thp: transhuge-memcg: commit tail pages at charge
[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 <linux/oom.h>
51 #include "internal.h"
52
53 #include <asm/uaccess.h>
54
55 #include <trace/events/vmscan.h>
56
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES      5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
60
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
64
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata = 1;
68 #else
69 static int really_do_swap_account __initdata = 0;
70 #endif
71
72 #else
73 #define do_swap_account         (0)
74 #endif
75
76 /*
77  * Per memcg event counter is incremented at every pagein/pageout. This counter
78  * is used for trigger some periodic events. This is straightforward and better
79  * than using jiffies etc. to handle periodic memcg event.
80  *
81  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
82  */
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
85
86 /*
87  * Statistics for memory cgroup.
88  */
89 enum mem_cgroup_stat_index {
90         /*
91          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92          */
93         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
94         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
95         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
96         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
97         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
98         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
99         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
100         /* incremented at every  pagein/pageout */
101         MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
102         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
103
104         MEM_CGROUP_STAT_NSTATS,
105 };
106
107 struct mem_cgroup_stat_cpu {
108         s64 count[MEM_CGROUP_STAT_NSTATS];
109 };
110
111 /*
112  * per-zone information in memory controller.
113  */
114 struct mem_cgroup_per_zone {
115         /*
116          * spin_lock to protect the per cgroup LRU
117          */
118         struct list_head        lists[NR_LRU_LISTS];
119         unsigned long           count[NR_LRU_LISTS];
120
121         struct zone_reclaim_stat reclaim_stat;
122         struct rb_node          tree_node;      /* RB tree node */
123         unsigned long long      usage_in_excess;/* Set to the value by which */
124                                                 /* the soft limit is exceeded*/
125         bool                    on_tree;
126         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
127                                                 /* use container_of        */
128 };
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
131
132 struct mem_cgroup_per_node {
133         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
134 };
135
136 struct mem_cgroup_lru_info {
137         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
138 };
139
140 /*
141  * Cgroups above their limits are maintained in a RB-Tree, independent of
142  * their hierarchy representation
143  */
144
145 struct mem_cgroup_tree_per_zone {
146         struct rb_root rb_root;
147         spinlock_t lock;
148 };
149
150 struct mem_cgroup_tree_per_node {
151         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
152 };
153
154 struct mem_cgroup_tree {
155         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
156 };
157
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
159
160 struct mem_cgroup_threshold {
161         struct eventfd_ctx *eventfd;
162         u64 threshold;
163 };
164
165 /* For threshold */
166 struct mem_cgroup_threshold_ary {
167         /* An array index points to threshold just below usage. */
168         int current_threshold;
169         /* Size of entries[] */
170         unsigned int size;
171         /* Array of thresholds */
172         struct mem_cgroup_threshold entries[0];
173 };
174
175 struct mem_cgroup_thresholds {
176         /* Primary thresholds array */
177         struct mem_cgroup_threshold_ary *primary;
178         /*
179          * Spare threshold array.
180          * This is needed to make mem_cgroup_unregister_event() "never fail".
181          * It must be able to store at least primary->size - 1 entries.
182          */
183         struct mem_cgroup_threshold_ary *spare;
184 };
185
186 /* for OOM */
187 struct mem_cgroup_eventfd_list {
188         struct list_head list;
189         struct eventfd_ctx *eventfd;
190 };
191
192 static void mem_cgroup_threshold(struct mem_cgroup *mem);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
194
195 /*
196  * The memory controller data structure. The memory controller controls both
197  * page cache and RSS per cgroup. We would eventually like to provide
198  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199  * to help the administrator determine what knobs to tune.
200  *
201  * TODO: Add a water mark for the memory controller. Reclaim will begin when
202  * we hit the water mark. May be even add a low water mark, such that
203  * no reclaim occurs from a cgroup at it's low water mark, this is
204  * a feature that will be implemented much later in the future.
205  */
206 struct mem_cgroup {
207         struct cgroup_subsys_state css;
208         /*
209          * the counter to account for memory usage
210          */
211         struct res_counter res;
212         /*
213          * the counter to account for mem+swap usage.
214          */
215         struct res_counter memsw;
216         /*
217          * Per cgroup active and inactive list, similar to the
218          * per zone LRU lists.
219          */
220         struct mem_cgroup_lru_info info;
221
222         /*
223           protect against reclaim related member.
224         */
225         spinlock_t reclaim_param_lock;
226
227         /*
228          * While reclaiming in a hierarchy, we cache the last child we
229          * reclaimed from.
230          */
231         int last_scanned_child;
232         /*
233          * Should the accounting and control be hierarchical, per subtree?
234          */
235         bool use_hierarchy;
236         atomic_t        oom_lock;
237         atomic_t        refcnt;
238
239         unsigned int    swappiness;
240         /* OOM-Killer disable */
241         int             oom_kill_disable;
242
243         /* set when res.limit == memsw.limit */
244         bool            memsw_is_minimum;
245
246         /* protect arrays of thresholds */
247         struct mutex thresholds_lock;
248
249         /* thresholds for memory usage. RCU-protected */
250         struct mem_cgroup_thresholds thresholds;
251
252         /* thresholds for mem+swap usage. RCU-protected */
253         struct mem_cgroup_thresholds memsw_thresholds;
254
255         /* For oom notifier event fd */
256         struct list_head oom_notify;
257
258         /*
259          * Should we move charges of a task when a task is moved into this
260          * mem_cgroup ? And what type of charges should we move ?
261          */
262         unsigned long   move_charge_at_immigrate;
263         /*
264          * percpu counter.
265          */
266         struct mem_cgroup_stat_cpu *stat;
267         /*
268          * used when a cpu is offlined or other synchronizations
269          * See mem_cgroup_read_stat().
270          */
271         struct mem_cgroup_stat_cpu nocpu_base;
272         spinlock_t pcp_counter_lock;
273 };
274
275 /* Stuffs for move charges at task migration. */
276 /*
277  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278  * left-shifted bitmap of these types.
279  */
280 enum move_type {
281         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
282         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
283         NR_MOVE_TYPE,
284 };
285
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct {
288         spinlock_t        lock; /* for from, to */
289         struct mem_cgroup *from;
290         struct mem_cgroup *to;
291         unsigned long precharge;
292         unsigned long moved_charge;
293         unsigned long moved_swap;
294         struct task_struct *moving_task;        /* a task moving charges */
295         struct mm_struct *mm;
296         wait_queue_head_t waitq;                /* a waitq for other context */
297 } mc = {
298         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
299         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
300 };
301
302 static bool move_anon(void)
303 {
304         return test_bit(MOVE_CHARGE_TYPE_ANON,
305                                         &mc.to->move_charge_at_immigrate);
306 }
307
308 static bool move_file(void)
309 {
310         return test_bit(MOVE_CHARGE_TYPE_FILE,
311                                         &mc.to->move_charge_at_immigrate);
312 }
313
314 /*
315  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
316  * limit reclaim to prevent infinite loops, if they ever occur.
