mem_cgroup *mem);
297 static void mem_cgroup_put(struct mem_cgroup *mem);
298 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
299 static void drain_all_stock_async(void);
300
301 static struct mem_cgroup_per_zone *
302 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
303 {
304         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
305 }
306
307 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
308 {
309         return &mem->css;
310 }
311
312 static struct mem_cgroup_per_zone *
313 page_cgroup_zoneinfo(struct page_cgroup *pc)
314 {
315         struct mem_cgroup *mem = pc->mem_cgroup;
316         int nid = page_cgroup_nid(pc);
317         int zid = page_cgroup_zid(pc);
318
319         if (!mem)
320                 return NULL;
321
322         return mem_cgroup_zoneinfo(mem, nid, zid);
323 }
324
325 static struct mem_cgroup_tree_per_zone *
326 soft_limit_tree_node_zone(int nid, int zid)
327 {
328         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
329 }
330
331 static struct mem_cgroup_tree_per_zone *
332 soft_limit_tree_from_page(struct page *page)
333 {
334         int nid = page_to_nid(page);
335         int zid = page_zonenum(page);
336
337         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
338 }
339
340 static void
341 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
342                                 struct mem_cgroup_per_zone *mz,
343                                 struct mem_cgroup_tree_per_zone *mctz,
344                                 unsigned long long new_usage_in_excess)
345 {
346         struct rb_node **p = &mctz->rb_root.rb_node;
347         struct rb_node *parent = NULL;
348         struct mem_cgroup_per_zone *mz_node;
349
350         if (mz->on_tree)
351                 return;
352
353         mz->usage_in_excess = new_usage_in_excess;
354         if (!mz->usage_in_excess)
355                 return;
356         while (*p) {
357                 parent = *p;
358                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
359                                         tree_node);
360                 if (mz->usage_in_excess < mz_node->usage_in_excess)
361                         p = &(*p)->rb_left;
362                 /*
363                  * We can't avoid mem cgroups that are over their soft
364                  * limit by the same amount
365                  */
366                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
367                         p = &(*p)->rb_right;
368         }
369         rb_link_node(&mz->tree_node, parent, p);
370         rb_insert_color(&mz->tree_node, &mctz->rb_root);
371         mz->on_tree = true;
372 }
373
374 static void
375 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
376                                 struct mem_cgroup_per_zone *mz,
377                                 struct mem_cgroup_tree_per_zone *mctz)
378 {
379         if (!mz->on_tree)
380                 return;
381         rb_erase(&mz->tree_node, &mctz->rb_root);
382         mz->on_tree = false;
383 }
384
385 static void
386 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
387                                 struct mem_cgroup_per_zone *mz,
388                                 struct mem_cgroup_tree_per_zone *mctz)
389 {
390         spin_lock(&mctz->lock);
391         __mem_cgroup_remove_exceeded(mem, mz, mctz);
392         spin_unlock(&mctz->lock);
393 }
394
395 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
396 {
397         bool ret = false;
398         s64 val;
399
400         val = this_cpu_read(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT]);
401         if (unlikely(val < 0)) {
402                 this_cpu_write(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT],
403                                 SOFTLIMIT_EVENTS_THRESH);
404                 ret = true;
405         }
406         return ret;
407 }
408
409 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
410 {
411         unsigned long long excess;
412         struct mem_cgroup_per_zone *mz;
413         struct mem_cgroup_tree_per_zone *mctz;
414         int nid = page_to_nid(page);
415         int zid = page_zonenum(page);
416         mctz = soft_limit_tree_from_page(page);
417
418         /*
419          * Necessary to update all ancestors when hierarchy is used.
420          * because their event counter is not touched.
421          */
422         for (; mem; mem = parent_mem_cgroup(mem)) {
423                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
424                 excess = res_counter_soft_limit_excess(&mem->res);
425                 /*
426                  * We have to update the tree if mz is on RB-tree or
427                  * mem is over its softlimit.
428                  */
429                 if (excess || mz->on_tree) {
430                         spin_lock(&mctz->lock);
431                         /* if on-tree, remove it */
432                         if (mz->on_tree)
433                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
434                         /*
435                          * Insert again. mz->usage_in_excess will be updated.
436                          * If excess is 0, no tree ops.
437                          */
438                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
439                         spin_unlock(&mctz->lock);
440                 }
441         }
442 }
443
444 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
445 {
446         int node, zone;
447         struct mem_cgroup_per_zone *mz;
448         struct mem_cgroup_tree_per_zone *mctz;
449
450         for_each_node_state(node, N_POSSIBLE) {
451                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
452                         mz = mem_cgroup_zoneinfo(mem, node, zone);
453                         mctz = soft_limit_tree_node_zone(node, zone);
454                         mem_cgroup_remove_exceeded(mem, mz, mctz);
455                 }
456         }
457 }
458
459 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
460 {
461         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
462 }
463
464 static struct mem_cgroup_per_zone *
465 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
466 {
467         struct rb_node *rightmost = NULL;
468         struct mem_cgroup_per_zone *mz;
469
470 retry:
471         mz = NULL;
472         rightmost = rb_last(&mctz->rb_root);
473         if (!rightmost)
474                 goto done;              /* Nothing to reclaim from */
475
476         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
477         /*
478          * Remove the node now but someone else can add it back,
479          * we will to add it back at the end of reclaim to its correct
480          * position in the tree.
481          */
482         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
483         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
484                 !css_tryget(&mz->mem->css))
485                 goto retry;
486 done:
487         return mz;
488 }
489
490 static struct mem_cgroup_per_zone *
491 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
492 {
493         struct mem_cgroup_per_zone *mz;
494
495         spin_lock(&mctz->lock);
496         mz = __mem_cgroup_largest_soft_limit_node(mctz);
497         spin_unlock(&mctz->lock);
498         return mz;
499 }
500
501 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
502                 enum mem_cgroup_stat_index idx)
503 {
504         int cpu;
505         s64 val = 0;
506
507         for_each_possible_cpu(cpu)
508                 val += per_cpu(mem->stat->count[idx], cpu);
509         return val;
510 }
511
512 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
513 {
514         s64 ret;
515
516         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
517         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
518         return ret;
519 }
520
521 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
522                                          bool charge)
523 {
524         int val = (charge) ? 1 : -1;
525         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
526 }
527
528 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
529                                          struct page_cgroup *pc,
530                                          bool charge)
531 {
532         int val = (charge) ? 1 : -1;
533
534         preempt_disable();
535
536         if (PageCgroupCache(pc))
537                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
538         else
539                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
540
541         if (charge)
542                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
543         else
544                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
545         __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT]);
546         __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS]);
547
548         preempt_enable();
549 }
550
551 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
552                                         enum lru_list idx)
553 {
554         int nid, zid;
555         struct mem_cgroup_per_zone *mz;
556         u64 total = 0;
557
558         for_each_online_node(nid)
559                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
560                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
561                         total += MEM_CGROUP_ZSTAT(mz, idx);
562                 }
563         return total;
564 }
565
566 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
567 {
568         return container_of(cgroup_subsys_state(cont,
569                                 mem_cgroup_subsys_id), struct mem_cgroup,
570                                 css);
571 }
572
573 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
574 {
575         /*
576          * mm_update_next_owner() may clear mm->owner to NULL
577          * if it races with swapoff, page migration, etc.
578          * So this can be called with p == NULL.
579          */
580         if (unlikely(!p))
581                 return NULL;
582
583         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
584                                 struct mem_cgroup, css);
585 }
586
587 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
588 {
589         struct mem_cgroup *mem = NULL;
590
591         if (!mm)
592                 return NULL;
593         /*
594          * Because we have no locks, mm->owner's may be being moved to other
595          * cgroup. We use css_tryget() here even if this looks
596          * pessimistic (rather than adding locks here).
597          */
598         rcu_read_lock();
599         do {
600                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
601                 if (unlikely(!mem))
602                         break;
603         } while (!css_tryget(&mem->css));
604         rcu_read_unlock();
605         return mem;
606 }
607
608 /*
609  * Call callback function against all cgroup under hierarchy tree.
610  */
611 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
612                           int (*func)(struct mem_cgroup *, void *))
613 {
614         int found, ret, nextid;
615         struct cgroup_subsys_state *css;
616         struct mem_cgroup *mem;
617
618         if (!root->use_hierarchy)
619                 return (*func)(root, data);
620
621         nextid = 1;
622         do {
623                 ret = 0;
624                 mem = NULL;
625
626                 rcu_read_lock();
627                 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
628                                    &found);
629                 if (css && css_tryget(css))
630                         mem = container_of(css, struct mem_cgroup, css);
631                 rcu_read_unlock();
632
633                 if (mem) {
634                         ret = (*func)(mem, data);
635                         css_put(&mem->css);
636                 }
637                 nextid = found + 1;
638         } while (!ret && css);
639
640         return ret;
641 }
642
643 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
644 {
645         return (mem == root_mem_cgroup);
646 }
647
648 /*
649  * Following LRU functions are allowed to be used without PCG_LOCK.
650  * Operations are called by routine of global LRU independently from memcg.
651  * What we have to take care of here is validness of pc->mem_cgroup.
652  *
653  * Changes to pc->mem_cgroup happens when
654  * 1. charge
655  * 2. moving account
656  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
657  * It is added to LRU before charge.
658  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
659  * When moving account, the page is not on LRU. It's isolated.
660  */
661
662 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
663 {
664         struct page_cgroup *pc;
665         struct mem_cgroup_per_zone *mz;
666
667         if (mem_cgroup_disabled())
668                 return;
669         pc = lookup_page_cgroup(page);
670         /* can happen while we handle swapcache. */
671         if (!TestClearPageCgroupAcctLRU(pc))
672                 return;
673         VM_BUG_ON(!pc->mem_cgroup);
674         /*
675          * We don't check PCG_USED bit. It's cleared when the "page" is finally
676          * removed from global LRU.
677          */
678         mz = page_cgroup_zoneinfo(pc);
679         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
680         if (mem_cgroup_is_root(pc->mem_cgroup))
681                 return;
682         VM_BUG_ON(list_empty(&pc->lru));
683         list_del_init(&pc->lru);
684         return;
685 }
686
687 void mem_cgroup_del_lru(struct page *page)
688 {
689         mem_cgroup_del_lru_list(page, page_lru(page));
690 }
691
692 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
693 {
694         struct mem_cgroup_per_zone *mz;
695         struct page_cgroup *pc;
696
697         if (mem_cgroup_disabled())
698                 return;
699
700         pc = lookup_page_cgroup(page);
701         /*
702          * Used bit is set without atomic ops but after smp_wmb().
703          * For making pc->mem_cgroup visible, insert smp_rmb() here.
704          */
705         smp_rmb();
706         /* unused or root page is not rotated. */
707         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
708                 return;
709         mz = page_cgroup_zoneinfo(pc);
710         list_move(&pc->lru, &mz->lists[lru]);
711 }
712
713 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
714 {
715         struct page_cgroup *pc;
716         struct mem_cgroup_per_zone *mz;
717
718         if (mem_cgroup_disabled())
719                 return;
720         pc = lookup_page_cgroup(page);
721         VM_BUG_ON(PageCgroupAcctLRU(pc));
722         /*
723          * Used bit is set without atomic ops but after smp_wmb().
724          * For making pc->mem_cgroup visible, insert smp_rmb() here.
725          */
726         smp_rmb();
727         if (!PageCgroupUsed(pc))
728                 return;
729
730         mz = page_cgroup_zoneinfo(pc);
731         MEM_CGROUP_ZSTAT(mz, lru) += 1;
732         SetPageCgroupAcctLRU(pc);
733         if (mem_cgroup_is_root(pc->mem_cgroup))
734                 return;
735         list_add(&pc->lru, &mz->lists[lru]);
736 }
737
738 /*
739  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
740  * lru because the page may.be reused after it's fully uncharged (because of
741  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
742  * it again. This function is only used to charge SwapCache. It's done under
743  * lock_page and expected that zone->lru_lock is never held.
744  */
745 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
746 {
747         unsigned long flags;
748         struct zone *zone = page_zone(page);
749         struct page_cgroup *pc = lookup_page_cgroup(page);
750
751         spin_lock_irqsave(&zone->lru_lock, flags);
752         /*
753          * Forget old LRU when this page_cgroup is *not* used. This Used bit
754          * is guarded by lock_page() because the page is SwapCache.
755          */
756         if (!PageCgroupUsed(pc))
757                 mem_cgroup_del_lru_list(page, page_lru(page));
758         spin_unlock_irqrestore(&zone->lru_lock, flags);
759 }
760
761 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
762 {
763         unsigned long flags;
764         struct zone *zone = page_zone(page);
765         struct page_cgroup *pc = lookup_page_cgroup(page);
766
767         spin_lock_irqsave(&zone->lru_lock, flags);
768         /* link when the page is linked to LRU but page_cgroup isn't */
769         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
770                 mem_cgroup_add_lru_list(page, page_lru(page));
771         spin_unlock_irqrestore(&zone->lru_lock, flags);
772 }
773
774
775 void mem_cgroup_move_lists(struct page *page,
776                            enum lru_list from, enum lru_list to)
777 {
778         if (mem_cgroup_disabled())
779                 return;
780         mem_cgroup_del_lru_list(page, from);
781         mem_cgroup_add_lru_list(page, to);
782 }
783
784 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
785 {
786         int ret;
787         struct mem_cgroup *curr = NULL;
788
789         task_lock(task);
790         rcu_read_lock();
791         curr = try_get_mem_cgroup_from_mm(task->mm);
792         rcu_read_unlock();
793         task_unlock(task);
794         if (!curr)
795                 return 0;
796         /*
797          * We should check use_hierarchy of "mem" not "curr". Because checking
798          * use_hierarchy of "curr" here make this function true if hierarchy is
799          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
800          * hierarchy(even if use_hierarchy is disabled in "mem").
801          */
802         if (mem->use_hierarchy)
803                 ret = css_is_ancestor(&curr->css, &mem->css);
804         else
805                 ret = (curr == mem);
806         css_put(&curr->css);
807         return ret;
808 }
809
810 /*
811  * prev_priority control...this will be used in memory reclaim path.