317  */
318 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
319 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
320
321 enum charge_type {
322         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
323         MEM_CGROUP_CHARGE_TYPE_MAPPED,
324         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
325         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
326         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
327         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
328         NR_CHARGE_TYPE,
329 };
330
331 /* only for here (for easy reading.) */
332 #define PCGF_CACHE      (1UL << PCG_CACHE)
333 #define PCGF_USED       (1UL << PCG_USED)
334 #define PCGF_LOCK       (1UL << PCG_LOCK)
335 /* Not used, but added here for completeness */
336 #define PCGF_ACCT       (1UL << PCG_ACCT)
337
338 /* for encoding cft->private value on file */
339 #define _MEM                    (0)
340 #define _MEMSWAP                (1)
341 #define _OOM_TYPE               (2)
342 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
343 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
344 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
345 /* Used for OOM nofiier */
346 #define OOM_CONTROL             (0)
347
348 /*
349  * Reclaim flags for mem_cgroup_hierarchical_reclaim
350  */
351 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
352 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
353 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
354 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
355 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
356 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
357
358 static void mem_cgroup_get(struct mem_cgroup *mem);
359 static void mem_cgroup_put(struct mem_cgroup *mem);
360 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
361 static void drain_all_stock_async(void);
362
363 static struct mem_cgroup_per_zone *
364 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
365 {
366         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
367 }
368
369 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
370 {
371         return &mem->css;
372 }
373
374 static struct mem_cgroup_per_zone *
375 page_cgroup_zoneinfo(struct page_cgroup *pc)
376 {
377         struct mem_cgroup *mem = pc->mem_cgroup;
378         int nid = page_cgroup_nid(pc);
379         int zid = page_cgroup_zid(pc);
380
381         if (!mem)
382                 return NULL;
383
384         return mem_cgroup_zoneinfo(mem, nid, zid);
385 }
386
387 static struct mem_cgroup_tree_per_zone *
388 soft_limit_tree_node_zone(int nid, int zid)
389 {
390         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
391 }
392
393 static struct mem_cgroup_tree_per_zone *
394 soft_limit_tree_from_page(struct page *page)
395 {
396         int nid = page_to_nid(page);
397         int zid = page_zonenum(page);
398
399         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
400 }
401
402 static void
403 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
404                                 struct mem_cgroup_per_zone *mz,
405                                 struct mem_cgroup_tree_per_zone *mctz,
406                                 unsigned long long new_usage_in_excess)
407 {
408         struct rb_node **p = &mctz->rb_root.rb_node;
409         struct rb_node *parent = NULL;
410         struct mem_cgroup_per_zone *mz_node;
411
412         if (mz->on_tree)
413                 return;
414
415         mz->usage_in_excess = new_usage_in_excess;
416         if (!mz->usage_in_excess)
417                 return;
418         while (*p) {
419                 parent = *p;
420                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
421                                         tree_node);
422                 if (mz->usage_in_excess < mz_node->usage_in_excess)
423                         p = &(*p)->rb_left;
424                 /*
425                  * We can't avoid mem cgroups that are over their soft
426                  * limit by the same amount
427                  */
428                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
429                         p = &(*p)->rb_right;
430         }
431         rb_link_node(&mz->tree_node, parent, p);
432         rb_insert_color(&mz->tree_node, &mctz->rb_root);
433         mz->on_tree = true;
434 }
435
436 static void
437 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
438                                 struct mem_cgroup_per_zone *mz,
439                                 struct mem_cgroup_tree_per_zone *mctz)
440 {
441         if (!mz->on_tree)
442                 return;
443         rb_erase(&mz->tree_node, &mctz->rb_root);
444         mz->on_tree = false;
445 }
446
447 static void
448 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
449                                 struct mem_cgroup_per_zone *mz,
450                                 struct mem_cgroup_tree_per_zone *mctz)
451 {
452         spin_lock(&mctz->lock);
453         __mem_cgroup_remove_exceeded(mem, mz, mctz);
454         spin_unlock(&mctz->lock);
455 }
456
457
458 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
459 {
460         unsigned long long excess;
461         struct mem_cgroup_per_zone *mz;
462         struct mem_cgroup_tree_per_zone *mctz;
463         int nid = page_to_nid(page);
464         int zid = page_zonenum(page);
465         mctz = soft_limit_tree_from_page(page);
466
467         /*
468          * Necessary to update all ancestors when hierarchy is used.
469          * because their event counter is not touched.
470          */
471         for (; mem; mem = parent_mem_cgroup(mem)) {
472                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
473                 excess = res_counter_soft_limit_excess(&mem->res);
474                 /*
475                  * We have to update the tree if mz is on RB-tree or
476                  * mem is over its softlimit.
477                  */
478                 if (excess || mz->on_tree) {
479                         spin_lock(&mctz->lock);
480                         /* if on-tree, remove it */
481                         if (mz->on_tree)
482                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
483                         /*
484                          * Insert again. mz->usage_in_excess will be updated.
485                          * If excess is 0, no tree ops.
486                          */
487                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
488                         spin_unlock(&mctz->lock);
489                 }
490         }
491 }
492
493 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
494 {
495         int node, zone;
496         struct mem_cgroup_per_zone *mz;
497         struct mem_cgroup_tree_per_zone *mctz;
498
499         for_each_node_state(node, N_POSSIBLE) {
500                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
501                         mz = mem_cgroup_zoneinfo(mem, node, zone);
502                         mctz = soft_limit_tree_node_zone(node, zone);
503                         mem_cgroup_remove_exceeded(mem, mz, mctz);
504                 }
505         }
506 }
507
508 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
509 {
510         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
511 }
512
513 static struct mem_cgroup_per_zone *
514 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
515 {
516         struct rb_node *rightmost = NULL;
517         struct mem_cgroup_per_zone *mz;
518
519 retry:
520         mz = NULL;
521         rightmost = rb_last(&mctz->rb_root);
522         if (!rightmost)
523                 goto done;              /* Nothing to reclaim from */
524
525         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
526         /*
527          * Remove the node now but someone else can add it back,
528          * we will to add it back at the end of reclaim to its correct
529          * position in the tree.
530          */
531         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
532         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
533                 !css_tryget(&mz->mem->css))
534                 goto retry;
535 done:
536         return mz;
537 }
538
539 static struct mem_cgroup_per_zone *
540 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
541 {
542         struct mem_cgroup_per_zone *mz;
543
544         spin_lock(&mctz->lock);
545         mz = __mem_cgroup_largest_soft_limit_node(mctz);
546         spin_unlock(&mctz->lock);
547         return mz;
548 }
549
550 /*
551  * Implementation Note: reading percpu statistics for memcg.
552  *
553  * Both of vmstat[] and percpu_counter has threshold and do periodic
554  * synchronization to implement "quick" read. There are trade-off between
555  * reading cost and precision of value. Then, we may have a chance to implement
556  * a periodic synchronizion of counter in memcg's counter.
557  *
558  * But this _read() function is used for user interface now. The user accounts
559  * memory usage by memory cgroup and he _always_ requires exact value because
560  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
561  * have to visit all online cpus and make sum. So, for now, unnecessary
562  * synchronization is not implemented. (just implemented for cpu hotplug)
563  *
564  * If there are kernel internal actions which can make use of some not-exact
565  * value, and reading all cpu value can be performance bottleneck in some
566  * common workload, threashold and synchonization as vmstat[] should be
567  * implemented.
568  */
569 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
570                 enum mem_cgroup_stat_index idx)
571 {
572         int cpu;
573         s64 val = 0;
574
575         get_online_cpus();
576         for_each_online_cpu(cpu)
577                 val += per_cpu(mem->stat->count[idx], cpu);
578 #ifdef CONFIG_HOTPLUG_CPU
579         spin_lock(&mem->pcp_counter_lock);
580         val += mem->nocpu_base.count[idx];
581         spin_unlock(&mem->pcp_counter_lock);
582 #endif
583         put_online_cpus();
584         return val;
585 }
586
587 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
588 {
589         s64 ret;
590
591         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
592         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
593         return ret;
594 }
595
596 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
597                                          bool charge)
598 {
599         int val = (charge) ? 1 : -1;
600         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
601 }
602
603 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
604                                          struct page_cgroup *pc,
605                                          bool charge)
606 {
607         int val = (charge) ? 1 : -1;
608
609         preempt_disable();
610
611         if (PageCgroupCache(pc))
612                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
613         else
614                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
615
616         if (charge)
617                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
618         else
619                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
620         __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
621
622         preempt_enable();
623 }
624
625 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
626                                         enum lru_list idx)
627 {
628         int nid, zid;
629         struct mem_cgroup_per_zone *mz;
630         u64 total = 0;
631
632         for_each_online_node(nid)
633                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
634                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
635                         total += MEM_CGROUP_ZSTAT(mz, idx);
636                 }
637         return total;
638 }
639
640 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
641 {
642         s64 val;
643
644         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
645
646         return !(val & ((1 << event_mask_shift) - 1));
647 }
648
649 /*
650  * Check events in order.