812  */
813 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
814 {
815         int prev_priority;
816
817         spin_lock(&mem->reclaim_param_lock);
818         prev_priority = mem->prev_priority;
819         spin_unlock(&mem->reclaim_param_lock);
820
821         return prev_priority;
822 }
823
824 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
825 {
826         spin_lock(&mem->reclaim_param_lock);
827         if (priority < mem->prev_priority)
828                 mem->prev_priority = priority;
829         spin_unlock(&mem->reclaim_param_lock);
830 }
831
832 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
833 {
834         spin_lock(&mem->reclaim_param_lock);
835         mem->prev_priority = priority;
836         spin_unlock(&mem->reclaim_param_lock);
837 }
838
839 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
840 {
841         unsigned long active;
842         unsigned long inactive;
843         unsigned long gb;
844         unsigned long inactive_ratio;
845
846         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
847         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
848
849         gb = (inactive + active) >> (30 - PAGE_SHIFT);
850         if (gb)
851                 inactive_ratio = int_sqrt(10 * gb);
852         else
853                 inactive_ratio = 1;
854
855         if (present_pages) {
856                 present_pages[0] = inactive;
857                 present_pages[1] = active;
858         }
859
860         return inactive_ratio;
861 }
862
863 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
864 {
865         unsigned long active;
866         unsigned long inactive;
867         unsigned long present_pages[2];
868         unsigned long inactive_ratio;
869
870         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
871
872         inactive = present_pages[0];
873         active = present_pages[1];
874
875         if (inactive * inactive_ratio < active)
876                 return 1;
877
878         return 0;
879 }
880
881 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
882 {
883         unsigned long active;
884         unsigned long inactive;
885
886         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
887         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
888
889         return (active > inactive);
890 }
891
892 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
893                                        struct zone *zone,
894                                        enum lru_list lru)
895 {
896         int nid = zone->zone_pgdat->node_id;
897         int zid = zone_idx(zone);
898         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
899
900         return MEM_CGROUP_ZSTAT(mz, lru);
901 }
902
903 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
904                                                       struct zone *zone)
905 {
906         int nid = zone->zone_pgdat->node_id;
907         int zid = zone_idx(zone);
908         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
909
910         return &mz->reclaim_stat;
911 }
912
913 struct zone_reclaim_stat *
914 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
915 {
916         struct page_cgroup *pc;
917         struct mem_cgroup_per_zone *mz;
918
919         if (mem_cgroup_disabled())
920                 return NULL;
921
922         pc = lookup_page_cgroup(page);
923         /*
924          * Used bit is set without atomic ops but after smp_wmb().
925          * For making pc->mem_cgroup visible, insert smp_rmb() here.
926          */
927         smp_rmb();
928         if (!PageCgroupUsed(pc))
929                 return NULL;
930
931         mz = page_cgroup_zoneinfo(pc);
932         if (!mz)
933                 return NULL;
934
935         return &mz->reclaim_stat;
936 }
937
938 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
939                                         struct list_head *dst,
940                                         unsigned long *scanned, int order,
941                                         int mode, struct zone *z,
942                                         struct mem_cgroup *mem_cont,
943                                         int active, int file)
944 {
945         unsigned long nr_taken = 0;
946         struct page *page;
947         unsigned long scan;
948         LIST_HEAD(pc_list);
949         struct list_head *src;
950         struct page_cgroup *pc, *tmp;
951         int nid = z->zone_pgdat->node_id;
952         int zid = zone_idx(z);
953         struct mem_cgroup_per_zone *mz;
954         int lru = LRU_FILE * file + active;
955         int ret;
956
957         BUG_ON(!mem_cont);
958         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
959         src = &mz->lists[lru];
960
961         scan = 0;
962         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
963                 if (scan >= nr_to_scan)
964                         break;
965
966                 page = pc->page;
967                 if (unlikely(!PageCgroupUsed(pc)))
968                         continue;
969                 if (unlikely(!PageLRU(page)))
970                         continue;
971
972                 scan++;
973                 ret = __isolate_lru_page(page, mode, file);
974                 switch (ret) {
975                 case 0:
976                         list_move(&page->lru, dst);
977                         mem_cgroup_del_lru(page);
978                         nr_taken++;
979                         break;
980                 case -EBUSY:
981                         /* we don't affect global LRU but rotate in our LRU */
982                         mem_cgroup_rotate_lru_list(page, page_lru(page));
983                         break;
984                 default:
985                         break;
986                 }
987         }
988
989         *scanned = scan;
990         return nr_taken;
991 }
992
993 #define mem_cgroup_from_res_counter(counter, member)    \
994         container_of(counter, struct mem_cgroup, member)
995
996 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
997 {
998         if (do_swap_account) {
999                 if (res_counter_check_under_limit(&mem->res) &&
1000                         res_counter_check_under_limit(&mem->memsw))
1001                         return true;
1002         } else
1003                 if (res_counter_check_under_limit(&mem->res))
1004                         return true;
1005         return false;
1006 }
1007
1008 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1009 {
1010         struct cgroup *cgrp = memcg->css.cgroup;
1011         unsigned int swappiness;
1012
1013         /* root ? */
1014         if (cgrp->parent == NULL)
1015                 return vm_swappiness;
1016
1017         spin_lock(&memcg->reclaim_param_lock);
1018         swappiness = memcg->swappiness;
1019         spin_unlock(&memcg->reclaim_param_lock);
1020
1021         return swappiness;
1022 }
1023
1024 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1025 {
1026         int *val = data;
1027         (*val)++;
1028         return 0;
1029 }
1030
1031 /**
1032  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1033  * @memcg: The memory cgroup that went over limit
1034  * @p: Task that is going to be killed
1035  *
1036  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1037  * enabled
1038  */
1039 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1040 {
1041         struct cgroup *task_cgrp;
1042         struct cgroup *mem_cgrp;
1043         /*
1044          * Need a buffer in BSS, can't rely on allocations. The code relies
1045          * on the assumption that OOM is serialized for memory controller.
1046          * If this assumption is broken, revisit this code.
1047          */
1048         static char memcg_name[PATH_MAX];
1049         int ret;
1050
1051         if (!memcg || !p)
1052                 return;
1053
1054
1055         rcu_read_lock();
1056
1057         mem_cgrp = memcg->css.cgroup;
1058         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1059
1060         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1061         if (ret < 0) {
1062                 /*
1063                  * Unfortunately, we are unable to convert to a useful name
1064                  * But we'll still print out the usage information
1065                  */
1066                 rcu_read_unlock();
1067                 goto done;
1068         }
1069         rcu_read_unlock();
1070
1071         printk(KERN_INFO "Task in %s killed", memcg_name);
1072
1073         rcu_read_lock();
1074         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1075         if (ret < 0) {
1076                 rcu_read_unlock();
1077                 goto done;
1078         }
1079         rcu_read_unlock();
1080
1081         /*
1082          * Continues from above, so we don't need an KERN_ level
1083          */
1084         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1085 done:
1086
1087         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1088                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1089                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1090                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1091         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1092                 "failcnt %llu\n",
1093                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1094                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1095                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1096 }
1097
1098 /*
1099  * This function returns the number of memcg under hierarchy tree. Returns
1100  * 1(self count) if no children.
1101  */
1102 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1103 {
1104         int num = 0;
1105         mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1106         return num;
1107 }
1108
1109 /*
1110  * Visit the first child (need not be the first child as per the ordering
1111  * of the cgroup list, since we track last_scanned_child) of @mem and use
1112  * that to reclaim free pages from.
1113  */
1114 static struct mem_cgroup *
1115 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1116 {
1117         struct mem_cgroup *ret = NULL;
1118         struct cgroup_subsys_state *css;
1119         int nextid, found;
1120
1121         if (!root_mem->use_hierarchy) {
1122                 css_get(&root_mem->css);
1123                 ret = root_mem;
1124         }
1125
1126         while (!ret) {
1127                 rcu_read_lock();
1128                 nextid = root_mem->last_scanned_child + 1;
1129                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1130                                    &found);
1131                 if (css && css_tryget(css))
1132                         ret = container_of(css, struct mem_cgroup, css);
1133
1134                 rcu_read_unlock();
1135                 /* Updates scanning parameter */
1136                 spin_lock(&root_mem->reclaim_param_lock);
1137                 if (!css) {
1138                         /* this means start scan from ID:1 */
1139                         root_mem->last_scanned_child = 0;
1140                 } else
1141                         root_mem->last_scanned_child = found;
1142                 spin_unlock(&root_mem->reclaim_param_lock);
1143         }
1144
1145         return ret;
1146 }
1147
1148 /*
1149  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1150  * we reclaimed from, so that we don't end up penalizing one child extensively
1151  * based on its position in the children list.
1152  *
1153  * root_mem is the original ancestor that we've been reclaim from.
1154  *
1155  * We give up and return to the caller when we visit root_mem twice.
1156  * (other groups can be removed while we're walking....)
1157  *
1158  * If shrink==true, for avoiding to free too much, this returns immedieately.
1159  */
1160 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1161                                                 struct zone *zone,
1162                                                 gfp_t gfp_mask,
1163                                                 unsigned long reclaim_options)
1164 {
1165         struct mem_cgroup *victim;
1166         int ret, total = 0;
1167         int loop = 0;
1168         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1169         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1170         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1171         unsigned long excess = mem_cgroup_get_excess(root_mem);
1172
1173         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1174         if (root_mem->memsw_is_minimum)
1175                 noswap = true;
1176
1177         while (1) {
1178                 victim = mem_cgroup_select_victim(root_mem);
1179                 if (victim == root_mem) {
1180                         loop++;
1181                         if (loop >= 1)
1182                                 drain_all_stock_async();
1183                         if (loop >= 2) {
1184                                 /*
1185                                  * If we have not been able to reclaim
1186                                  * anything, it might because there are
1187                                  * no reclaimable pages under this hierarchy
1188                                  */
1189                                 if (!check_soft || !total) {
1190                                         css_put(&victim->css);
1191                                         break;
1192                                 }
1193                                 /*
1194                                  * We want to do more targetted reclaim.
1195                                  * excess >> 2 is not to excessive so as to
1196                                  * reclaim too much, nor too less that we keep
1197                                  * coming back to reclaim from this cgroup
1198                                  */
1199                                 if (total >= (excess >> 2) ||
1200                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1201                                         css_put(&victim->css);
1202                                         break;
1203                                 }
1204                         }
1205                 }
1206                 if (!mem_cgroup_local_usage(victim)) {
1207                         /* this cgroup's local usage == 0 */
1208                         css_put(&victim->css);
1209                         continue;
1210                 }
1211                 /* we use swappiness of local cgroup */
1212                 if (check_soft)
1213                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1214                                 noswap, get_swappiness(victim), zone,
1215                                 zone->zone_pgdat->node_id);
1216                 else
1217                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1218                                                 noswap, get_swappiness(victim));
1219                 css_put(&victim->css);
1220                 /*
1221                  * At shrinking usage, we can't check we should stop here or
1222                  * reclaim more. It's depends on callers. last_scanned_child
1223                  * will work enough for keeping fairness under tree.
1224                  */
1225                 if (shrink)
1226                         return ret;
1227                 total += ret;
1228                 if (check_soft) {
1229                         if (res_counter_check_under_soft_limit(&root_mem->res))
1230                                 return total;
1231                 } else if (mem_cgroup_check_under_limit(root_mem))
1232                         return 1 + total;
1233         }
1234         return total;
1235 }
1236
1237 bool mem_cgroup_oom_called(struct task_struct *task)
1238 {
1239         bool ret = false;
1240         struct mem_cgroup *mem;
1241         struct mm_struct *mm;
1242
1243         rcu_read_lock();
1244         mm = task->mm;
1245         if (!mm)
1246                 mm = &init_mm;
1247         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1248         if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1249                 ret = true;
1250         rcu_read_unlock();
1251         return ret;
1252 }
1253
1254 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
1255 {
1256         mem->last_oom_jiffies = jiffies;
1257         return 0;
1258 }
1259
1260 static void record_last_oom(struct mem_cgroup *mem)
1261 {
1262         mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
1263 }
1264
1265 /*
1266  * Currently used to update mapped file statistics, but the routine can be
1267  * generalized to update other statistics as well.
1268  */
1269 void mem_cgroup_update_file_mapped(struct page *page, int val)
1270 {
1271         struct mem_cgroup *mem;
1272         struct page_cgroup *pc;
1273
1274         pc = lookup_page_cgroup(page);
1275         if (unlikely(!pc))
1276                 return;
1277
1278         lock_page_cgroup(pc);
1279         mem = pc->mem_cgroup;
1280         if (!mem)
1281                 goto done;
1282
1283         if (!PageCgroupUsed(pc))
1284                 goto done;
1285
1286         /*
1287          * Preemption is already disabled. We can use __this_cpu_xxx
1288          */
1289         __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], val);
1290
1291 done:
1292         unlock_page_cgroup(pc);
1293 }
1294
1295 /*
1296  * size of first charge trial. "32" comes from vmscan.c's magic value.
1297  * TODO: maybe necessary to use big numbers in big irons.
1298  */
1299 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1300 struct memcg_stock_pcp {
1301         struct mem_cgroup *cached; /* this never be root cgroup */
1302         int charge;
1303         struct work_struct work;
1304 };
1305 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1306 static atomic_t memcg_drain_count;
1307
1308 /*
1309  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1310  * from local stock and true is returned. If the stock is 0 or charges from a
1311  * cgroup which is not current target, returns false. This stock will be
1312  * refilled.
1313  */
1314 static bool consume_stock(struct mem_cgroup *mem)
1315 {
1316         struct memcg_stock_pcp *stock;
1317         bool ret = true;
1318
1319         stock = &get_cpu_var(memcg_stock);
1320         if (mem == stock->cached && stock->charge)
1321                 stock->charge -= PAGE_SIZE;
1322         else /* need to call res_counter_charge */
1323                 ret = false;
1324         put_cpu_var(memcg_stock);
1325         return ret;
1326 }
1327
1328 /*
1329  * Returns stocks cached in percpu to res_counter and reset cached information.
1330  */
1331 static void drain_stock(struct memcg_stock_pcp *stock)
1332 {
1333         struct mem_cgroup *old = stock->cached;
1334
1335         if (stock->charge) {
1336                 res_counter_uncharge(&old->res, stock->charge);
1337                 if (do_swap_account)
1338                         res_counter_uncharge(&old->memsw, stock->charge);
1339         }
1340         stock->cached = NULL;
1341         stock->charge = 0;
1342 }
1343
1344 /*
1345  * This must be called under preempt disabled or must be called by
1346  * a thread which is pinned to local cpu.
1347  */
1348 static void drain_local_stock(struct work_struct *dummy)
1349 {
1350         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1351         drain_stock(stock);
1352 }
1353
1354 /*
1355  * Cache charges(val) which is from res_counter, to local per_cpu area.
1356  * This will be consumed by consumt_stock() function, later.
1357  */
1358 static void refill_stock(struct mem_cgroup *mem, int val)
1359 {
1360         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1361
1362         if (stock->cached != mem) { /* reset if necessary */
1363                 drain_stock(stock);
1364                 stock->cached = mem;
1365         }
1366         stock->charge += val;
1367         put_cpu_var(memcg_stock);
1368 }
1369
1370 /*
1371  * Tries to drain stocked charges in other cpus. This function is asynchronous
1372  * and just put a work per cpu for draining localy on each cpu. Caller can
1373  * expects some charges will be back to res_counter later but cannot wait for
1374  * it.
1375  */
1376 static void drain_all_stock_async(void)
1377 {
1378         int cpu;
1379         /* This function is for scheduling "drain" in asynchronous way.
1380          * The result of "drain" is not directly handled by callers. Then,
1381          * if someone is calling drain, we don't have to call drain more.
1382          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1383          * there is a race. We just do loose check here.