651  *
652  */
653 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
654 {
655         /* threshold event is triggered in finer grain than soft limit */
656         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
657                 mem_cgroup_threshold(mem);
658                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
659                         mem_cgroup_update_tree(mem, page);
660         }
661 }
662
663 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
664 {
665         return container_of(cgroup_subsys_state(cont,
666                                 mem_cgroup_subsys_id), struct mem_cgroup,
667                                 css);
668 }
669
670 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
671 {
672         /*
673          * mm_update_next_owner() may clear mm->owner to NULL
674          * if it races with swapoff, page migration, etc.
675          * So this can be called with p == NULL.
676          */
677         if (unlikely(!p))
678                 return NULL;
679
680         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
681                                 struct mem_cgroup, css);
682 }
683
684 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
685 {
686         struct mem_cgroup *mem = NULL;
687
688         if (!mm)
689                 return NULL;
690         /*
691          * Because we have no locks, mm->owner's may be being moved to other
692          * cgroup. We use css_tryget() here even if this looks
693          * pessimistic (rather than adding locks here).
694          */
695         rcu_read_lock();
696         do {
697                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
698                 if (unlikely(!mem))
699                         break;
700         } while (!css_tryget(&mem->css));
701         rcu_read_unlock();
702         return mem;
703 }
704
705 /* The caller has to guarantee "mem" exists before calling this */
706 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
707 {
708         struct cgroup_subsys_state *css;
709         int found;
710
711         if (!mem) /* ROOT cgroup has the smallest ID */
712                 return root_mem_cgroup; /*css_put/get against root is ignored*/
713         if (!mem->use_hierarchy) {
714                 if (css_tryget(&mem->css))
715                         return mem;
716                 return NULL;
717         }
718         rcu_read_lock();
719         /*
720          * searching a memory cgroup which has the smallest ID under given
721          * ROOT cgroup. (ID >= 1)
722          */
723         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
724         if (css && css_tryget(css))
725                 mem = container_of(css, struct mem_cgroup, css);
726         else
727                 mem = NULL;
728         rcu_read_unlock();
729         return mem;
730 }
731
732 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
733                                         struct mem_cgroup *root,
734                                         bool cond)
735 {
736         int nextid = css_id(&iter->css) + 1;
737         int found;
738         int hierarchy_used;
739         struct cgroup_subsys_state *css;
740
741         hierarchy_used = iter->use_hierarchy;
742
743         css_put(&iter->css);
744         /* If no ROOT, walk all, ignore hierarchy */
745         if (!cond || (root && !hierarchy_used))
746                 return NULL;
747
748         if (!root)
749                 root = root_mem_cgroup;
750
751         do {
752                 iter = NULL;
753                 rcu_read_lock();
754
755                 css = css_get_next(&mem_cgroup_subsys, nextid,
756                                 &root->css, &found);
757                 if (css && css_tryget(css))
758                         iter = container_of(css, struct mem_cgroup, css);
759                 rcu_read_unlock();
760                 /* If css is NULL, no more cgroups will be found */
761                 nextid = found + 1;
762         } while (css && !iter);
763
764         return iter;
765 }
766 /*
767  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
768  * be careful that "break" loop is not allowed. We have reference count.
769  * Instead of that modify "cond" to be false and "continue" to exit the loop.
770  */
771 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
772         for (iter = mem_cgroup_start_loop(root);\
773              iter != NULL;\
774              iter = mem_cgroup_get_next(iter, root, cond))
775
776 #define for_each_mem_cgroup_tree(iter, root) \
777         for_each_mem_cgroup_tree_cond(iter, root, true)
778
779 #define for_each_mem_cgroup_all(iter) \
780         for_each_mem_cgroup_tree_cond(iter, NULL, true)
781
782
783 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
784 {
785         return (mem == root_mem_cgroup);
786 }
787
788 /*
789  * Following LRU functions are allowed to be used without PCG_LOCK.
790  * Operations are called by routine of global LRU independently from memcg.
791  * What we have to take care of here is validness of pc->mem_cgroup.
792  *
793  * Changes to pc->mem_cgroup happens when
794  * 1. charge
795  * 2. moving account
796  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
797  * It is added to LRU before charge.
798  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
799  * When moving account, the page is not on LRU. It's isolated.
800  */
801
802 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
803 {
804         struct page_cgroup *pc;
805         struct mem_cgroup_per_zone *mz;
806
807         if (mem_cgroup_disabled())
808                 return;
809         pc = lookup_page_cgroup(page);
810         /* can happen while we handle swapcache. */
811         if (!TestClearPageCgroupAcctLRU(pc))
812                 return;
813         VM_BUG_ON(!pc->mem_cgroup);
814         /*
815          * We don't check PCG_USED bit. It's cleared when the "page" is finally
816          * removed from global LRU.
817          */
818         mz = page_cgroup_zoneinfo(pc);
819         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
820         if (mem_cgroup_is_root(pc->mem_cgroup))
821                 return;
822         VM_BUG_ON(list_empty(&pc->lru));
823         list_del_init(&pc->lru);
824         return;
825 }
826
827 void mem_cgroup_del_lru(struct page *page)
828 {
829         mem_cgroup_del_lru_list(page, page_lru(page));
830 }
831
832 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
833 {
834         struct mem_cgroup_per_zone *mz;
835         struct page_cgroup *pc;
836
837         if (mem_cgroup_disabled())
838                 return;
839
840         pc = lookup_page_cgroup(page);
841         /*
842          * Used bit is set without atomic ops but after smp_wmb().
843          * For making pc->mem_cgroup visible, insert smp_rmb() here.
844          */
845         smp_rmb();
846         /* unused or root page is not rotated. */
847         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
848                 return;
849         mz = page_cgroup_zoneinfo(pc);
850         list_move(&pc->lru, &mz->lists[lru]);
851 }
852
853 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
854 {
855         struct page_cgroup *pc;
856         struct mem_cgroup_per_zone *mz;
857
858         if (mem_cgroup_disabled())
859                 return;
860         pc = lookup_page_cgroup(page);
861         VM_BUG_ON(PageCgroupAcctLRU(pc));
862         /*
863          * Used bit is set without atomic ops but after smp_wmb().
864          * For making pc->mem_cgroup visible, insert smp_rmb() here.
865          */
866         smp_rmb();
867         if (!PageCgroupUsed(pc))
868                 return;
869
870         mz = page_cgroup_zoneinfo(pc);
871         MEM_CGROUP_ZSTAT(mz, lru) += 1;
872         SetPageCgroupAcctLRU(pc);
873         if (mem_cgroup_is_root(pc->mem_cgroup))
874                 return;
875         list_add(&pc->lru, &mz->lists[lru]);
876 }
877
878 /*
879  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
880  * lru because the page may.be reused after it's fully uncharged (because of
881  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
882  * it again. This function is only used to charge SwapCache. It's done under
883  * lock_page and expected that zone->lru_lock is never held.