1384          */
1385         if (atomic_read(&memcg_drain_count))
1386                 return;
1387         /* Notify other cpus that system-wide "drain" is running */
1388         atomic_inc(&memcg_drain_count);
1389         get_online_cpus();
1390         for_each_online_cpu(cpu) {
1391                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1392                 schedule_work_on(cpu, &stock->work);
1393         }
1394         put_online_cpus();
1395         atomic_dec(&memcg_drain_count);
1396         /* We don't wait for flush_work */
1397 }
1398
1399 /* This is a synchronous drain interface. */
1400 static void drain_all_stock_sync(void)
1401 {
1402         /* called when force_empty is called */
1403         atomic_inc(&memcg_drain_count);
1404         schedule_on_each_cpu(drain_local_stock);
1405         atomic_dec(&memcg_drain_count);
1406 }
1407
1408 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1409                                         unsigned long action,
1410                                         void *hcpu)
1411 {
1412         int cpu = (unsigned long)hcpu;
1413         struct memcg_stock_pcp *stock;
1414
1415         if (action != CPU_DEAD)
1416                 return NOTIFY_OK;
1417         stock = &per_cpu(memcg_stock, cpu);
1418         drain_stock(stock);
1419         return NOTIFY_OK;
1420 }
1421
1422 /*
1423  * Unlike exported interface, "oom" parameter is added. if oom==true,
1424  * oom-killer can be invoked.
1425  */
1426 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1427                         gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1428 {
1429         struct mem_cgroup *mem, *mem_over_limit;
1430         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1431         struct res_counter *fail_res;
1432         int csize = CHARGE_SIZE;
1433
1434         if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1435                 /* Don't account this! */
1436                 *memcg = NULL;
1437                 return 0;
1438         }
1439
1440         /*
1441          * We always charge the cgroup the mm_struct belongs to.
1442          * The mm_struct's mem_cgroup changes on task migration if the
1443          * thread group leader migrates. It's possible that mm is not
1444          * set, if so charge the init_mm (happens for pagecache usage).
1445          */
1446         mem = *memcg;
1447         if (likely(!mem)) {
1448                 mem = try_get_mem_cgroup_from_mm(mm);
1449                 *memcg = mem;
1450         } else {
1451                 css_get(&mem->css);
1452         }
1453         if (unlikely(!mem))
1454                 return 0;
1455
1456         VM_BUG_ON(css_is_removed(&mem->css));
1457         if (mem_cgroup_is_root(mem))
1458                 goto done;
1459
1460         while (1) {
1461                 int ret = 0;
1462                 unsigned long flags = 0;
1463
1464                 if (consume_stock(mem))
1465                         goto done;
1466
1467                 ret = res_counter_charge(&mem->res, csize, &fail_res);
1468                 if (likely(!ret)) {
1469                         if (!do_swap_account)
1470                                 break;
1471                         ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1472                         if (likely(!ret))
1473                                 break;
1474                         /* mem+swap counter fails */
1475                         res_counter_uncharge(&mem->res, csize);
1476                         flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1477                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1478                                                                         memsw);
1479                 } else
1480                         /* mem counter fails */
1481                         mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1482                                                                         res);
1483
1484                 /* reduce request size and retry */
1485                 if (csize > PAGE_SIZE) {
1486                         csize = PAGE_SIZE;
1487                         continue;
1488                 }
1489                 if (!(gfp_mask & __GFP_WAIT))
1490                         goto nomem;
1491
1492                 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1493                                                 gfp_mask, flags);
1494                 if (ret)
1495                         continue;
1496
1497                 /*
1498                  * try_to_free_mem_cgroup_pages() might not give us a full
1499                  * picture of reclaim. Some pages are reclaimed and might be
1500                  * moved to swap cache or just unmapped from the cgroup.
1501                  * Check the limit again to see if the reclaim reduced the
1502                  * current usage of the cgroup before giving up
1503                  *
1504                  */
1505                 if (mem_cgroup_check_under_limit(mem_over_limit))
1506                         continue;
1507
1508                 /* try to avoid oom while someone is moving charge */
1509                 if (mc.moving_task && current != mc.moving_task) {
1510                         struct mem_cgroup *from, *to;
1511                         bool do_continue = false;
1512                         /*
1513                          * There is a small race that "from" or "to" can be
1514                          * freed by rmdir, so we use css_tryget().
1515                          */
1516                         rcu_read_lock();
1517                         from = mc.from;
1518                         to = mc.to;
1519                         if (from && css_tryget(&from->css)) {
1520                                 if (mem_over_limit->use_hierarchy)
1521                                         do_continue = css_is_ancestor(
1522                                                         &from->css,
1523                                                         &mem_over_limit->css);
1524                                 else
1525                                         do_continue = (from == mem_over_limit);
1526                                 css_put(&from->css);
1527                         }
1528                         if (!do_continue && to && css_tryget(&to->css)) {
1529                                 if (mem_over_limit->use_hierarchy)
1530                                         do_continue = css_is_ancestor(
1531                                                         &to->css,
1532                                                         &mem_over_limit->css);
1533                                 else
1534                                         do_continue = (to == mem_over_limit);
1535                                 css_put(&to->css);
1536                         }
1537                         rcu_read_unlock();
1538                         if (do_continue) {
1539                                 DEFINE_WAIT(wait);
1540                                 prepare_to_wait(&mc.waitq, &wait,
1541                                                         TASK_INTERRUPTIBLE);
1542                                 /* moving charge context might have finished. */
1543                                 if (mc.moving_task)
1544                                         schedule();
1545                                 finish_wait(&mc.waitq, &wait);
1546                                 continue;
1547                         }
1548                 }
1549
1550                 if (!nr_retries--) {
1551                         if (oom) {
1552                                 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1553                                 record_last_oom(mem_over_limit);
1554                         }
1555                         goto nomem;
1556                 }
1557         }
1558         if (csize > PAGE_SIZE)
1559                 refill_stock(mem, csize - PAGE_SIZE);
1560 done:
1561         return 0;
1562 nomem:
1563         css_put(&mem->css);
1564         return -ENOMEM;
1565 }
1566
1567 /*
1568  * Somemtimes we have to undo a charge we got by try_charge().
1569  * This function is for that and do uncharge, put css's refcnt.
1570  * gotten by try_charge().
1571  */
1572 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1573                                                         unsigned long count)
1574 {
1575         if (!mem_cgroup_is_root(mem)) {
1576                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1577                 if (do_swap_account)
1578                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1579                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1580                 WARN_ON_ONCE(count > INT_MAX);
1581                 __css_put(&mem->css, (int)count);
1582         }
1583         /* we don't need css_put for root */
1584 }
1585
1586 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1587 {
1588         __mem_cgroup_cancel_charge(mem, 1);
1589 }
1590
1591 /*
1592  * A helper function to get mem_cgroup from ID. must be called under
1593  * rcu_read_lock(). The caller must check css_is_removed() or some if
1594  * it's concern. (dropping refcnt from swap can be called against removed
1595  * memcg.)
1596  */
1597 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1598 {
1599         struct cgroup_subsys_state *css;
1600
1601         /* ID 0 is unused ID */
1602         if (!id)
1603                 return NULL;
1604         css = css_lookup(&mem_cgroup_subsys, id);
1605         if (!css)
1606                 return NULL;
1607         return container_of(css, struct mem_cgroup, css);
1608 }
1609
1610 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1611 {
1612         struct mem_cgroup *mem = NULL;
1613         struct page_cgroup *pc;
1614         unsigned short id;
1615         swp_entry_t ent;
1616
1617         VM_BUG_ON(!PageLocked(page));
1618
1619         pc = lookup_page_cgroup(page);
1620         lock_page_cgroup(pc);
1621         if (PageCgroupUsed(pc)) {
1622                 mem = pc->mem_cgroup;
1623                 if (mem && !css_tryget(&mem->css))
1624                         mem = NULL;
1625         } else if (PageSwapCache(page)) {
1626                 ent.val = page_private(page);
1627                 id = lookup_swap_cgroup(ent);
1628                 rcu_read_lock();
1629                 mem = mem_cgroup_lookup(id);
1630                 if (mem && !css_tryget(&mem->css))
1631                         mem = NULL;
1632                 rcu_read_unlock();
1633         }
1634         unlock_page_cgroup(pc);
1635         return mem;
1636 }
1637
1638 /*
1639  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1640  * USED state. If already USED, uncharge and return.
1641  */
1642
1643 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1644                                      struct page_cgroup *pc,
1645                                      enum charge_type ctype)
1646 {
1647         /* try_charge() can return NULL to *memcg, taking care of it. */
1648         if (!mem)
1649                 return;
1650
1651         lock_page_cgroup(pc);
1652         if (unlikely(PageCgroupUsed(pc))) {
1653                 unlock_page_cgroup(pc);
1654                 mem_cgroup_cancel_charge(mem);
1655                 return;
1656         }
1657
1658         pc->mem_cgroup = mem;
1659         /*
1660          * We access a page_cgroup asynchronously without lock_page_cgroup().
1661          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1662          * is accessed after testing USED bit. To make pc->mem_cgroup visible
1663          * before USED bit, we need memory barrier here.
1664          * See mem_cgroup_add_lru_list(), etc.
1665          */
1666         smp_wmb();
1667         switch (ctype) {
1668         case MEM_CGROUP_CHARGE_TYPE_CACHE:
1669         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1670                 SetPageCgroupCache(pc);
1671                 SetPageCgroupUsed(pc);
1672                 break;
1673         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1674                 ClearPageCgroupCache(pc);
1675                 SetPageCgroupUsed(pc);
1676                 break;
1677         default:
1678                 break;
1679         }
1680
1681         mem_cgroup_charge_statistics(mem, pc, true);
1682
1683         unlock_page_cgroup(pc);
1684         /*
1685          * "charge_statistics" updated event counter. Then, check it.
1686          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1687          * if they exceeds softlimit.
1688          */
1689         if (mem_cgroup_soft_limit_check(mem))
1690                 mem_cgroup_update_tree(mem, pc->page);
1691         if (mem_cgroup_threshold_check(mem))
1692                 mem_cgroup_threshold(mem);
1693
1694 }
1695
1696 /**
1697  * __mem_cgroup_move_account - move account of the page
1698  * @pc: page_cgroup of the page.
1699  * @from: mem_cgroup which the page is moved from.
1700  * @to: mem_cgroup which the page is moved to. @from != @to.
1701  * @uncharge: whether we should call uncharge and css_put against @from.
1702  *
1703  * The caller must confirm following.
1704  * - page is not on LRU (isolate_page() is useful.)
1705  * - the pc is locked, used, and ->mem_cgroup points to @from.
1706  *
1707  * This function doesn't do "charge" nor css_get to new cgroup. It should be
1708  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1709  * true, this function does "uncharge" from old cgroup, but it doesn't if
1710  * @uncharge is false, so a caller should do "uncharge".
1711  */
1712
1713 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1714         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1715 {
1716         struct page *page;
1717
1718         VM_BUG_ON(from == to);
1719         VM_BUG_ON(PageLRU(pc->page));
1720         VM_BUG_ON(!PageCgroupLocked(pc));
1721         VM_BUG_ON(!PageCgroupUsed(pc));
1722         VM_BUG_ON(pc->mem_cgroup != from);
1723
1724         page = pc->page;
1725         if (page_mapped(page) && !PageAnon(page)) {
1726                 /* Update mapped_file data for mem_cgroup */
1727                 preempt_disable();
1728                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1729                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1730                 preempt_enable();
1731         }
1732         mem_cgroup_charge_statistics(from, pc, false);
1733         if (uncharge)
1734                 /* This is not "cancel", but cancel_charge does all we need. */
1735                 mem_cgroup_cancel_charge(from);
1736
1737         /* caller should have done css_get */
1738         pc->mem_cgroup = to;
1739         mem_cgroup_charge_statistics(to, pc, true);
1740         /*
1741          * We charges against "to" which may not have any tasks. Then, "to"
1742          * can be under rmdir(). But in current implementation, caller of
1743          * this function is just force_empty() and move charge, so it's
1744          * garanteed that "to" is never removed. So, we don't check rmdir
1745          * status here.
1746          */
1747 }
1748
1749 /*
1750  * check whether the @pc is valid for moving account and call
1751  * __mem_cgroup_move_account()
1752  */
1753 static int mem_cgroup_move_account(struct page_cgroup *pc,
1754                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1755 {
1756         int ret = -EINVAL;
1757         lock_page_cgroup(pc);
1758         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1759                 __mem_cgroup_move_account(pc, from, to, uncharge);
1760                 ret = 0;
1761         }
1762         unlock_page_cgroup(pc);
1763         return ret;
1764 }
1765
1766 /*
1767  * move charges to its parent.
1768  */
1769
1770 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1771                                   struct mem_cgroup *child,
1772                                   gfp_t gfp_mask)
1773 {
1774         struct page *page = pc->page;
1775         struct cgroup *cg = child->css.cgroup;
1776         struct cgroup *pcg = cg->parent;
1777         struct mem_cgroup *parent;
1778         int ret;
1779
1780         /* Is ROOT ? */
1781         if (!pcg)
1782                 return -EINVAL;
1783
1784         ret = -EBUSY;
1785         if (!get_page_unless_zero(page))
1786                 goto out;
1787         if (isolate_lru_page(page))
1788                 goto put;
1789
1790         parent = mem_cgroup_from_cont(pcg);
1791         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1792         if (ret || !parent)
1793                 goto put_back;
1794
1795         ret = mem_cgroup_move_account(pc, child, parent, true);
1796         if (ret)
1797                 mem_cgroup_cancel_charge(parent);
1798 put_back:
1799         putback_lru_page(page);
1800 put:
1801         put_page(page);
1802 out:
1803         return ret;
1804 }
1805
1806 /*
1807  * Charge the memory controller for page usage.
1808  * Return
1809  * 0 if the charge was successful
1810  * < 0 if the cgroup is over its limit
1811  */
1812 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1813                                 gfp_t gfp_mask, enum charge_type ctype,
1814                                 struct mem_cgroup *memcg)
1815 {
1816         struct mem_cgroup *mem;
1817         struct page_cgroup *pc;
1818         int ret;
1819
1820         pc = lookup_page_cgroup(page);
1821         /* can happen at boot */
1822         if (unlikely(!pc))
1823                 return 0;
1824         prefetchw(pc);
1825
1826         mem = memcg;
1827         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1828         if (ret || !mem)
1829                 return ret;
1830
1831         __mem_cgroup_commit_charge(mem, pc, ctype);
1832         return 0;
1833 }
1834
1835 int mem_cgroup_newpage_charge(struct page *page,
1836                               struct mm_struct *mm, gfp_t gfp_mask)
1837 {
1838         if (mem_cgroup_disabled())
1839                 return 0;
1840         if (PageCompound(page))
1841                 return 0;
1842         /*
1843          * If already mapped, we don't have to account.
1844          * If page cache, page->mapping has address_space.
1845          * But page->mapping may have out-of-use anon_vma pointer,
1846          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1847          * is NULL.
1848          */
1849         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1850                 return 0;
1851         if (unlikely(!mm))
1852                 mm = &init_mm;
1853         return mem_cgroup_charge_common(page, mm, gfp_mask,
1854                                 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1855 }
1856
1857 static void
1858 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1859                                         enum charge_type ctype);
1860
1861 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1862                                 gfp_t gfp_mask)
1863 {
1864         struct mem_cgroup *mem = NULL;
1865         int ret;
1866
1867         if (mem_cgroup_disabled())
1868                 return 0;
1869         if (PageCompound(page))
1870                 return 0;
1871         /*
1872          * Corner case handling. This is called from add_to_page_cache()
1873          * in usual. But some FS (shmem) precharges this page before calling it
1874          * and call add_to_page_cache() with GFP_NOWAIT.
1875          *
1876          * For GFP_NOWAIT case, the page may be pre-charged before calling
1877          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1878          * charge twice. (It works but has to pay a bit larger cost.)
1879          * And when the page is SwapCache, it should take swap information
1880          * into account. This is under lock_page() now.
1881          */
1882         if (!(gfp_mask & __GFP_WAIT)) {
1883                 struct page_cgroup *pc;
1884
1885
1886                 pc = lookup_page_cgroup(page);
1887                 if (!pc)
1888                         return 0;
1889                 lock_page_cgroup(pc);
1890                 if (PageCgroupUsed(pc)) {
1891                         unlock_page_cgroup(pc);
1892                         return 0;
1893                 }
1894                 unlock_page_cgroup(pc);
1895         }
1896
1897         if (unlikely(!mm && !mem))
1898                 mm = &init_mm;
1899
1900         if (page_is_file_cache(page))
1901                 return mem_cgroup_charge_common(page, mm, gfp_mask,
1902                                 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1903
1904         /* shmem */
1905         if (PageSwapCache(page)) {
1906                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1907                 if (!ret)
1908                         __mem_cgroup_commit_charge_swapin(page, mem,
1909                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
1910         } else
1911                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1912                                         MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1913
1914         return ret;
1915 }
1916
1917 /*
1918  * While swap-in, try_charge -> commit or cancel, the page is locked.
1919  * And when try_charge() successfully returns, one refcnt to memcg without
1920  * struct page_cgroup is acquired. This refcnt will be consumed by
1921  * "commit()" or removed by "cancel()"
1922  */
1923 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1924                                  struct page *page,
1925                                  gfp_t mask, struct mem_cgroup **ptr)
1926 {
1927         struct mem_cgroup *mem;
1928         int ret;
1929
1930         if (mem_cgroup_disabled())
1931                 return 0;
1932
1933         if (!do_swap_account)
1934                 goto charge_cur_mm;
1935         /*
1936          * A racing thread's fault, or swapoff, may have already updated
1937          * the pte, and even removed page from swap cache: in those cases
1938          * do_swap_page()'s pte_same() test will fail; but there's also a
1939          * KSM case which does need to charge the page.
1940          */
1941         if (!PageSwapCache(page))
1942                 goto charge_cur_mm;
1943         mem = try_get_mem_cgroup_from_page(page);
1944         if (!mem)
1945                 goto charge_cur_mm;
1946         *ptr = mem;
1947         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1948         /* drop extra refcnt from tryget */
1949         css_put(&mem->css);
1950         return ret;
1951 charge_cur_mm:
1952         if (unlikely(!mm))
1953                 mm = &init_mm;
1954         return __mem_cgroup_try_charge(mm, mask, ptr, true);
1955 }
1956
1957 static void
1958 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1959                                         enum charge_type ctype)
1960 {
1961         struct page_cgroup *pc;
1962
1963         if (mem_cgroup_disabled())
1964                 return;
1965         if (!ptr)
1966                 return;
1967         cgroup_exclude_rmdir(&ptr->css);
1968         pc = lookup_page_cgroup(page);
1969         mem_cgroup_lru_del_before_commit_swapcache(page);
1970         __mem_cgroup_commit_charge(ptr, pc, ctype);
1971         mem_cgroup_lru_add_after_commit_swapcache(page);
1972         /*
1973          * Now swap is on-memory. This means this page may be
1974          * counted both as mem and swap....double count.
1975          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1976          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1977          * may call delete_from_swap_cache() before reach here.
1978          */
1979         if (do_swap_account && PageSwapCache(page)) {
1980                 swp_entry_t ent = {.val = page_private(page)};
1981                 unsigned short id;
1982                 struct mem_cgroup *memcg;
1983
1984                 id = swap_cgroup_record(ent, 0);
1985                 rcu_read_lock();
1986                 memcg = mem_cgroup_lookup(id);
1987                 if (memcg) {
1988                         /*
1989                          * This recorded memcg can be obsolete one. So, avoid
1990                          * calling css_tryget
1991                          */
1992                         if (!mem_cgroup_is_root(memcg))
1993                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1994                         mem_cgroup_swap_statistics(memcg, false);
1995                         mem_cgroup_put(memcg);
1996                 }
1997                 rcu_read_unlock();
1998         }
1999         /*
2000          * At swapin, we may charge account against cgroup which has no tasks.
2001          * So, rmdir()->pre_destroy() can be called while we do this charge.
2002          * In that case, we need to call pre_destroy() again. check it here.
2003          */
2004         cgroup_release_and_wakeup_rmdir(&ptr->css);
2005 }
2006
2007 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2008 {
2009         __mem_cgroup_commit_charge_swapin(page, ptr,
2010                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2011 }
2012
2013 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2014 {
2015         if (mem_cgroup_disabled())
2016                 return;
2017         if (!mem)
2018                 return;
2019         mem_cgroup_cancel_charge(mem);
2020 }
2021
2022 static void
2023 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2024 {
2025         struct memcg_batch_info *batch = NULL;
2026         bool uncharge_memsw = true;
2027         /* If swapout, usage of swap doesn't decrease */
2028         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2029                 uncharge_memsw = false;
2030         /*
2031          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2032          * In those cases, all pages freed continously can be expected to be in
2033          * the same cgroup and we have chance to coalesce uncharges.
2034          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2035          * because we want to do uncharge as soon as possible.
2036          */
2037         if (!current->memcg_batch.do_batch || test_thread_flag(TIF_MEMDIE))
2038                 goto direct_uncharge;
2039
2040         batch = &current->memcg_batch;
2041         /*
2042          * In usual, we do css_get() when we remember memcg pointer.
2043          * But in this case, we keep res->usage until end of a series of
2044          * uncharges. Then, it's ok to ignore memcg's refcnt.
2045          */
2046         if (!batch->memcg)
2047                 batch->memcg = mem;
2048         /*
2049          * In typical case, batch->memcg == mem. This means we can
2050          * merge a series of uncharges to an uncharge of res_counter.
2051          * If not, we uncharge res_counter ony by one.
2052          */
2053         if (batch->memcg != mem)
2054                 goto direct_uncharge;
2055         /* remember freed charge and uncharge it later */
2056         batch->bytes += PAGE_SIZE;
2057         if (uncharge_memsw)
2058                 batch->memsw_bytes += PAGE_SIZE;
2059         return;
2060 direct_uncharge:
2061         res_counter_uncharge(&mem->res, PAGE_SIZE);
2062         if (uncharge_memsw)
2063                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2064         return;
2065 }
2066
2067 /*
2068  * uncharge if !page_mapped(page)
2069  */
2070 static struct mem_cgroup *
2071 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2072 {
2073         struct page_cgroup *pc;
2074         struct mem_cgroup *mem = NULL;
2075         struct mem_cgroup_per_zone *mz;
2076
2077         if (mem_cgroup_disabled())
2078                 return NULL;
2079
2080         if (PageSwapCache(page))
2081                 return NULL;
2082
2083         /*
2084          * Check if our page_cgroup is valid
2085          */
2086         pc = lookup_page_cgroup(page);
2087         if (unlikely(!pc || !PageCgroupUsed(pc)))
2088                 return NULL;
2089
2090         lock_page_cgroup(pc);
2091
2092         mem = pc->mem_cgroup;
2093
2094         if (!PageCgroupUsed(pc))
2095                 goto unlock_out;
2096
2097         switch (ctype) {
2098         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2099         case MEM_CGROUP_CHARGE_TYPE_DROP:
2100                 if (page_mapped(page))
2101                         goto unlock_out;
2102                 break;
2103         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2104                 if (!PageAnon(page)) {  /* Shared memory */
2105                         if (page->mapping && !page_is_file_cache(page))
2106                                 goto unlock_out;
2107                 } else if (page_mapped(page)) /* Anon */
2108                                 goto unlock_out;
2109                 break;
2110         default:
2111                 break;
2112         }
2113
2114         if (!mem_cgroup_is_root(mem))
2115                 __do_uncharge(mem, ctype);
2116         if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2117                 mem_cgroup_swap_statistics(mem, true);
2118         mem_cgroup_charge_statistics(mem, pc, false);
2119
2120         ClearPageCgroupUsed(pc);
2121         /*
2122          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2123          * freed from LRU. This is safe because uncharged page is expected not
2124          * to be reused (freed soon). Exception is SwapCache, it's handled by
2125          * special functions.
2126          */
2127
2128         mz = page_cgroup_zoneinfo(pc);
2129         unlock_page_cgroup(pc);
2130
2131         if (mem_cgroup_soft_limit_check(mem))
2132                 mem_cgroup_update_tree(mem, page);
2133         if (mem_cgroup_threshold_check(mem))
2134                 mem_cgroup_threshold(mem);
2135         /* at swapout, this memcg will be accessed to record to swap */
2136         if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2137                 css_put(&mem->css);
2138
2139         return mem;
2140
2141 unlock_out:
2142         unlock_page_cgroup(pc);
2143         return NULL;
2144 }
2145
2146 void mem_cgroup_uncharge_page(struct page *page)
2147 {
2148         /* early check. */
2149         if (page_mapped(page))
2150                 return;
2151         if (page->mapping && !PageAnon(page))
2152                 return;
2153         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2154 }
2155
2156 void mem_cgroup_uncharge_cache_page(struct page *page)
2157 {
2158         VM_BUG_ON(page_mapped(page));
2159         VM_BUG_ON(page->mapping);
2160         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2161 }
2162
2163 /*
2164  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2165  * In that cases, pages are freed continuously and we can expect pages
2166  * are in the same memcg. All these calls itself limits the number of
2167  * pages freed at once, then uncharge_start/end() is called properly.
2168  * This may be called prural(2) times in a context,
2169  */
2170
2171 void mem_cgroup_uncharge_start(void)
2172 {
2173         current->memcg_batch.do_batch++;
2174         /* We can do nest. */
2175         if (current->memcg_batch.do_batch == 1) {
2176                 current->memcg_batch.memcg = NULL;
2177                 current->memcg_batch.bytes = 0;
2178                 current->memcg_batch.memsw_bytes = 0;
2179         }
2180 }
2181
2182 void mem_cgroup_uncharge_end(void)
2183 {
2184         struct memcg_batch_info *batch = &current->memcg_batch;
2185
2186         if (!batch->do_batch)
2187                 return;
2188
2189         batch->do_batch--;
2190         if (batch->do_batch) /* If stacked, do nothing. */
2191                 return;
2192
2193         if (!batch->memcg)
2194                 return;
2195         /*
2196          * This "batch->memcg" is valid without any css_get/put etc...
2197          * bacause we hide charges behind us.
2198          */
2199         if (batch->bytes)
2200                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2201         if (batch->memsw_bytes)
2202                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2203         /* forget this pointer (for sanity check) */
2204         batch->memcg = NULL;
2205 }
2206
2207 #ifdef CONFIG_SWAP
2208 /*
2209  * called after __delete_from_swap_cache() and drop "page" account.
2210  * memcg information is recorded to swap_cgroup of "ent"
2211  */
2212 void
2213 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2214 {
2215         struct mem_cgroup *memcg;
2216         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2217
2218         if (!swapout) /* this was a swap cache but the swap is unused ! */
2219                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2220
2221         memcg = __mem_cgroup_uncharge_common(page, ctype);
2222
2223         /* record memcg information */
2224         if (do_swap_account && swapout && memcg) {
2225                 swap_cgroup_record(ent, css_id(&memcg->css));
2226                 mem_cgroup_get(memcg);
2227         }
2228         if (swapout && memcg)
2229                 css_put(&memcg->css);
2230 }
2231 #endif
2232
2233 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2234 /*
2235  * called from swap_entry_free(). remove record in swap_cgroup and
2236  * uncharge "memsw" account.
2237  */
2238 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2239 {
2240         struct mem_cgroup *memcg;
2241         unsigned short id;
2242
2243         if (!do_swap_account)
2244                 return;
2245
2246         id = swap_cgroup_record(ent, 0);
2247         rcu_read_lock();
2248         memcg = mem_cgroup_lookup(id);
2249         if (memcg) {
2250                 /*
2251                  * We uncharge this because swap is freed.
2252                  * This memcg can be obsolete one. We avoid calling css_tryget
2253                  */
2254                 if (!mem_cgroup_is_root(memcg))
2255                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2256                 mem_cgroup_swap_statistics(memcg, false);
2257                 mem_cgroup_put(memcg);
2258         }
2259         rcu_read_unlock();
2260 }
2261
2262 /**
2263  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2264  * @entry: swap entry to be moved
2265  * @from:  mem_cgroup which the entry is moved from
2266  * @to:  mem_cgroup which the entry is moved to
2267  * @need_fixup: whether we should fixup res_counters and refcounts.
2268  *
2269  * It succeeds only when the swap_cgroup's record for this entry is the same
2270  * as the mem_cgroup's id of @from.
2271  *
2272  * Returns 0 on success, -EINVAL on failure.
2273  *
2274  * The caller must have charged to @to, IOW, called res_counter_charge() about
2275  * both res and memsw, and called css_get().
2276  */
2277 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2278                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2279 {
2280         unsigned short old_id, new_id;
2281
2282         old_id = css_id(&from->css);
2283         new_id = css_id(&to->css);
2284
2285         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2286                 mem_cgroup_swap_statistics(from, false);
2287                 mem_cgroup_swap_statistics(to, true);
2288                 /*
2289                  * This function is only called from task migration context now.
2290                  * It postpones res_counter and refcount handling till the end
2291                  * of task migration(mem_cgroup_clear_mc()) for performance
2292                  * improvement. But we cannot postpone mem_cgroup_get(to)
2293                  * because if the process that has been moved to @to does
2294                  * swap-in, the refcount of @to might be decreased to 0.
2295                  */
2296                 mem_cgroup_get(to);
2297                 if (need_fixup) {
2298                         if (!mem_cgroup_is_root(from))
2299                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2300                         mem_cgroup_put(from);
2301                         /*
2302                          * we charged both to->res and to->memsw, so we should
2303                          * uncharge to->res.