884  */
885 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
886 {
887         unsigned long flags;
888         struct zone *zone = page_zone(page);
889         struct page_cgroup *pc = lookup_page_cgroup(page);
890
891         spin_lock_irqsave(&zone->lru_lock, flags);
892         /*
893          * Forget old LRU when this page_cgroup is *not* used. This Used bit
894          * is guarded by lock_page() because the page is SwapCache.
895          */
896         if (!PageCgroupUsed(pc))
897                 mem_cgroup_del_lru_list(page, page_lru(page));
898         spin_unlock_irqrestore(&zone->lru_lock, flags);
899 }
900
901 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
902 {
903         unsigned long flags;
904         struct zone *zone = page_zone(page);
905         struct page_cgroup *pc = lookup_page_cgroup(page);
906
907         spin_lock_irqsave(&zone->lru_lock, flags);
908         /* link when the page is linked to LRU but page_cgroup isn't */
909         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
910                 mem_cgroup_add_lru_list(page, page_lru(page));
911         spin_unlock_irqrestore(&zone->lru_lock, flags);
912 }
913
914
915 void mem_cgroup_move_lists(struct page *page,
916                            enum lru_list from, enum lru_list to)
917 {
918         if (mem_cgroup_disabled())
919                 return;
920         mem_cgroup_del_lru_list(page, from);
921         mem_cgroup_add_lru_list(page, to);
922 }
923
924 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
925 {
926         int ret;
927         struct mem_cgroup *curr = NULL;
928         struct task_struct *p;
929
930         p = find_lock_task_mm(task);
931         if (!p)
932                 return 0;
933         curr = try_get_mem_cgroup_from_mm(p->mm);
934         task_unlock(p);
935         if (!curr)
936                 return 0;
937         /*
938          * We should check use_hierarchy of "mem" not "curr". Because checking
939          * use_hierarchy of "curr" here make this function true if hierarchy is
940          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
941          * hierarchy(even if use_hierarchy is disabled in "mem").
942          */
943         if (mem->use_hierarchy)
944                 ret = css_is_ancestor(&curr->css, &mem->css);
945         else
946                 ret = (curr == mem);
947         css_put(&curr->css);
948         return ret;
949 }
950
951 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
952 {
953         unsigned long active;
954         unsigned long inactive;
955         unsigned long gb;
956         unsigned long inactive_ratio;
957
958         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
959         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
960
961         gb = (inactive + active) >> (30 - PAGE_SHIFT);
962         if (gb)
963                 inactive_ratio = int_sqrt(10 * gb);
964         else
965                 inactive_ratio = 1;
966
967         if (present_pages) {
968                 present_pages[0] = inactive;
969                 present_pages[1] = active;
970         }
971
972         return inactive_ratio;
973 }
974
975 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
976 {
977         unsigned long active;
978         unsigned long inactive;
979         unsigned long present_pages[2];
980         unsigned long inactive_ratio;
981
982         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
983
984         inactive = present_pages[0];
985         active = present_pages[1];
986
987         if (inactive * inactive_ratio < active)
988                 return 1;
989
990         return 0;
991 }
992
993 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
994 {
995         unsigned long active;
996         unsigned long inactive;
997
998         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
999         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1000
1001         return (active > inactive);
1002 }
1003
1004 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
1005                                        struct zone *zone,
1006                                        enum lru_list lru)
1007 {
1008         int nid = zone_to_nid(zone);
1009         int zid = zone_idx(zone);
1010         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1011
1012         return MEM_CGROUP_ZSTAT(mz, lru);
1013 }
1014
1015 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1016                                                       struct zone *zone)
1017 {
1018         int nid = zone_to_nid(zone);
1019         int zid = zone_idx(zone);
1020         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1021
1022         return &mz->reclaim_stat;
1023 }
1024
1025 struct zone_reclaim_stat *
1026 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1027 {
1028         struct page_cgroup *pc;
1029         struct mem_cgroup_per_zone *mz;
1030         int page_size = PAGE_SIZE;
1031
1032         if (PageTransHuge(page))
1033                 page_size <<= compound_order(page);
1034
1035         if (mem_cgroup_disabled())
1036                 return NULL;
1037
1038         pc = lookup_page_cgroup(page);
1039         /*
1040          * Used bit is set without atomic ops but after smp_wmb().
1041          * For making pc->mem_cgroup visible, insert smp_rmb() here.
1042          */
1043         smp_rmb();
1044         if (!PageCgroupUsed(pc))
1045                 return NULL;
1046
1047         mz = page_cgroup_zoneinfo(pc);
1048         if (!mz)
1049                 return NULL;
1050
1051         return &mz->reclaim_stat;
1052 }
1053
1054 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1055                                         struct list_head *dst,
1056                                         unsigned long *scanned, int order,
1057                                         int mode, struct zone *z,
1058                                         struct mem_cgroup *mem_cont,
1059                                         int active, int file)
1060 {
1061         unsigned long nr_taken = 0;
1062         struct page *page;
1063         unsigned long scan;
1064         LIST_HEAD(pc_list);
1065         struct list_head *src;
1066         struct page_cgroup *pc, *tmp;
1067         int nid = zone_to_nid(z);
1068         int zid = zone_idx(z);
1069         struct mem_cgroup_per_zone *mz;
1070         int lru = LRU_FILE * file + active;
1071         int ret;
1072
1073         BUG_ON(!mem_cont);
1074         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1075         src = &mz->lists[lru];
1076
1077         scan = 0;
1078         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1079                 if (scan >= nr_to_scan)
1080                         break;
1081
1082                 page = pc->page;
1083                 if (unlikely(!PageCgroupUsed(pc)))
1084                         continue;
1085                 if (unlikely(!PageLRU(page)))
1086                         continue;
1087
1088                 scan++;
1089                 ret = __isolate_lru_page(page, mode, file);
1090                 switch (ret) {
1091                 case 0:
1092                         list_move(&page->lru, dst);
1093                         mem_cgroup_del_lru(page);
1094                         nr_taken++;
1095                         break;
1096                 case -EBUSY:
1097                         /* we don't affect global LRU but rotate in our LRU */
1098                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1099                         break;
1100                 default:
1101                         break;
1102                 }
1103         }
1104
1105         *scanned = scan;
1106
1107         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1108                                       0, 0, 0, mode);
1109
1110         return nr_taken;
1111 }
1112
1113 #define mem_cgroup_from_res_counter(counter, member)    \
1114         container_of(counter, struct mem_cgroup, member)
1115
1116 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1117 {
1118         if (do_swap_account) {
1119                 if (res_counter_check_under_limit(&mem->res) &&
1120                         res_counter_check_under_limit(&mem->memsw))
1121                         return true;
1122         } else
1123                 if (res_counter_check_under_limit(&mem->res))
1124                         return true;
1125         return false;
1126 }
1127
1128 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1129 {
1130         struct cgroup *cgrp = memcg->css.cgroup;
1131         unsigned int swappiness;
1132
1133         /* root ? */
1134         if (cgrp->parent == NULL)
1135                 return vm_swappiness;
1136
1137         spin_lock(&memcg->reclaim_param_lock);
1138         swappiness = memcg->swappiness;
1139         spin_unlock(&memcg->reclaim_param_lock);
1140
1141         return swappiness;
1142 }
1143
1144 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1145 {
1146         int cpu;
1147
1148         get_online_cpus();
1149         spin_lock(&mem->pcp_counter_lock);
1150         for_each_online_cpu(cpu)
1151                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1152         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1153         spin_unlock(&mem->pcp_counter_lock);
1154         put_online_cpus();
1155
1156         synchronize_rcu();
1157 }
1158
1159 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1160 {
1161         int cpu;
1162
1163         if (!mem)
1164                 return;
1165         get_online_cpus();
1166         spin_lock(&mem->pcp_counter_lock);
1167         for_each_online_cpu(cpu)
1168                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1169         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1170         spin_unlock(&mem->pcp_counter_lock);
1171         put_online_cpus();
1172 }
1173 /*
1174  * 2 routines for checking "mem" is under move_account() or not.