2304                          */
2305                         if (!mem_cgroup_is_root(to))
2306                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2307                         css_put(&to->css);
2308                 }
2309                 return 0;
2310         }
2311         return -EINVAL;
2312 }
2313 #else
2314 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2315                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2316 {
2317         return -EINVAL;
2318 }
2319 #endif
2320
2321 /*
2322  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2323  * page belongs to.
2324  */
2325 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2326 {
2327         struct page_cgroup *pc;
2328         struct mem_cgroup *mem = NULL;
2329         int ret = 0;
2330
2331         if (mem_cgroup_disabled())
2332                 return 0;
2333
2334         pc = lookup_page_cgroup(page);
2335         lock_page_cgroup(pc);
2336         if (PageCgroupUsed(pc)) {
2337                 mem = pc->mem_cgroup;
2338                 css_get(&mem->css);
2339         }
2340         unlock_page_cgroup(pc);
2341
2342         if (mem) {
2343                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
2344                 css_put(&mem->css);
2345         }
2346         *ptr = mem;
2347         return ret;
2348 }
2349
2350 /* remove redundant charge if migration failed*/
2351 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2352                 struct page *oldpage, struct page *newpage)
2353 {
2354         struct page *target, *unused;
2355         struct page_cgroup *pc;
2356         enum charge_type ctype;
2357
2358         if (!mem)
2359                 return;
2360         cgroup_exclude_rmdir(&mem->css);
2361         /* at migration success, oldpage->mapping is NULL. */
2362         if (oldpage->mapping) {
2363                 target = oldpage;
2364                 unused = NULL;
2365         } else {
2366                 target = newpage;
2367                 unused = oldpage;
2368         }
2369
2370         if (PageAnon(target))
2371                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2372         else if (page_is_file_cache(target))
2373                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2374         else
2375                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2376
2377         /* unused page is not on radix-tree now. */
2378         if (unused)
2379                 __mem_cgroup_uncharge_common(unused, ctype);
2380
2381         pc = lookup_page_cgroup(target);
2382         /*
2383          * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2384          * So, double-counting is effectively avoided.
2385          */
2386         __mem_cgroup_commit_charge(mem, pc, ctype);
2387
2388         /*
2389          * Both of oldpage and newpage are still under lock_page().
2390          * Then, we don't have to care about race in radix-tree.
2391          * But we have to be careful that this page is unmapped or not.
2392          *
2393          * There is a case for !page_mapped(). At the start of
2394          * migration, oldpage was mapped. But now, it's zapped.
2395          * But we know *target* page is not freed/reused under us.
2396          * mem_cgroup_uncharge_page() does all necessary checks.
2397          */
2398         if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2399                 mem_cgroup_uncharge_page(target);
2400         /*
2401          * At migration, we may charge account against cgroup which has no tasks
2402          * So, rmdir()->pre_destroy() can be called while we do this charge.
2403          * In that case, we need to call pre_destroy() again. check it here.
2404          */
2405         cgroup_release_and_wakeup_rmdir(&mem->css);
2406 }
2407
2408 /*
2409  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2410  * Calling hierarchical_reclaim is not enough because we should update
2411  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2412  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2413  * not from the memcg which this page would be charged to.
2414  * try_charge_swapin does all of these works properly.
2415  */
2416 int mem_cgroup_shmem_charge_fallback(struct page *page,
2417                             struct mm_struct *mm,
2418                             gfp_t gfp_mask)
2419 {
2420         struct mem_cgroup *mem = NULL;
2421         int ret;
2422
2423         if (mem_cgroup_disabled())
2424                 return 0;
2425
2426         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2427         if (!ret)
2428                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2429
2430         return ret;
2431 }
2432
2433 static DEFINE_MUTEX(set_limit_mutex);
2434
2435 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2436                                 unsigned long long val)
2437 {
2438         int retry_count;
2439         u64 memswlimit;
2440         int ret = 0;
2441         int children = mem_cgroup_count_children(memcg);
2442         u64 curusage, oldusage;
2443
2444         /*
2445          * For keeping hierarchical_reclaim simple, how long we should retry
2446          * is depends on callers. We set our retry-count to be function
2447          * of # of children which we should visit in this loop.
2448          */
2449         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2450
2451         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2452
2453         while (retry_count) {
2454                 if (signal_pending(current)) {
2455                         ret = -EINTR;
2456                         break;
2457                 }
2458                 /*
2459                  * Rather than hide all in some function, I do this in
2460                  * open coded manner. You see what this really does.
2461                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2462                  */
2463                 mutex_lock(&set_limit_mutex);
2464                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2465                 if (memswlimit < val) {
2466                         ret = -EINVAL;
2467                         mutex_unlock(&set_limit_mutex);
2468                         break;
2469                 }
2470                 ret = res_counter_set_limit(&memcg->res, val);
2471                 if (!ret) {
2472                         if (memswlimit == val)
2473                                 memcg->memsw_is_minimum = true;
2474                         else
2475                                 memcg->memsw_is_minimum = false;
2476                 }
2477                 mutex_unlock(&set_limit_mutex);
2478
2479                 if (!ret)
2480                         break;
2481
2482                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2483                                                 MEM_CGROUP_RECLAIM_SHRINK);
2484                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2485                 /* Usage is reduced ? */
2486                 if (curusage >= oldusage)
2487                         retry_count--;
2488                 else
2489                         oldusage = curusage;
2490         }
2491
2492         return ret;
2493 }
2494
2495 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2496                                         unsigned long long val)
2497 {
2498         int retry_count;
2499         u64 memlimit, oldusage, curusage;
2500         int children = mem_cgroup_count_children(memcg);
2501         int ret = -EBUSY;
2502
2503         /* see mem_cgroup_resize_res_limit */
2504         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2505         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2506         while (retry_count) {
2507                 if (signal_pending(current)) {
2508                         ret = -EINTR;
2509                         break;
2510                 }
2511                 /*
2512                  * Rather than hide all in some function, I do this in
2513                  * open coded manner. You see what this really does.
2514                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2515                  */
2516                 mutex_lock(&set_limit_mutex);
2517                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2518                 if (memlimit > val) {
2519                         ret = -EINVAL;
2520                         mutex_unlock(&set_limit_mutex);
2521                         break;
2522                 }
2523                 ret = res_counter_set_limit(&memcg->memsw, val);
2524                 if (!ret) {
2525                         if (memlimit == val)
2526                                 memcg->memsw_is_minimum = true;
2527                         else
2528                                 memcg->memsw_is_minimum = false;
2529                 }
2530                 mutex_unlock(&set_limit_mutex);
2531
2532                 if (!ret)
2533                         break;
2534
2535                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2536                                                 MEM_CGROUP_RECLAIM_NOSWAP |
2537                                                 MEM_CGROUP_RECLAIM_SHRINK);
2538                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2539                 /* Usage is reduced ? */
2540                 if (curusage >= oldusage)
2541                         retry_count--;
2542                 else
2543                         oldusage = curusage;
2544         }
2545         return ret;
2546 }
2547
2548 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2549                                                 gfp_t gfp_mask, int nid,
2550                                                 int zid)
2551 {
2552         unsigned long nr_reclaimed = 0;
2553         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2554         unsigned long reclaimed;
2555         int loop = 0;
2556         struct mem_cgroup_tree_per_zone *mctz;
2557         unsigned long long excess;
2558
2559         if (order > 0)
2560                 return 0;
2561
2562         mctz = soft_limit_tree_node_zone(nid, zid);
2563         /*
2564          * This loop can run a while, specially if mem_cgroup's continuously
2565          * keep exceeding their soft limit and putting the system under
2566          * pressure
2567          */
2568         do {
2569                 if (next_mz)
2570                         mz = next_mz;
2571                 else
2572                         mz = mem_cgroup_largest_soft_limit_node(mctz);
2573                 if (!mz)
2574                         break;
2575
2576                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2577                                                 gfp_mask,
2578                                                 MEM_CGROUP_RECLAIM_SOFT);
2579                 nr_reclaimed += reclaimed;
2580                 spin_lock(&mctz->lock);
2581
2582                 /*
2583                  * If we failed to reclaim anything from this memory cgroup
2584                  * it is time to move on to the next cgroup
2585                  */
2586                 next_mz = NULL;
2587                 if (!reclaimed) {
2588                         do {
2589                                 /*
2590                                  * Loop until we find yet another one.
2591                                  *
2592                                  * By the time we get the soft_limit lock
2593                                  * again, someone might have aded the
2594                                  * group back on the RB tree. Iterate to
2595                                  * make sure we get a different mem.
2596                                  * mem_cgroup_largest_soft_limit_node returns
2597                                  * NULL if no other cgroup is present on
2598                                  * the tree
2599                                  */
2600                                 next_mz =
2601                                 __mem_cgroup_largest_soft_limit_node(mctz);
2602                                 if (next_mz == mz) {
2603                                         css_put(&next_mz->mem->css);
2604                                         next_mz = NULL;
2605                                 } else /* next_mz == NULL or other memcg */
2606                                         break;
2607                         } while (1);
2608                 }
2609                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2610                 excess = res_counter_soft_limit_excess(&mz->mem->res);
2611                 /*
2612                  * One school of thought says that we should not add
2613                  * back the node to the tree if reclaim returns 0.
2614                  * But our reclaim could return 0, simply because due
2615                  * to priority we are exposing a smaller subset of
2616                  * memory to reclaim from. Consider this as a longer
2617                  * term TODO.
2618                  */
2619                 /* If excess == 0, no tree ops */
2620                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2621                 spin_unlock(&mctz->lock);
2622                 css_put(&mz->mem->css);
2623                 loop++;
2624                 /*
2625                  * Could not reclaim anything and there are no more
2626                  * mem cgroups to try or we seem to be looping without
2627                  * reclaiming anything.
2628                  */
2629                 if (!nr_reclaimed &&
2630                         (next_mz == NULL ||
2631                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2632                         break;
2633         } while (!nr_reclaimed);
2634         if (next_mz)
2635                 css_put(&next_mz->mem->css);
2636         return nr_reclaimed;
2637 }
2638
2639 /*
2640  * This routine traverse page_cgroup in given list and drop them all.
2641  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2642  */
2643 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2644                                 int node, int zid, enum lru_list lru)
2645 {
2646         struct zone *zone;
2647         struct mem_cgroup_per_zone *mz;
2648         struct page_cgroup *pc, *busy;
2649         unsigned long flags, loop;
2650         struct list_head *list;
2651         int ret = 0;
2652
2653         zone = &NODE_DATA(node)->node_zones[zid];
2654         mz = mem_cgroup_zoneinfo(mem, node, zid);
2655         list = &mz->lists[lru];
2656
2657         loop = MEM_CGROUP_ZSTAT(mz, lru);
2658         /* give some margin against EBUSY etc...*/
2659         loop += 256;
2660         busy = NULL;
2661         while (loop--) {
2662                 ret = 0;
2663                 spin_lock_irqsave(&zone->lru_lock, flags);
2664                 if (list_empty(list)) {
2665                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2666                         break;
2667                 }
2668                 pc = list_entry(list->prev, struct page_cgroup, lru);
2669                 if (busy == pc) {
2670                         list_move(&pc->lru, list);
2671                         busy = NULL;
2672                         spin_unlock_irqrestore(&zone->lru_lock, flags);
2673                         continue;
2674                 }
2675                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2676
2677                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2678                 if (ret == -ENOMEM)
2679                         break;
2680
2681                 if (ret == -EBUSY || ret == -EINVAL) {
2682                         /* found lock contention or "pc" is obsolete. */
2683                         busy = pc;
2684                         cond_resched();
2685                 } else
2686                         busy = NULL;
2687         }
2688
2689         if (!ret && !list_empty(list))
2690                 return -EBUSY;
2691         return ret;
2692 }
2693
2694 /*
2695  * make mem_cgroup's charge to be 0 if there is no task.
2696  * This enables deleting this mem_cgroup.
2697  */
2698 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2699 {
2700         int ret;
2701         int node, zid, shrink;
2702         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2703         struct cgroup *cgrp = mem->css.cgroup;
2704
2705         css_get(&mem->css);
2706
2707         shrink = 0;
2708         /* should free all ? */
2709         if (free_all)
2710                 goto try_to_free;
2711 move_account:
2712         do {
2713                 ret = -EBUSY;
2714                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2715                         goto out;
2716                 ret = -EINTR;
2717                 if (signal_pending(current))
2718                         goto out;
2719                 /* This is for making all *used* pages to be on LRU. */
2720                 lru_add_drain_all();
2721                 drain_all_stock_sync();
2722                 ret = 0;
2723                 for_each_node_state(node, N_HIGH_MEMORY) {
2724                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2725                                 enum lru_list l;
2726                                 for_each_lru(l) {
2727                                         ret = mem_cgroup_force_empty_list(mem,
2728                                                         node, zid, l);
2729                                         if (ret)
2730                                                 break;
2731                                 }
2732                         }
2733                         if (ret)
2734                                 break;
2735                 }
2736                 /* it seems parent cgroup doesn't have enough mem */
2737                 if (ret == -ENOMEM)
2738                         goto try_to_free;
2739                 cond_resched();
2740         /* "ret" should also be checked to ensure all lists are empty. */
2741         } while (mem->res.usage > 0 || ret);
2742 out:
2743         css_put(&mem->css);
2744         return ret;
2745
2746 try_to_free:
2747         /* returns EBUSY if there is a task or if we come here twice. */
2748         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2749                 ret = -EBUSY;
2750                 goto out;
2751         }
2752         /* we call try-to-free pages for make this cgroup empty */
2753         lru_add_drain_all();
2754         /* try to free all pages in this cgroup */
2755         shrink = 1;
2756         while (nr_retries && mem->res.usage > 0) {
2757                 int progress;
2758
2759                 if (signal_pending(current)) {
2760                         ret = -EINTR;
2761                         goto out;
2762                 }
2763                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2764                                                 false, get_swappiness(mem));
2765                 if (!progress) {
2766                         nr_retries--;
2767                         /* maybe some writeback is necessary */
2768                         congestion_wait(BLK_RW_ASYNC, HZ/10);
2769                 }
2770
2771         }
2772         lru_add_drain();
2773         /* try move_account...there may be some *locked* pages. */
2774         goto move_account;
2775 }
2776
2777 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2778 {
2779         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2780 }
2781
2782
2783 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2784 {
2785         return mem_cgroup_from_cont(cont)->use_hierarchy;
2786 }
2787
2788 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2789                                         u64 val)
2790 {
2791         int retval = 0;
2792         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2793         struct cgroup *parent = cont->parent;
2794         struct mem_cgroup *parent_mem = NULL;
2795
2796         if (parent)
2797                 parent_mem = mem_cgroup_from_cont(parent);
2798
2799         cgroup_lock();
2800         /*
2801          * If parent's use_hierarchy is set, we can't make any modifications
2802          * in the child subtrees. If it is unset, then the change can
2803          * occur, provided the current cgroup has no children.
2804          *
2805          * For the root cgroup, parent_mem is NULL, we allow value to be
2806          * set if there are no children.