1175  *
1176  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1177  *                        for avoiding race in accounting. If true,
1178  *                        pc->mem_cgroup may be overwritten.
1179  *
1180  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1181  *                        under hierarchy of moving cgroups. This is for
1182  *                        waiting at hith-memory prressure caused by "move".
1183  */
1184
1185 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1186 {
1187         VM_BUG_ON(!rcu_read_lock_held());
1188         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1189 }
1190
1191 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1192 {
1193         struct mem_cgroup *from;
1194         struct mem_cgroup *to;
1195         bool ret = false;
1196         /*
1197          * Unlike task_move routines, we access mc.to, mc.from not under
1198          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1199          */
1200         spin_lock(&mc.lock);
1201         from = mc.from;
1202         to = mc.to;
1203         if (!from)
1204                 goto unlock;
1205         if (from == mem || to == mem
1206             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1207             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1208                 ret = true;
1209 unlock:
1210         spin_unlock(&mc.lock);
1211         return ret;
1212 }
1213
1214 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1215 {
1216         if (mc.moving_task && current != mc.moving_task) {
1217                 if (mem_cgroup_under_move(mem)) {
1218                         DEFINE_WAIT(wait);
1219                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1220                         /* moving charge context might have finished. */
1221                         if (mc.moving_task)
1222                                 schedule();
1223                         finish_wait(&mc.waitq, &wait);
1224                         return true;
1225                 }
1226         }
1227         return false;
1228 }
1229
1230 /**
1231  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1232  * @memcg: The memory cgroup that went over limit
1233  * @p: Task that is going to be killed
1234  *
1235  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1236  * enabled
1237  */
1238 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1239 {
1240         struct cgroup *task_cgrp;
1241         struct cgroup *mem_cgrp;
1242         /*
1243          * Need a buffer in BSS, can't rely on allocations. The code relies
1244          * on the assumption that OOM is serialized for memory controller.
1245          * If this assumption is broken, revisit this code.
1246          */
1247         static char memcg_name[PATH_MAX];
1248         int ret;
1249
1250         if (!memcg || !p)
1251                 return;
1252
1253
1254         rcu_read_lock();
1255
1256         mem_cgrp = memcg->css.cgroup;
1257         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1258
1259         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1260         if (ret < 0) {
1261                 /*
1262                  * Unfortunately, we are unable to convert to a useful name
1263                  * But we'll still print out the usage information
1264                  */
1265                 rcu_read_unlock();
1266                 goto done;
1267         }
1268         rcu_read_unlock();
1269
1270         printk(KERN_INFO "Task in %s killed", memcg_name);
1271
1272         rcu_read_lock();
1273         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1274         if (ret < 0) {
1275                 rcu_read_unlock();
1276                 goto done;
1277         }
1278         rcu_read_unlock();
1279
1280         /*
1281          * Continues from above, so we don't need an KERN_ level
1282          */
1283         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1284 done:
1285
1286         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1287                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1288                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1289                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1290         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1291                 "failcnt %llu\n",
1292                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1293                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1294                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1295 }
1296
1297 /*
1298  * This function returns the number of memcg under hierarchy tree. Returns
1299  * 1(self count) if no children.
1300  */
1301 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1302 {
1303         int num = 0;
1304         struct mem_cgroup *iter;
1305
1306         for_each_mem_cgroup_tree(iter, mem)
1307                 num++;
1308         return num;
1309 }
1310
1311 /*
1312  * Return the memory (and swap, if configured) limit for a memcg.
1313  */
1314 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1315 {
1316         u64 limit;
1317         u64 memsw;
1318
1319         limit = res_counter_read_u64(&memcg->res, RES_LIMIT) +
1320                         total_swap_pages;
1321         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1322         /*
1323          * If memsw is finite and limits the amount of swap space available
1324          * to this memcg, return that limit.
1325          */
1326         return min(limit, memsw);
1327 }
1328
1329 /*
1330  * Visit the first child (need not be the first child as per the ordering
1331  * of the cgroup list, since we track last_scanned_child) of @mem and use
1332  * that to reclaim free pages from.
1333  */
1334 static struct mem_cgroup *
1335 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1336 {
1337         struct mem_cgroup *ret = NULL;
1338         struct cgroup_subsys_state *css;
1339         int nextid, found;
1340
1341         if (!root_mem->use_hierarchy) {
1342                 css_get(&root_mem->css);
1343                 ret = root_mem;
1344         }
1345
1346         while (!ret) {
1347                 rcu_read_lock();
1348                 nextid = root_mem->last_scanned_child + 1;
1349                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1350                                    &found);
1351                 if (css && css_tryget(css))
1352                         ret = container_of(css, struct mem_cgroup, css);
1353
1354                 rcu_read_unlock();
1355                 /* Updates scanning parameter */
1356                 spin_lock(&root_mem->reclaim_param_lock);
1357                 if (!css) {
1358                         /* this means start scan from ID:1 */
1359                         root_mem->last_scanned_child = 0;
1360                 } else
1361                         root_mem->last_scanned_child = found;
1362                 spin_unlock(&root_mem->reclaim_param_lock);
1363         }
1364
1365         return ret;
1366 }
1367
1368 /*
1369  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1370  * we reclaimed from, so that we don't end up penalizing one child extensively
1371  * based on its position in the children list.
1372  *
1373  * root_mem is the original ancestor that we've been reclaim from.
1374  *
1375  * We give up and return to the caller when we visit root_mem twice.
1376  * (other groups can be removed while we're walking....)
1377  *
1378  * If shrink==true, for avoiding to free too much, this returns immedieately.
1379  */
1380 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1381                                                 struct zone *zone,
1382                                                 gfp_t gfp_mask,
1383                                                 unsigned long reclaim_options)
1384 {
1385         struct mem_cgroup *victim;
1386         int ret, total = 0;
1387         int loop = 0;
1388         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1389         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1390         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1391         unsigned long excess = mem_cgroup_get_excess(root_mem);
1392
1393         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1394         if (root_mem->memsw_is_minimum)
1395                 noswap = true;
1396
1397         while (1) {
1398                 victim = mem_cgroup_select_victim(root_mem);
1399                 if (victim == root_mem) {
1400                         loop++;
1401                         if (loop >= 1)
1402                                 drain_all_stock_async();
1403                         if (loop >= 2) {
1404                                 /*
1405                                  * If we have not been able to reclaim
1406                                  * anything, it might because there are
1407                                  * no reclaimable pages under this hierarchy
1408                                  */
1409                                 if (!check_soft || !total) {
1410                                         css_put(&victim->css);
1411                                         break;
1412                                 }
1413                                 /*
1414                                  * We want to do more targetted reclaim.
1415                                  * excess >> 2 is not to excessive so as to
1416                                  * reclaim too much, nor too less that we keep
1417                                  * coming back to reclaim from this cgroup
1418                                  */
1419                                 if (total >= (excess >> 2) ||
1420                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1421                                         css_put(&victim->css);
1422                                         break;
1423                                 }
1424                         }
1425                 }
1426                 if (!mem_cgroup_local_usage(victim)) {
1427                         /* this cgroup's local usage == 0 */
1428                         css_put(&victim->css);
1429                         continue;
1430                 }
1431                 /* we use swappiness of local cgroup */
1432                 if (check_soft)
1433                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1434                                 noswap, get_swappiness(victim), zone);
1435                 else
1436                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1437                                                 noswap, get_swappiness(victim));
1438                 css_put(&victim->css);
1439                 /*
1440                  * At shrinking usage, we can't check we should stop here or
1441                  * reclaim more. It's depends on callers. last_scanned_child
1442                  * will work enough for keeping fairness under tree.
1443                  */
1444                 if (shrink)
1445                         return ret;
1446                 total += ret;
1447                 if (check_soft) {
1448                         if (res_counter_check_under_soft_limit(&root_mem->res))
1449                                 return total;
1450                 } else if (mem_cgroup_check_under_limit(root_mem))
1451                         return 1 + total;
1452         }
1453         return total;
1454 }
1455
1456 /*
1457  * Check OOM-Killer is already running under our hierarchy.
1458  * If someone is running, return false.
1459  */
1460 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1461 {
1462         int x, lock_count = 0;
1463         struct mem_cgroup *iter;
1464
1465         for_each_mem_cgroup_tree(iter, mem) {
1466                 x = atomic_inc_return(&iter->oom_lock);
1467                 lock_count = max(x, lock_count);
1468         }
1469
1470         if (lock_count == 1)
1471                 return true;
1472         return false;
1473 }
1474
1475 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1476 {
1477         struct mem_cgroup *iter;
1478
1479         /*
1480          * When a new child is created while the hierarchy is under oom,
1481          * mem_cgroup_oom_lock() may not be called. We have to use
1482          * atomic_add_unless() here.
1483          */
1484         for_each_mem_cgroup_tree(iter, mem)
1485                 atomic_add_unless(&iter->oom_lock, -1, 0);
1486         return 0;
1487 }
1488
1489
1490 static DEFINE_MUTEX(memcg_oom_mutex);
1491 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1492
1493 struct oom_wait_info {
1494         struct mem_cgroup *mem;
1495         wait_queue_t    wait;
1496 };
1497
1498 static int memcg_oom_wake_function(wait_queue_t *wait,
1499         unsigned mode, int sync, void *arg)
1500 {
1501         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1502         struct oom_wait_info *oom_wait_info;
1503
1504         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1505
1506         if (oom_wait_info->mem == wake_mem)
1507                 goto wakeup;
1508         /* if no hierarchy, no match */
1509         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1510                 return 0;
1511         /*
1512          * Both of oom_wait_info->mem and wake_mem are stable under us.
1513          * Then we can use css_is_ancestor without taking care of RCU.
1514          */
1515         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1516             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1517                 return 0;
1518
1519 wakeup:
1520         return autoremove_wake_function(wait, mode, sync, arg);
1521 }
1522
1523 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1524 {
1525         /* for filtering, pass "mem" as argument. */
1526         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1527 }
1528
1529 static void memcg_oom_recover(struct mem_cgroup *mem)
1530 {
1531         if (mem && atomic_read(&mem->oom_lock))
1532                 memcg_wakeup_oom(mem);
1533 }
1534
1535 /*
1536  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1537  */
1538 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1539 {
1540         struct oom_wait_info owait;
1541         bool locked, need_to_kill;
1542
1543         owait.mem = mem;
1544         owait.wait.flags = 0;
1545         owait.wait.func = memcg_oom_wake_function;
1546         owait.wait.private = current;
1547         INIT_LIST_HEAD(&owait.wait.task_list);
1548         need_to_kill = true;
1549         /* At first, try to OOM lock hierarchy under mem.*/
1550         mutex_lock(&memcg_oom_mutex);
1551         locked = mem_cgroup_oom_lock(mem);
1552         /*
1553          * Even if signal_pending(), we can't quit charge() loop without
1554          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1555          * under OOM is always welcomed, use TASK_KILLABLE here.
1556          */
1557         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1558         if (!locked || mem->oom_kill_disable)
1559                 need_to_kill = false;
1560         if (locked)
1561                 mem_cgroup_oom_notify(mem);
1562         mutex_unlock(&memcg_oom_mutex);
1563
1564         if (need_to_kill) {
1565                 finish_wait(&memcg_oom_waitq, &owait.wait);
1566                 mem_cgroup_out_of_memory(mem, mask);
1567         } else {
1568                 schedule();
1569                 finish_wait(&memcg_oom_waitq, &owait.wait);
1570         }
1571         mutex_lock(&memcg_oom_mutex);
1572         mem_cgroup_oom_unlock(mem);
1573         memcg_wakeup_oom(mem);
1574         mutex_unlock(&memcg_oom_mutex);
1575
1576         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1577                 return false;
1578         /* Give chance to dying process */
1579         schedule_timeout(1);
1580         return true;
1581 }
1582
1583 /*
1584  * Currently used to update mapped file statistics, but the routine can be
1585  * generalized to update other statistics as well.
1586  *
1587  * Notes: Race condition
1588  *
1589  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1590  * it tends to be costly. But considering some conditions, we doesn't need
1591  * to do so _always_.
1592  *
1593  * Considering "charge", lock_page_cgroup() is not required because all
1594  * file-stat operations happen after a page is attached to radix-tree. There
1595  * are no race with "charge".
1596  *
1597  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1598  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1599  * if there are race with "uncharge". Statistics itself is properly handled
1600  * by flags.
1601  *
1602  * Considering "move", this is an only case we see a race. To make the race
1603  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1604  * possibility of race condition. If there is, we take a lock.
1605  */
1606
1607 static void mem_cgroup_update_file_stat(struct page *page, int idx, int val)
1608 {
1609         struct mem_cgroup *mem;
1610         struct page_cgroup *pc = lookup_page_cgroup(page);
1611         bool need_unlock = false;
1612
1613         if (unlikely(!pc))
1614                 return;
1615
1616         rcu_read_lock();
1617         mem = pc->mem_cgroup;
1618         if (unlikely(!mem || !PageCgroupUsed(pc)))
1619                 goto out;
1620         /* pc->mem_cgroup is unstable ? */
1621         if (unlikely(mem_cgroup_stealed(mem))) {
1622                 /* take a lock against to access pc->mem_cgroup */
1623                 lock_page_cgroup(pc);
1624                 need_unlock = true;
1625                 mem = pc->mem_cgroup;
1626                 if (!mem || !PageCgroupUsed(pc))
1627                         goto out;
1628         }
1629
1630         this_cpu_add(mem->stat->count[idx], val);
1631
1632         switch (idx) {
1633         case MEM_CGROUP_STAT_FILE_MAPPED:
1634                 if (val > 0)
1635                         SetPageCgroupFileMapped(pc);
1636                 else if (!page_mapped(page))
1637                         ClearPageCgroupFileMapped(pc);
1638                 break;
1639         default:
1640                 BUG();
1641         }
1642
1643 out:
1644         if (unlikely(need_unlock))
1645                 unlock_page_cgroup(pc);
1646         rcu_read_unlock();
1647         return;
1648 }
1649
1650 void mem_cgroup_update_file_mapped(struct page *page, int val)
1651 {
1652         mem_cgroup_update_file_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, val);
1653 }
1654
1655 /*
1656  * size of first charge trial. "32" comes from vmscan.c's magic value.
1657  * TODO: maybe necessary to use big numbers in big irons.
1658  */
1659 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1660 struct memcg_stock_pcp {
1661         struct mem_cgroup *cached; /* this never be root cgroup */
1662         int charge;
1663         struct work_struct work;
1664 };
1665 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1666 static atomic_t memcg_drain_count;
1667
1668 /*
1669  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1670  * from local stock and true is returned. If the stock is 0 or charges from a
1671  * cgroup which is not current target, returns false. This stock will be
1672  * refilled.