2807          */
2808         if ((!parent_mem || !parent_mem->use_hierarchy) &&
2809                                 (val == 1 || val == 0)) {
2810                 if (list_empty(&cont->children))
2811                         mem->use_hierarchy = val;
2812                 else
2813                         retval = -EBUSY;
2814         } else
2815                 retval = -EINVAL;
2816         cgroup_unlock();
2817
2818         return retval;
2819 }
2820
2821 struct mem_cgroup_idx_data {
2822         s64 val;
2823         enum mem_cgroup_stat_index idx;
2824 };
2825
2826 static int
2827 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2828 {
2829         struct mem_cgroup_idx_data *d = data;
2830         d->val += mem_cgroup_read_stat(mem, d->idx);
2831         return 0;
2832 }
2833
2834 static void
2835 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2836                                 enum mem_cgroup_stat_index idx, s64 *val)
2837 {
2838         struct mem_cgroup_idx_data d;
2839         d.idx = idx;
2840         d.val = 0;
2841         mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2842         *val = d.val;
2843 }
2844
2845 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
2846 {
2847         u64 idx_val, val;
2848
2849         if (!mem_cgroup_is_root(mem)) {
2850                 if (!swap)
2851                         return res_counter_read_u64(&mem->res, RES_USAGE);
2852                 else
2853                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
2854         }
2855
2856         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
2857         val = idx_val;
2858         mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
2859         val += idx_val;
2860
2861         if (swap) {
2862                 mem_cgroup_get_recursive_idx_stat(mem,
2863                                 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2864                 val += idx_val;
2865         }
2866
2867         return val << PAGE_SHIFT;
2868 }
2869
2870 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2871 {
2872         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2873         u64 val;
2874         int type, name;
2875
2876         type = MEMFILE_TYPE(cft->private);
2877         name = MEMFILE_ATTR(cft->private);
2878         switch (type) {
2879         case _MEM:
2880                 if (name == RES_USAGE)
2881                         val = mem_cgroup_usage(mem, false);
2882                 else
2883                         val = res_counter_read_u64(&mem->res, name);
2884                 break;
2885         case _MEMSWAP:
2886                 if (name == RES_USAGE)
2887                         val = mem_cgroup_usage(mem, true);
2888                 else
2889                         val = res_counter_read_u64(&mem->memsw, name);
2890                 break;
2891         default:
2892                 BUG();
2893                 break;
2894         }
2895         return val;
2896 }
2897 /*
2898  * The user of this function is...
2899  * RES_LIMIT.
2900  */
2901 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2902                             const char *buffer)
2903 {
2904         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2905         int type, name;
2906         unsigned long long val;
2907         int ret;
2908
2909         type = MEMFILE_TYPE(cft->private);
2910         name = MEMFILE_ATTR(cft->private);
2911         switch (name) {
2912         case RES_LIMIT:
2913                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2914                         ret = -EINVAL;
2915                         break;
2916                 }
2917                 /* This function does all necessary parse...reuse it */
2918                 ret = res_counter_memparse_write_strategy(buffer, &val);
2919                 if (ret)
2920                         break;
2921                 if (type == _MEM)
2922                         ret = mem_cgroup_resize_limit(memcg, val);
2923                 else
2924                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
2925                 break;
2926         case RES_SOFT_LIMIT:
2927                 ret = res_counter_memparse_write_strategy(buffer, &val);
2928                 if (ret)
2929                         break;
2930                 /*
2931                  * For memsw, soft limits are hard to implement in terms
2932                  * of semantics, for now, we support soft limits for
2933                  * control without swap
2934                  */
2935                 if (type == _MEM)
2936                         ret = res_counter_set_soft_limit(&memcg->res, val);
2937                 else
2938                         ret = -EINVAL;
2939                 break;
2940         default:
2941                 ret = -EINVAL; /* should be BUG() ? */
2942                 break;
2943         }
2944         return ret;
2945 }
2946
2947 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2948                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2949 {
2950         struct cgroup *cgroup;
2951         unsigned long long min_limit, min_memsw_limit, tmp;
2952
2953         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2954         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2955         cgroup = memcg->css.cgroup;
2956         if (!memcg->use_hierarchy)
2957                 goto out;
2958
2959         while (cgroup->parent) {
2960                 cgroup = cgroup->parent;
2961                 memcg = mem_cgroup_from_cont(cgroup);
2962                 if (!memcg->use_hierarchy)
2963                         break;
2964                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2965                 min_limit = min(min_limit, tmp);
2966                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2967                 min_memsw_limit = min(min_memsw_limit, tmp);
2968         }
2969 out:
2970         *mem_limit = min_limit;
2971         *memsw_limit = min_memsw_limit;
2972         return;
2973 }
2974
2975 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2976 {
2977         struct mem_cgroup *mem;
2978         int type, name;
2979
2980         mem = mem_cgroup_from_cont(cont);
2981         type = MEMFILE_TYPE(event);
2982         name = MEMFILE_ATTR(event);
2983         switch (name) {
2984         case RES_MAX_USAGE:
2985                 if (type == _MEM)
2986                         res_counter_reset_max(&mem->res);
2987                 else
2988                         res_counter_reset_max(&mem->memsw);
2989                 break;
2990         case RES_FAILCNT:
2991                 if (type == _MEM)
2992                         res_counter_reset_failcnt(&mem->res);
2993                 else
2994                         res_counter_reset_failcnt(&mem->memsw);
2995                 break;
2996         }
2997
2998         return 0;
2999 }
3000
3001 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3002                                         struct cftype *cft)
3003 {
3004         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3005 }
3006
3007 #ifdef CONFIG_MMU
3008 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3009                                         struct cftype *cft, u64 val)
3010 {
3011         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3012
3013         if (val >= (1 << NR_MOVE_TYPE))
3014                 return -EINVAL;
3015         /*
3016          * We check this value several times in both in can_attach() and
3017          * attach(), so we need cgroup lock to prevent this value from being
3018          * inconsistent.
3019          */
3020         cgroup_lock();
3021         mem->move_charge_at_immigrate = val;
3022         cgroup_unlock();
3023
3024         return 0;
3025 }
3026 #else
3027 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3028                                         struct cftype *cft, u64 val)
3029 {
3030         return -ENOSYS;
3031 }
3032 #endif
3033
3034
3035 /* For read statistics */
3036 enum {
3037         MCS_CACHE,
3038         MCS_RSS,
3039         MCS_FILE_MAPPED,
3040         MCS_PGPGIN,
3041         MCS_PGPGOUT,
3042         MCS_SWAP,
3043         MCS_INACTIVE_ANON,
3044         MCS_ACTIVE_ANON,
3045         MCS_INACTIVE_FILE,
3046         MCS_ACTIVE_FILE,
3047         MCS_UNEVICTABLE,
3048         NR_MCS_STAT,
3049 };
3050
3051 struct mcs_total_stat {
3052         s64 stat[NR_MCS_STAT];
3053 };
3054
3055 struct {
3056         char *local_name;
3057         char *total_name;
3058 } memcg_stat_strings[NR_MCS_STAT] = {
3059         {"cache", "total_cache"},
3060         {"rss", "total_rss"},
3061         {"mapped_file", "total_mapped_file"},
3062         {"pgpgin", "total_pgpgin"},
3063         {"pgpgout", "total_pgpgout"},
3064         {"swap", "total_swap"},
3065         {"inactive_anon", "total_inactive_anon"},
3066         {"active_anon", "total_active_anon"},
3067         {"inactive_file", "total_inactive_file"},
3068         {"active_file", "total_active_file"},
3069         {"unevictable", "total_unevictable"}
3070 };
3071
3072
3073 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3074 {
3075         struct mcs_total_stat *s = data;
3076         s64 val;
3077
3078         /* per cpu stat */
3079         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3080         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3081         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3082         s->stat[MCS_RSS] += val * PAGE_SIZE;
3083         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3084         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3085         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3086         s->stat[MCS_PGPGIN] += val;
3087         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3088         s->stat[MCS_PGPGOUT] += val;
3089         if (do_swap_account) {
3090                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3091                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3092         }
3093
3094         /* per zone stat */
3095         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3096         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3097         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3098         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3099         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3100         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3101         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3102         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3103         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3104         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3105         return 0;
3106 }
3107
3108 static void
3109 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3110 {
3111         mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3112 }
3113
3114 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3115                                  struct cgroup_map_cb *cb)
3116 {
3117         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3118         struct mcs_total_stat mystat;
3119         int i;
3120
3121         memset(&mystat, 0, sizeof(mystat));
3122         mem_cgroup_get_local_stat(mem_cont, &mystat);
3123
3124         for (i = 0; i < NR_MCS_STAT; i++) {
3125                 if (i == MCS_SWAP && !do_swap_account)
3126                         continue;
3127                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3128         }
3129
3130         /* Hierarchical information */
3131         {
3132                 unsigned long long limit, memsw_limit;
3133                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3134                 cb->fill(cb, "hierarchical_memory_limit", limit);
3135                 if (do_swap_account)
3136                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3137         }
3138
3139         memset(&mystat, 0, sizeof(mystat));
3140         mem_cgroup_get_total_stat(mem_cont, &mystat);
3141         for (i = 0; i < NR_MCS_STAT; i++) {
3142                 if (i == MCS_SWAP && !do_swap_account)
3143                         continue;
3144                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3145         }
3146
3147 #ifdef CONFIG_DEBUG_VM
3148         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3149
3150         {
3151                 int nid, zid;
3152                 struct mem_cgroup_per_zone *mz;
3153                 unsigned long recent_rotated[2] = {0, 0};
3154                 unsigned long recent_scanned[2] = {0, 0};
3155
3156                 for_each_online_node(nid)
3157                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3158                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3159
3160                                 recent_rotated[0] +=
3161                                         mz->reclaim_stat.recent_rotated[0];
3162                                 recent_rotated[1] +=
3163                                         mz->reclaim_stat.recent_rotated[1];
3164                                 recent_scanned[0] +=
3165                                         mz->reclaim_stat.recent_scanned[0];
3166                                 recent_scanned[1] +=
3167                                         mz->reclaim_stat.recent_scanned[1];
3168                         }
3169                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3170                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3171                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3172                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3173         }
3174 #endif
3175
3176         return 0;
3177 }
3178
3179 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3180 {
3181         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3182
3183         return get_swappiness(memcg);
3184 }
3185
3186 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3187                                        u64 val)
3188 {
3189         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3190         struct mem_cgroup *parent;
3191
3192         if (val > 100)
3193                 return -EINVAL;
3194
3195         if (cgrp->parent == NULL)
3196                 return -EINVAL;
3197
3198         parent = mem_cgroup_from_cont(cgrp->parent);
3199
3200         cgroup_lock();
3201
3202         /* If under hierarchy, only empty-root can set this value */
3203         if ((parent->use_hierarchy) ||
3204             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3205                 cgroup_unlock();
3206                 return -EINVAL;
3207         }
3208
3209         spin_lock(&memcg->reclaim_param_lock);
3210         memcg->swappiness = val;
3211         spin_unlock(&memcg->reclaim_param_lock);
3212
3213         cgroup_unlock();
3214
3215         return 0;
3216 }
3217
3218 static bool mem_cgroup_threshold_check(struct mem_cgroup *mem)
3219 {
3220         bool ret = false;
3221         s64 val;
3222
3223         val = this_cpu_read(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS]);
3224         if (unlikely(val < 0)) {
3225                 this_cpu_write(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS],
3226                                 THRESHOLDS_EVENTS_THRESH);
3227                 ret = true;
3228         }
3229         return ret;
3230 }
3231
3232 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3233 {
3234         struct mem_cgroup_threshold_ary *t;
3235         u64 usage;
3236         int i;
3237
3238         rcu_read_lock();
3239         if (!swap)
3240                 t = rcu_dereference(memcg->thresholds);
3241         else
3242                 t = rcu_dereference(memcg->memsw_thresholds);
3243
3244         if (!t)
3245                 goto unlock;
3246
3247         usage = mem_cgroup_usage(memcg, swap);
3248
3249         /*
3250          * current_threshold points to threshold just below usage.
3251          * If it's not true, a threshold was crossed after last
3252          * call of __mem_cgroup_threshold().
3253          */
3254         i = atomic_read(&t->current_threshold);
3255
3256         /*
3257          * Iterate backward over array of thresholds starting from
3258          * current_threshold and check if a threshold is crossed.
3259          * If none of thresholds below usage is crossed, we read
3260          * only one element of the array here.
3261          */
3262         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3263                 eventfd_signal(t->entries[i].eventfd, 1);
3264
3265         /* i = current_threshold + 1 */
3266         i++;
3267
3268         /*
3269          * Iterate forward over array of thresholds starting from
3270          * current_threshold+1 and check if a threshold is crossed.
3271          * If none of thresholds above usage is crossed, we read
3272          * only one element of the array here.
3273          */
3274         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3275                 eventfd_signal(t->entries[i].eventfd, 1);
3276
3277         /* Update current_threshold */
3278         atomic_set(&t->current_threshold, i - 1);
3279 unlock:
3280         rcu_read_unlock();
3281 }
3282
3283 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3284 {
3285         __mem_cgroup_threshold(memcg, false);
3286         if (do_swap_account)
3287                 __mem_cgroup_threshold(memcg, true);
3288 }
3289
3290 static int compare_thresholds(const void *a, const void *b)
3291 {
3292         const struct mem_cgroup_threshold *_a = a;
3293         const struct mem_cgroup_threshold *_b = b;
3294
3295         return _a->threshold - _b->threshold;
3296 }
3297
3298 static int mem_cgroup_register_event(struct cgroup *cgrp, struct cftype *cft,
3299                 struct eventfd_ctx *eventfd, const char *args)
3300 {
3301         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3302         struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3303         int type = MEMFILE_TYPE(cft->private);
3304         u64 threshold, usage;
3305         int size;
3306         int i, ret;
3307
3308         ret = res_counter_memparse_write_strategy(args, &threshold);
3309         if (ret)
3310                 return ret;
3311
3312         mutex_lock(&memcg->thresholds_lock);
3313         if (type == _MEM)
3314                 thresholds = memcg->thresholds;
3315         else if (type == _MEMSWAP)
3316                 thresholds = memcg->memsw_thresholds;
3317         else
3318                 BUG();
3319
3320         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3321
3322         /* Check if a threshold crossed before adding a new one */
3323         if (thresholds)
3324                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3325
3326         if (thresholds)
3327                 size = thresholds->size + 1;
3328         else
3329                 size = 1;
3330
3331         /* Allocate memory for new array of thresholds */
3332         thresholds_new = kmalloc(sizeof(*thresholds_new) +
3333                         size * sizeof(struct mem_cgroup_threshold),
3334                         GFP_KERNEL);
3335         if (!thresholds_new) {
3336                 ret = -ENOMEM;
3337                 goto unlock;
3338         }
3339         thresholds_new->size = size;
3340
3341         /* Copy thresholds (if any) to new array */
3342         if (thresholds)
3343                 memcpy(thresholds_new->entries, thresholds->entries,
3344                                 thresholds->size *
3345                                 sizeof(struct mem_cgroup_threshold));
3346         /* Add new threshold */
3347         thresholds_new->entries[size - 1].eventfd = eventfd;
3348         thresholds_new->entries[size - 1].threshold = threshold;
3349
3350         /* Sort thresholds. Registering of new threshold isn't time-critical */
3351         sort(thresholds_new->entries, size,
3352                         sizeof(struct mem_cgroup_threshold),
3353                         compare_thresholds, NULL);
3354
3355         /* Find current threshold */
3356         atomic_set(&thresholds_new->current_threshold, -1);
3357         for (i = 0; i < size; i++) {
3358                 if (thresholds_new->entries[i].threshold < usage) {
3359                         /*
3360                          * thresholds_new->current_threshold will not be used
3361                          * until rcu_assign_pointer(), so it's safe to increment
3362                          * it here.