1673  */
1674 static bool consume_stock(struct mem_cgroup *mem)
1675 {
1676         struct memcg_stock_pcp *stock;
1677         bool ret = true;
1678
1679         stock = &get_cpu_var(memcg_stock);
1680         if (mem == stock->cached && stock->charge)
1681                 stock->charge -= PAGE_SIZE;
1682         else /* need to call res_counter_charge */
1683                 ret = false;
1684         put_cpu_var(memcg_stock);
1685         return ret;
1686 }
1687
1688 /*
1689  * Returns stocks cached in percpu to res_counter and reset cached information.
1690  */
1691 static void drain_stock(struct memcg_stock_pcp *stock)
1692 {
1693         struct mem_cgroup *old = stock->cached;
1694
1695         if (stock->charge) {
1696                 res_counter_uncharge(&old->res, stock->charge);
1697                 if (do_swap_account)
1698                         res_counter_uncharge(&old->memsw, stock->charge);
1699         }
1700         stock->cached = NULL;
1701         stock->charge = 0;
1702 }
1703
1704 /*
1705  * This must be called under preempt disabled or must be called by
1706  * a thread which is pinned to local cpu.
1707  */
1708 static void drain_local_stock(struct work_struct *dummy)
1709 {
1710         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1711         drain_stock(stock);
1712 }
1713
1714 /*
1715  * Cache charges(val) which is from res_counter, to local per_cpu area.
1716  * This will be consumed by consume_stock() function, later.
1717  */
1718 static void refill_stock(struct mem_cgroup *mem, int val)
1719 {
1720         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1721
1722         if (stock->cached != mem) { /* reset if necessary */
1723                 drain_stock(stock);
1724                 stock->cached = mem;
1725         }
1726         stock->charge += val;
1727         put_cpu_var(memcg_stock);
1728 }
1729
1730 /*
1731  * Tries to drain stocked charges in other cpus. This function is asynchronous
1732  * and just put a work per cpu for draining localy on each cpu. Caller can
1733  * expects some charges will be back to res_counter later but cannot wait for
1734  * it.
1735  */
1736 static void drain_all_stock_async(void)
1737 {
1738         int cpu;
1739         /* This function is for scheduling "drain" in asynchronous way.
1740          * The result of "drain" is not directly handled by callers. Then,
1741          * if someone is calling drain, we don't have to call drain more.
1742          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1743          * there is a race. We just do loose check here.
1744          */
1745         if (atomic_read(&memcg_drain_count))
1746                 return;
1747         /* Notify other cpus that system-wide "drain" is running */
1748         atomic_inc(&memcg_drain_count);
1749         get_online_cpus();
1750         for_each_online_cpu(cpu) {
1751                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1752                 schedule_work_on(cpu, &stock->work);
1753         }
1754         put_online_cpus();
1755         atomic_dec(&memcg_drain_count);
1756         /* We don't wait for flush_work */
1757 }
1758
1759 /* This is a synchronous drain interface. */
1760 static void drain_all_stock_sync(void)
1761 {
1762         /* called when force_empty is called */
1763         atomic_inc(&memcg_drain_count);
1764         schedule_on_each_cpu(drain_local_stock);
1765         atomic_dec(&memcg_drain_count);
1766 }
1767
1768 /*
1769  * This function drains percpu counter value from DEAD cpu and
1770  * move it to local cpu. Note that this function can be preempted.
1771  */
1772 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1773 {
1774         int i;
1775
1776         spin_lock(&mem->pcp_counter_lock);
1777         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1778                 s64 x = per_cpu(mem->stat->count[i], cpu);
1779
1780                 per_cpu(mem->stat->count[i], cpu) = 0;
1781                 mem->nocpu_base.count[i] += x;
1782         }
1783         /* need to clear ON_MOVE value, works as a kind of lock. */
1784         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1785         spin_unlock(&mem->pcp_counter_lock);
1786 }
1787
1788 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1789 {
1790         int idx = MEM_CGROUP_ON_MOVE;
1791
1792         spin_lock(&mem->pcp_counter_lock);
1793         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1794         spin_unlock(&mem->pcp_counter_lock);
1795 }
1796
1797 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1798                                         unsigned long action,
1799                                         void *hcpu)
1800 {
1801         int cpu = (unsigned long)hcpu;
1802         struct memcg_stock_pcp *stock;
1803         struct mem_cgroup *iter;
1804
1805         if ((action == CPU_ONLINE)) {
1806                 for_each_mem_cgroup_all(iter)
1807                         synchronize_mem_cgroup_on_move(iter, cpu);
1808                 return NOTIFY_OK;
1809         }
1810
1811         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1812                 return NOTIFY_OK;
1813
1814         for_each_mem_cgroup_all(iter)
1815                 mem_cgroup_drain_pcp_counter(iter, cpu);
1816
1817         stock = &per_cpu(memcg_stock, cpu);
1818         drain_stock(stock);
1819         return NOTIFY_OK;
1820 }
1821
1822
1823 /* See __mem_cgroup_try_charge() for details */
1824 enum {
1825         CHARGE_OK,              /* success */
1826         CHARGE_RETRY,           /* need to retry but retry is not bad */
1827         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1828         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1829         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1830 };
1831
1832 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1833                                 int csize, bool oom_check)
1834 {
1835         struct mem_cgroup *mem_over_limit;
1836         struct res_counter *fail_res;
1837         unsigned long flags = 0;
1838         int ret;
1839
1840         ret = res_counter_charge(&mem->res, csize, &fail_res);
1841
1842         if (likely(!ret)) {
1843                 if (!do_swap_account)
1844                         return CHARGE_OK;
1845                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1846                 if (likely(!ret))
1847                         return CHARGE_OK;
1848
1849                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1850                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1851         } else
1852                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1853
1854         if (csize > PAGE_SIZE) /* change csize and retry */
1855                 return CHARGE_RETRY;
1856
1857         if (!(gfp_mask & __GFP_WAIT))
1858                 return CHARGE_WOULDBLOCK;
1859
1860         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1861                                         gfp_mask, flags);
1862         /*
1863          * try_to_free_mem_cgroup_pages() might not give us a full
1864          * picture of reclaim. Some pages are reclaimed and might be
1865          * moved to swap cache or just unmapped from the cgroup.
1866          * Check the limit again to see if the reclaim reduced the
1867          * current usage of the cgroup before giving up
1868          */
1869         if (ret || mem_cgroup_check_under_limit(mem_over_limit))
1870                 return CHARGE_RETRY;
1871
1872         /*
1873          * At task move, charge accounts can be doubly counted. So, it's
1874          * better to wait until the end of task_move if something is going on.
1875          */
1876         if (mem_cgroup_wait_acct_move(mem_over_limit))
1877                 return CHARGE_RETRY;
1878
1879         /* If we don't need to call oom-killer at el, return immediately */
1880         if (!oom_check)
1881                 return CHARGE_NOMEM;
1882         /* check OOM */
1883         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1884                 return CHARGE_OOM_DIE;
1885
1886         return CHARGE_RETRY;
1887 }
1888
1889 /*
1890  * Unlike exported interface, "oom" parameter is added. if oom==true,
1891  * oom-killer can be invoked.
1892  */
1893 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1894                                    gfp_t gfp_mask,
1895                                    struct mem_cgroup **memcg, bool oom,
1896                                    int page_size)
1897 {
1898         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1899         struct mem_cgroup *mem = NULL;
1900         int ret;
1901         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1902
1903         /*
1904          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1905          * in system level. So, allow to go ahead dying process in addition to
1906          * MEMDIE process.
1907          */
1908         if (unlikely(test_thread_flag(TIF_MEMDIE)
1909                      || fatal_signal_pending(current)))
1910                 goto bypass;
1911
1912         /*
1913          * We always charge the cgroup the mm_struct belongs to.