3363                          */
3364                         atomic_inc(&thresholds_new->current_threshold);
3365                 }
3366         }
3367
3368         /*
3369          * We need to increment refcnt to be sure that all thresholds
3370          * will be unregistered before calling __mem_cgroup_free()
3371          */
3372         mem_cgroup_get(memcg);
3373
3374         if (type == _MEM)
3375                 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3376         else
3377                 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3378
3379         /* To be sure that nobody uses thresholds before freeing it */
3380         synchronize_rcu();
3381
3382         kfree(thresholds);
3383 unlock:
3384         mutex_unlock(&memcg->thresholds_lock);
3385
3386         return ret;
3387 }
3388
3389 static int mem_cgroup_unregister_event(struct cgroup *cgrp, struct cftype *cft,
3390                 struct eventfd_ctx *eventfd)
3391 {
3392         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3393         struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3394         int type = MEMFILE_TYPE(cft->private);
3395         u64 usage;
3396         int size = 0;
3397         int i, j, ret;
3398
3399         mutex_lock(&memcg->thresholds_lock);
3400         if (type == _MEM)
3401                 thresholds = memcg->thresholds;
3402         else if (type == _MEMSWAP)
3403                 thresholds = memcg->memsw_thresholds;
3404         else
3405                 BUG();
3406
3407         /*
3408          * Something went wrong if we trying to unregister a threshold
3409          * if we don't have thresholds
3410          */
3411         BUG_ON(!thresholds);
3412
3413         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3414
3415         /* Check if a threshold crossed before removing */
3416         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3417
3418         /* Calculate new number of threshold */
3419         for (i = 0; i < thresholds->size; i++) {
3420                 if (thresholds->entries[i].eventfd != eventfd)
3421                         size++;
3422         }
3423
3424         /* Set thresholds array to NULL if we don't have thresholds */
3425         if (!size) {
3426                 thresholds_new = NULL;
3427                 goto assign;
3428         }
3429
3430         /* Allocate memory for new array of thresholds */
3431         thresholds_new = kmalloc(sizeof(*thresholds_new) +
3432                         size * sizeof(struct mem_cgroup_threshold),
3433                         GFP_KERNEL);
3434         if (!thresholds_new) {
3435                 ret = -ENOMEM;
3436                 goto unlock;
3437         }
3438         thresholds_new->size = size;
3439
3440         /* Copy thresholds and find current threshold */
3441         atomic_set(&thresholds_new->current_threshold, -1);
3442         for (i = 0, j = 0; i < thresholds->size; i++) {
3443                 if (thresholds->entries[i].eventfd == eventfd)
3444                         continue;
3445
3446                 thresholds_new->entries[j] = thresholds->entries[i];
3447                 if (thresholds_new->entries[j].threshold < usage) {
3448                         /*
3449                          * thresholds_new->current_threshold will not be used
3450                          * until rcu_assign_pointer(), so it's safe to increment
3451                          * it here.
3452                          */
3453                         atomic_inc(&thresholds_new->current_threshold);
3454                 }
3455                 j++;
3456         }
3457
3458 assign:
3459         if (type == _MEM)
3460                 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3461         else
3462                 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3463
3464         /* To be sure that nobody uses thresholds before freeing it */
3465         synchronize_rcu();
3466
3467         for (i = 0; i < thresholds->size - size; i++)
3468                 mem_cgroup_put(memcg);
3469
3470         kfree(thresholds);
3471 unlock:
3472         mutex_unlock(&memcg->thresholds_lock);
3473
3474         return ret;
3475 }
3476
3477 static struct cftype mem_cgroup_files[] = {
3478         {
3479                 .name = "usage_in_bytes",
3480                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3481                 .read_u64 = mem_cgroup_read,
3482                 .register_event = mem_cgroup_register_event,
3483                 .unregister_event = mem_cgroup_unregister_event,
3484         },
3485         {
3486                 .name = "max_usage_in_bytes",
3487                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3488                 .trigger = mem_cgroup_reset,
3489                 .read_u64 = mem_cgroup_read,
3490         },
3491         {
3492                 .name = "limit_in_bytes",
3493                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3494                 .write_string = mem_cgroup_write,
3495                 .read_u64 = mem_cgroup_read,
3496         },
3497         {
3498                 .name = "soft_limit_in_bytes",
3499                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3500                 .write_string = mem_cgroup_write,
3501                 .read_u64 = mem_cgroup_read,
3502         },
3503         {
3504                 .name = "failcnt",
3505                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3506                 .trigger = mem_cgroup_reset,
3507                 .read_u64 = mem_cgroup_read,
3508         },
3509         {
3510                 .name = "stat",
3511                 .read_map = mem_control_stat_show,
3512         },
3513         {
3514                 .name = "force_empty",
3515                 .trigger = mem_cgroup_force_empty_write,
3516         },
3517         {
3518                 .name = "use_hierarchy",
3519                 .write_u64 = mem_cgroup_hierarchy_write,
3520                 .read_u64 = mem_cgroup_hierarchy_read,
3521         },
3522         {
3523                 .name = "swappiness",
3524                 .read_u64 = mem_cgroup_swappiness_read,
3525                 .write_u64 = mem_cgroup_swappiness_write,
3526         },
3527         {
3528                 .name = "move_charge_at_immigrate",
3529                 .read_u64 = mem_cgroup_move_charge_read,
3530                 .write_u64 = mem_cgroup_move_charge_write,
3531         },
3532 };
3533
3534 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3535 static struct cftype memsw_cgroup_files[] = {
3536         {
3537                 .name = "memsw.usage_in_bytes",
3538                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3539                 .read_u64 = mem_cgroup_read,
3540                 .register_event = mem_cgroup_register_event,
3541                 .unregister_event = mem_cgroup_unregister_event,
3542         },
3543         {
3544                 .name = "memsw.max_usage_in_bytes",
3545                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3546                 .trigger = mem_cgroup_reset,
3547                 .read_u64 = mem_cgroup_read,
3548         },
3549         {
3550                 .name = "memsw.limit_in_bytes",
3551                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3552                 .write_string = mem_cgroup_write,
3553                 .read_u64 = mem_cgroup_read,
3554         },
3555         {
3556                 .name = "memsw.failcnt",
3557                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3558                 .trigger = mem_cgroup_reset,
3559                 .read_u64 = mem_cgroup_read,
3560         },
3561 };
3562
3563 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3564 {
3565         if (!do_swap_account)
3566                 return 0;
3567         return cgroup_add_files(cont, ss, memsw_cgroup_files,
3568                                 ARRAY_SIZE(memsw_cgroup_files));
3569 };
3570 #else
3571 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3572 {
3573         return 0;
3574 }
3575 #endif
3576
3577 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3578 {
3579         struct mem_cgroup_per_node *pn;
3580         struct mem_cgroup_per_zone *mz;
3581         enum lru_list l;
3582         int zone, tmp = node;
3583         /*
3584          * This routine is called against possible nodes.
3585          * But it's BUG to call kmalloc() against offline node.
3586          *
3587          * TODO: this routine can waste much memory for nodes which will
3588          *       never be onlined. It's better to use memory hotplug callback
3589          *       function.
3590          */
3591         if (!node_state(node, N_NORMAL_MEMORY))
3592                 tmp = -1;
3593         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3594         if (!pn)
3595                 return 1;
3596
3597         mem->info.nodeinfo[node] = pn;
3598         memset(pn, 0, sizeof(*pn));
3599
3600         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3601                 mz = &pn->zoneinfo[zone];
3602                 for_each_lru(l)
3603                         INIT_LIST_HEAD(&mz->lists[l]);
3604                 mz->usage_in_excess = 0;
3605                 mz->on_tree = false;
3606                 mz->mem = mem;
3607         }
3608         return 0;
3609 }
3610
3611 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3612 {
3613         kfree(mem->info.nodeinfo[node]);
3614 }
3615
3616 static struct mem_cgroup *mem_cgroup_alloc(void)
3617 {
3618         struct mem_cgroup *mem;
3619         int size = sizeof(struct mem_cgroup);
3620
3621         /* Can be very big if MAX_NUMNODES is very big */
3622         if (size < PAGE_SIZE)
3623                 mem = kmalloc(size, GFP_KERNEL);
3624         else
3625                 mem = vmalloc(size);
3626
3627         if (mem)
3628                 memset(mem, 0, size);
3629         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3630         if (!mem->stat) {
3631                 if (size < PAGE_SIZE)
3632                         kfree(mem);
3633                 else
3634                         vfree(mem);
3635                 mem = NULL;
3636         }
3637         return mem;
3638 }
3639
3640 /*
3641  * At destroying mem_cgroup, references from swap_cgroup can remain.
3642  * (scanning all at force_empty is too costly...)
3643  *
3644  * Instead of clearing all references at force_empty, we remember
3645  * the number of reference from swap_cgroup and free mem_cgroup when
3646  * it goes down to 0.
3647  *
3648  * Removal of cgroup itself succeeds regardless of refs from swap.
3649  */
3650
3651 static void __mem_cgroup_free(struct mem_cgroup *mem)
3652 {
3653         int node;
3654
3655         mem_cgroup_remove_from_trees(mem);
3656         free_css_id(&mem_cgroup_subsys, &mem->css);
3657
3658         for_each_node_state(node, N_POSSIBLE)
3659                 free_mem_cgroup_per_zone_info(mem, node);
3660
3661         free_percpu(mem->stat);
3662         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
3663                 kfree(mem);
3664         else
3665                 vfree(mem);
3666 }
3667
3668 static void mem_cgroup_get(struct mem_cgroup *mem)
3669 {
3670         atomic_inc(&mem->refcnt);
3671 }
3672
3673 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
3674 {
3675         if (atomic_sub_and_test(count, &mem->refcnt)) {
3676                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3677                 __mem_cgroup_free(mem);
3678                 if (parent)
3679                         mem_cgroup_put(parent);
3680         }
3681 }
3682
3683 static void mem_cgroup_put(struct mem_cgroup *mem)
3684 {
3685         __mem_cgroup_put(mem, 1);
3686 }
3687
3688 /*
3689  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3690  */
3691 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3692 {
3693         if (!mem->res.parent)
3694                 return NULL;
3695         return mem_cgroup_from_res_counter(mem->res.parent, res);
3696 }
3697
3698 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3699 static void __init enable_swap_cgroup(void)
3700 {
3701         if (!mem_cgroup_disabled() && really_do_swap_account)
3702                 do_swap_account = 1;
3703 }
3704 #else
3705 static void __init enable_swap_cgroup(void)
3706 {
3707 }
3708 #endif
3709
3710 static int mem_cgroup_soft_limit_tree_init(void)
3711 {
3712         struct mem_cgroup_tree_per_node *rtpn;
3713         struct mem_cgroup_tree_per_zone *rtpz;
3714         int tmp, node, zone;
3715
3716         for_each_node_state(node, N_POSSIBLE) {
3717                 tmp = node;
3718                 if (!node_state(node, N_NORMAL_MEMORY))
3719                         tmp = -1;
3720                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3721                 if (!rtpn)
3722                         return 1;
3723
3724                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3725
3726                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3727                         rtpz = &rtpn->rb_tree_per_zone[zone];
3728                         rtpz->rb_root = RB_ROOT;
3729                         spin_lock_init(&rtpz->lock);
3730                 }
3731         }
3732         return 0;
3733 }
3734
3735 static struct cgroup_subsys_state * __ref
3736 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3737 {
3738         struct mem_cgroup *mem, *parent;
3739         long error = -ENOMEM;
3740         int node;
3741
3742         mem = mem_cgroup_alloc();
3743         if (!mem)
3744                 return ERR_PTR(error);
3745
3746         for_each_node_state(node, N_POSSIBLE)
3747                 if (alloc_mem_cgroup_per_zone_info(mem, node))
3748                         goto free_out;
3749
3750         /* root ? */
3751         if (cont->parent == NULL) {
3752                 int cpu;
3753                 enable_swap_cgroup();
3754                 parent = NULL;
3755                 root_mem_cgroup = mem;
3756                 if (mem_cgroup_soft_limit_tree_init())
3757                         goto free_out;
3758                 for_each_possible_cpu(cpu) {
3759                         struct memcg_stock_pcp *stock =
3760                                                 &per_cpu(memcg_stock, cpu);
3761                         INIT_WORK(&stock->work, drain_local_stock);
3762                 }
3763                 hotcpu_notifier(memcg_stock_cpu_callback, 0);
3764         } else {
3765                 parent = mem_cgroup_from_cont(cont->parent);
3766                 mem->use_hierarchy = parent->use_hierarchy;
3767         }
3768
3769         if (parent && parent->use_hierarchy) {
3770                 res_counter_init(&mem->res, &parent->res);
3771                 res_counter_init(&mem->memsw, &parent->memsw);
3772                 /*
3773                  * We increment refcnt of the parent to ensure that we can
3774                  * safely access it on res_counter_charge/uncharge.
3775                  * This refcnt will be decremented when freeing this
3776                  * mem_cgroup(see mem_cgroup_put).
3777                  */
3778                 mem_cgroup_get(parent);
3779         } else {
3780                 res_counter_init(&mem->res, NULL);
3781                 res_counter_init(&mem->memsw, NULL);
3782         }
3783         mem->last_scanned_child = 0;
3784         spin_lock_init(&mem->reclaim_param_lock);
3785
3786         if (parent)
3787                 mem->swappiness = get_swappiness(parent);
3788         atomic_set(&mem->refcnt, 1);
3789         mem->move_charge_at_immigrate = 0;
3790         mutex_init(&mem->thresholds_lock);
3791         return &mem->css;
3792 free_out:
3793         __mem_cgroup_free(mem);
3794         root_mem_cgroup = NULL;
3795         return ERR_PTR(error);
3796 }
3797
3798 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
3799                                         struct cgroup *cont)
3800 {
3801         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3802
3803         return mem_cgroup_force_empty(mem, false);
3804 }
3805
3806 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
3807                                 struct cgroup *cont)
3808 {
3809         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3810
3811         mem_cgroup_put(mem);
3812 }
3813
3814 static int mem_cgroup_populate(struct cgroup_subsys *ss,
3815                                 struct cgroup *cont)
3816 {
3817         int ret;
3818
3819         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
3820                                 ARRAY_SIZE(mem_cgroup_files));
3821
3822         if (!ret)
3823                 ret = register_memsw_files(cont, ss);
3824         return ret;
3825 }
3826
3827 #ifdef CONFIG_MMU
3828 /* Handlers for move charge at task migration. */
3829 #define PRECHARGE_COUNT_AT_ONCE 256
3830 static int mem_cgroup_do_precharge(unsigned long count)
3831 {
3832         int ret = 0;
3833         int batch_count = PRECHARGE_COUNT_AT_ONCE;
3834         struct mem_cgroup *mem = mc.to;
3835
3836         if (mem_cgroup_is_root(mem)) {
3837                 mc.precharge += count;
3838                 /* we don't need css_get for root */
3839                 return ret;
3840         }
3841         /* try to charge at once */
3842         if (count > 1) {
3843                 struct res_counter *dummy;
3844                 /*
3845                  * "mem" cannot be under rmdir() because we've already checked
3846                  * by cgroup_lock_live_cgroup() that it is not removed and we
3847                  * are still under the same cgroup_mutex. So we can postpone
3848                  * css_get().
3849                  */
3850                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
3851                         goto one_by_one;
3852                 if (do_swap_account && res_counter_charge(&mem->memsw,
3853                                                 PAGE_SIZE * count, &dummy)) {
3854                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
3855                         goto one_by_one;
3856                 }
3857                 mc.precharge += count;
3858                 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
3859                 WARN_ON_ONCE(count > INT_MAX);
3860                 __css_get(&mem->css, (int)count);
3861                 return ret;
3862         }
3863 one_by_one:
3864         /* fall back to one by one charge */
3865         while (count--) {
3866                 if (signal_pending(current)) {
3867                         ret = -EINTR;
3868                         break;
3869                 }
3870                 if (!batch_count--) {
3871                         batch_count = PRECHARGE_COUNT_AT_ONCE;
3872                         cond_resched();
3873                 }
3874                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
3875                 if (ret || !mem)
3876                         /* mem_cgroup_clear_mc() will do uncharge later */
3877                         return -ENOMEM;
3878                 mc.precharge++;
3879         }
3880         return ret;
3881 }
3882 #else   /* !CONFIG_MMU */
3883 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
3884                                 struct cgroup *cgroup,
3885                                 struct task_struct *p,
3886                                 bool threadgroup)
3887 {
3888         return 0;
3889 }
3890 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
3891                                 struct cgroup *cgroup,
3892                                 struct task_struct *p,
3893                                 bool threadgroup)
3894 {
3895 }
3896 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
3897                                 struct cgroup *cont,
3898                                 struct cgroup *old_cont,
3899                                 struct task_struct *p,
3900                                 bool threadgroup)
3901 {
3902 }
3903 #endif
3904
3905 /**
3906  * is_target_pte_for_mc - check a pte whether it is valid for move charge
3907  * @vma: the vma the pte to be checked belongs
3908  * @addr: the address corresponding to the pte to be checked
3909  * @ptent: the pte to be checked
3910  * @target: the pointer the target page or swap ent will be stored(can be NULL)
3911  *
3912  * Returns
3913  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
3914  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
3915  *     move charge. if @target is not NULL, the page is stored in target->page
3916  *     with extra refcnt got(Callers should handle it).
3917  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
3918  *     target for charge migration. if @target is not NULL, the entry is stored
3919  *     in target->ent.
3920  *
3921  * Called with pte lock held.
3922  */
3923 union mc_target {
3924         struct page     *page;
3925         swp_entry_t     ent;
3926 };
3927
3928 enum mc_target_type {
3929         MC_TARGET_NONE, /* not used */
3930         MC_TARGET_PAGE,
3931         MC_TARGET_SWAP,
3932 };
3933
3934 static int is_target_pte_for_mc(struct vm_area_struct *vma,
3935                 unsigned long addr, pte_t ptent, union mc_target *target)
3936 {
3937         struct page *page = NULL;
3938         struct page_cgroup *pc;
3939         int ret = 0;
3940         swp_entry_t ent = { .val = 0 };
3941         int usage_count = 0;
3942         bool move_anon = test_bit(MOVE_CHARGE_TYPE_ANON,
3943                                         &mc.to->move_charge_at_immigrate);
3944
3945         if (!pte_present(ptent)) {
3946                 /* TODO: handle swap of shmes/tmpfs */
3947                 if (pte_none(ptent) || pte_file(ptent))
3948                         return 0;
3949                 else if (is_swap_pte(ptent)) {
3950                         ent = pte_to_swp_entry(ptent);
3951                         if (!move_anon || non_swap_entry(ent))
3952                                 return 0;
3953                         usage_count = mem_cgroup_count_swap_user(ent, &page);
3954                 }
3955         } else {
3956                 page = vm_normal_page(vma, addr, ptent);
3957                 if (!page || !page_mapped(page))
3958                         return 0;
3959                 /*
3960                  * TODO: We don't move charges of file(including shmem/tmpfs)
3961                  * pages for now.
3962                  */
3963                 if (!move_anon || !PageAnon(page))
3964                         return 0;
3965                 if (!get_page_unless_zero(page))
3966                         return 0;
3967                 usage_count = page_mapcount(page);
3968         }
3969         if (usage_count > 1) {
3970                 /*
3971                  * TODO: We don't move charges of shared(used by multiple
3972                  * processes) pages for now.
3973                  */
3974                 if (page)
3975                         put_page(page);
3976                 return 0;
3977         }
3978         if (page) {
3979                 pc = lookup_page_cgroup(page);
3980                 /*
3981                  * Do only loose check w/o page_cgroup lock.
3982                  * mem_cgroup_move_account() checks the pc is valid or not under
3983                  * the lock.
3984                  */
3985                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
3986                         ret = MC_TARGET_PAGE;
3987                         if (target)
3988                                 target->page = page;
3989                 }
3990                 if (!ret || !target)
3991                         put_page(page);
3992         }
3993         /* throught */
3994         if (ent.val && do_swap_account && !ret &&
3995                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
3996                 ret = MC_TARGET_SWAP;
3997                 if (target)
3998                         target->ent = ent;
3999         }
4000         return ret;
4001 }
4002
4003 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4004                                         unsigned long addr, unsigned long end,
4005                                         struct mm_walk *walk)
4006 {
4007         struct vm_area_struct *vma = walk->private;
4008         pte_t *pte;
4009         spinlock_t *ptl;
4010
4011         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4012         for (; addr != end; pte++, addr += PAGE_SIZE)
4013                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4014                         mc.precharge++; /* increment precharge temporarily */
4015         pte_unmap_unlock(pte - 1, ptl);
4016         cond_resched();
4017
4018         return 0;
4019 }
4020
4021 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4022 {
4023         unsigned long precharge;
4024         struct vm_area_struct *vma;
4025
4026         down_read(&mm->mmap_sem);
4027         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4028                 struct mm_walk mem_cgroup_count_precharge_walk = {
4029                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4030                         .mm = mm,
4031                         .private = vma,
4032                 };
4033                 if (is_vm_hugetlb_page(vma))
4034                         continue;
4035                 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4036                 if (vma->vm_flags & VM_SHARED)
4037                         continue;
4038                 walk_page_range(vma->vm_start, vma->vm_end,
4039                                         &mem_cgroup_count_precharge_walk);
4040         }
4041         up_read(&mm->mmap_sem);
4042
4043         precharge = mc.precharge;
4044         mc.precharge = 0;
4045
4046         return precharge;
4047 }
4048
4049 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4050 {
4051         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4052 }
4053
4054 static void mem_cgroup_clear_mc(void)
4055 {
4056         /* we must uncharge all the leftover precharges from mc.to */
4057         if (mc.precharge) {
4058                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4059                 mc.precharge = 0;
4060         }
4061         /*
4062          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4063          * we must uncharge here.
4064          */
4065         if (mc.moved_charge) {
4066                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4067                 mc.moved_charge = 0;
4068         }
4069         /* we must fixup refcnts and charges */
4070         if (mc.moved_swap) {
4071                 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4072                 /* uncharge swap account from the old cgroup */
4073                 if (!mem_cgroup_is_root(mc.from))
4074                         res_counter_uncharge(&mc.from->memsw,
4075                                                 PAGE_SIZE * mc.moved_swap);
4076                 __mem_cgroup_put(mc.from, mc.moved_swap);
4077
4078                 if (!mem_cgroup_is_root(mc.to)) {
4079                         /*
4080                          * we charged both to->res and to->memsw, so we should
4081                          * uncharge to->res.
4082                          */
4083                         res_counter_uncharge(&mc.to->res,
4084                                                 PAGE_SIZE * mc.moved_swap);
4085                         VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4086                         __css_put(&mc.to->css, mc.moved_swap);
4087                 }
4088                 /* we've already done mem_cgroup_get(mc.to) */
4089
4090                 mc.moved_swap = 0;
4091         }
4092         mc.from = NULL;
4093         mc.to = NULL;
4094         mc.moving_task = NULL;
4095         wake_up_all(&mc.waitq);
4096 }
4097
4098 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4099                                 struct cgroup *cgroup,
4100                                 struct task_struct *p,
4101                                 bool threadgroup)
4102 {
4103         int ret = 0;
4104         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4105
4106         if (mem->move_charge_at_immigrate) {
4107                 struct mm_struct *mm;
4108                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4109
4110                 VM_BUG_ON(from == mem);
4111
4112                 mm = get_task_mm(p);
4113                 if (!mm)
4114                         return 0;
4115                 /* We move charges only when we move a owner of the mm */
4116                 if (mm->owner == p) {
4117                         VM_BUG_ON(mc.from);
4118                         VM_BUG_ON(mc.to);
4119                         VM_BUG_ON(mc.precharge);
4120                         VM_BUG_ON(mc.moved_charge);
4121                         VM_BUG_ON(mc.moved_swap);
4122                         VM_BUG_ON(mc.moving_task);
4123                         mc.from = from;
4124                         mc.to = mem;
4125                         mc.precharge = 0;
4126                         mc.moved_charge = 0;
4127                         mc.moved_swap = 0;
4128                         mc.moving_task = current;
4129
4130                         ret = mem_cgroup_precharge_mc(mm);
4131                         if (ret)
4132                                 mem_cgroup_clear_mc();
4133                 }
4134                 mmput(mm);
4135         }
4136         return ret;
4137 }
4138
4139 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4140                                 struct cgroup *cgroup,
4141                                 struct task_struct *p,
4142                                 bool threadgroup)
4143 {
4144         mem_cgroup_clear_mc();
4145 }
4146
4147 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4148                                 unsigned long addr, unsigned long end,
4149                                 struct mm_walk *walk)
4150 {
4151         int ret = 0;
4152         struct vm_area_struct *vma = walk->private;
4153         pte_t *pte;
4154         spinlock_t *ptl;
4155
4156 retry:
4157         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4158         for (; addr != end; addr += PAGE_SIZE) {
4159                 pte_t ptent = *(pte++);
4160                 union mc_target target;
4161                 int type;
4162                 struct page *page;
4163                 struct page_cgroup *pc;
4164                 swp_entry_t ent;
4165
4166                 if (!mc.precharge)
4167                         break;
4168
4169                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4170                 switch (type) {
4171                 case MC_TARGET_PAGE:
4172                         page = target.page;
4173                         if (isolate_lru_page(page))
4174                                 goto put;
4175                         pc = lookup_page_cgroup(page);
4176                         if (!mem_cgroup_move_account(pc,
4177                                                 mc.from, mc.to, false)) {
4178                                 mc.precharge--;
4179                                 /* we uncharge from mc.from later. */
4180                                 mc.moved_charge++;
4181                         }
4182                         putback_lru_page(page);
4183 put:                    /* is_target_pte_for_mc() gets the page */
4184                         put_page(page);
4185                         break;
4186                 case MC_TARGET_SWAP:
4187                         ent = target.ent;
4188                         if (!mem_cgroup_move_swap_account(ent,
4189                                                 mc.from, mc.to, false)) {
4190                                 mc.precharge--;
4191                                 /* we fixup refcnts and charges later. */
4192                                 mc.moved_swap++;
4193                         }
4194                         break;
4195                 default:
4196                         break;
4197                 }
4198         }
4199         pte_unmap_unlock(pte - 1, ptl);
4200         cond_resched();
4201
4202         if (addr != end) {
4203                 /*
4204                  * We have consumed all precharges we got in can_attach().
4205                  * We try charge one by one, but don't do any additional
4206                  * charges to mc.to if we have failed in charge once in attach()
4207                  * phase.
4208                  */
4209                 ret = mem_cgroup_do_precharge(1);
4210                 if (!ret)
4211                         goto retry;
4212         }
4213
4214         return ret;
4215 }
4216
4217 static void mem_cgroup_move_charge(struct mm_struct *mm)
4218 {
4219         struct vm_area_struct *vma;
4220
4221         lru_add_drain_all();
4222         down_read(&mm->mmap_sem);
4223         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4224                 int ret;
4225                 struct mm_walk mem_cgroup_move_charge_walk = {
4226                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4227                         .mm = mm,
4228                         .private = vma,
4229                 };
4230                 if (is_vm_hugetlb_page(vma))
4231                         continue;
4232                 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4233                 if (vma->vm_flags & VM_SHARED)
4234                         continue;
4235                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4236                                                 &mem_cgroup_move_charge_walk);
4237                 if (ret)
4238                         /*
4239                          * means we have consumed all precharges and failed in
4240                          * doing additional charge. Just abandon here.
4241                          */
4242                         break;
4243         }
4244         up_read(&mm->mmap_sem);
4245 }
4246
4247 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4248                                 struct cgroup *cont,
4249                                 struct cgroup *old_cont,
4250                                 struct task_struct *p,
4251                                 bool threadgroup)
4252 {
4253         struct mm_struct *mm;
4254
4255         if (!mc.to)
4256                 /* no need to move charge */
4257                 return;
4258
4259         mm = get_task_mm(p);
4260         if (mm) {
4261                 mem_cgroup_move_charge(mm);
4262                 mmput(mm);
4263         }
4264         mem_cgroup_clear_mc();
4265 }
4266
4267 struct cgroup_subsys mem_cgroup_subsys = {
4268         .name = "memory",
4269         .subsys_id = mem_cgroup_subsys_id,
4270         .create = mem_cgroup_create,
4271         .pre_destroy = mem_cgroup_pre_destroy,
4272         .destroy = mem_cgroup_destroy,
4273         .populate = mem_cgroup_populate,
4274         .can_attach = mem_cgroup_can_attach,
4275         .cancel_attach = mem_cgroup_cancel_attach,
4276         .attach = mem_cgroup_move_task,
4277         .early_init = 0,
4278         .use_id = 1,
4279 };
4280
4281 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4282
4283 static int __init disable_swap_account(char *s)
4284 {
4285         really_do_swap_account = 0;
4286         return 1;
4287 }
4288 __setup("noswapaccount", disable_swap_account);
4289 #endif