1914          * The mm_struct's mem_cgroup changes on task migration if the
1915          * thread group leader migrates. It's possible that mm is not
1916          * set, if so charge the init_mm (happens for pagecache usage).
1917          */
1918         if (!*memcg && !mm)
1919                 goto bypass;
1920 again:
1921         if (*memcg) { /* css should be a valid one */
1922                 mem = *memcg;
1923                 VM_BUG_ON(css_is_removed(&mem->css));
1924                 if (mem_cgroup_is_root(mem))
1925                         goto done;
1926                 if (page_size == PAGE_SIZE && consume_stock(mem))
1927                         goto done;
1928                 css_get(&mem->css);
1929         } else {
1930                 struct task_struct *p;
1931
1932                 rcu_read_lock();
1933                 p = rcu_dereference(mm->owner);
1934                 /*
1935                  * Because we don't have task_lock(), "p" can exit.
1936                  * In that case, "mem" can point to root or p can be NULL with
1937                  * race with swapoff. Then, we have small risk of mis-accouning.
1938                  * But such kind of mis-account by race always happens because
1939                  * we don't have cgroup_mutex(). It's overkill and we allo that
1940                  * small race, here.
1941                  * (*) swapoff at el will charge against mm-struct not against
1942                  * task-struct. So, mm->owner can be NULL.
1943                  */
1944                 mem = mem_cgroup_from_task(p);
1945                 if (!mem || mem_cgroup_is_root(mem)) {
1946                         rcu_read_unlock();
1947                         goto done;
1948                 }
1949                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1950                         /*
1951                          * It seems dagerous to access memcg without css_get().
1952                          * But considering how consume_stok works, it's not
1953                          * necessary. If consume_stock success, some charges
1954                          * from this memcg are cached on this cpu. So, we
1955                          * don't need to call css_get()/css_tryget() before
1956                          * calling consume_stock().
1957                          */
1958                         rcu_read_unlock();
1959                         goto done;
1960                 }
1961                 /* after here, we may be blocked. we need to get refcnt */
1962                 if (!css_tryget(&mem->css)) {
1963                         rcu_read_unlock();
1964                         goto again;
1965                 }
1966                 rcu_read_unlock();
1967         }
1968
1969         do {
1970                 bool oom_check;
1971
1972                 /* If killed, bypass charge */
1973                 if (fatal_signal_pending(current)) {
1974                         css_put(&mem->css);
1975                         goto bypass;
1976                 }
1977
1978                 oom_check = false;
1979                 if (oom && !nr_oom_retries) {
1980                         oom_check = true;
1981                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1982                 }
1983
1984                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1985
1986                 switch (ret) {
1987                 case CHARGE_OK:
1988                         break;
1989                 case CHARGE_RETRY: /* not in OOM situation but retry */
1990                         csize = page_size;
1991                         css_put(&mem->css);
1992                         mem = NULL;
1993                         goto again;
1994                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1995                         css_put(&mem->css);
1996                         goto nomem;
1997                 case CHARGE_NOMEM: /* OOM routine works */
1998                         if (!oom) {
1999                                 css_put(&mem->css);
2000                                 goto nomem;
2001                         }
2002                         /* If oom, we never return -ENOMEM */
2003                         nr_oom_retries--;
2004                         break;
2005                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2006                         css_put(&mem->css);
2007                         goto bypass;
2008                 }
2009         } while (ret != CHARGE_OK);
2010
2011         if (csize > page_size)
2012                 refill_stock(mem, csize - page_size);
2013         css_put(&mem->css);
2014 done:
2015         *memcg = mem;
2016         return 0;
2017 nomem:
2018         *memcg = NULL;
2019         return -ENOMEM;
2020 bypass:
2021         *memcg = NULL;
2022         return 0;
2023 }
2024
2025 /*
2026  * Somemtimes we have to undo a charge we got by try_charge().
2027  * This function is for that and do uncharge, put css's refcnt.
2028  * gotten by try_charge().
2029  */
2030 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2031                                                         unsigned long count)
2032 {
2033         if (!mem_cgroup_is_root(mem)) {
2034                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2035                 if (do_swap_account)
2036                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2037         }
2038 }
2039
2040 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2041                                      int page_size)
2042 {
2043         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2044 }
2045
2046 /*
2047  * A helper function to get mem_cgroup from ID. must be called under
2048  * rcu_read_lock(). The caller must check css_is_removed() or some if
2049  * it's concern. (dropping refcnt from swap can be called against removed
2050  * memcg.)
2051  */
2052 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2053 {
2054         struct cgroup_subsys_state *css;
2055
2056         /* ID 0 is unused ID */
2057         if (!id)
2058                 return NULL;
2059         css = css_lookup(&mem_cgroup_subsys, id);
2060         if (!css)
2061                 return NULL;
2062         return container_of(css, struct mem_cgroup, css);
2063 }
2064
2065 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2066 {
2067         struct mem_cgroup *mem = NULL;
2068         struct page_cgroup *pc;
2069         unsigned short id;
2070         swp_entry_t ent;
2071
2072         VM_BUG_ON(!PageLocked(page));
2073
2074         pc = lookup_page_cgroup(page);
2075         lock_page_cgroup(pc);
2076         if (PageCgroupUsed(pc)) {
2077                 mem = pc->mem_cgroup;
2078                 if (mem && !css_tryget(&mem->css))
2079                         mem = NULL;
2080         } else if (PageSwapCache(page)) {
2081                 ent.val = page_private(page);
2082                 id = lookup_swap_cgroup(ent);
2083                 rcu_read_lock();
2084                 mem = mem_cgroup_lookup(id);
2085                 if (mem && !css_tryget(&mem->css))
2086                         mem = NULL;
2087                 rcu_read_unlock();
2088         }
2089         unlock_page_cgroup(pc);
2090         return mem;
2091 }
2092
2093 /*
2094  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
2095  * USED state. If already USED, uncharge and return.
2096  */
2097 static void ____mem_cgroup_commit_charge(struct mem_cgroup *mem,
2098                                          struct page_cgroup *pc,
2099                                          enum charge_type ctype)
2100 {
2101         pc->mem_cgroup = mem;
2102         /*
2103          * We access a page_cgroup asynchronously without lock_page_cgroup().
2104          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2105          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2106          * before USED bit, we need memory barrier here.
2107          * See mem_cgroup_add_lru_list(), etc.
2108          */
2109         smp_wmb();
2110         switch (ctype) {
2111         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2112         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2113                 SetPageCgroupCache(pc);
2114                 SetPageCgroupUsed(pc);
2115                 break;
2116         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2117                 ClearPageCgroupCache(pc);
2118                 SetPageCgroupUsed(pc);
2119                 break;
2120         default:
2121                 break;
2122         }
2123
2124         mem_cgroup_charge_statistics(mem, pc, true);
2125 }
2126
2127 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2128                                        struct page_cgroup *pc,
2129                                        enum charge_type ctype,
2130                                        int page_size)
2131 {
2132         int i;
2133         int count = page_size >> PAGE_SHIFT;
2134
2135         /* try_charge() can return NULL to *memcg, taking care of it. */
2136         if (!mem)
2137                 return;
2138
2139         lock_page_cgroup(pc);
2140         if (unlikely(PageCgroupUsed(pc))) {
2141                 unlock_page_cgroup(pc);
2142                 mem_cgroup_cancel_charge(mem, page_size);
2143                 return;
2144         }
2145
2146         /*
2147          * we don't need page_cgroup_lock about tail pages, becase they are not
2148          * accessed by any other context at this point.
2149          */
2150         for (i = 0; i < count; i++)
2151                 ____mem_cgroup_commit_charge(mem, pc + i, ctype);
2152
2153         unlock_page_cgroup(pc);
2154         /*
2155          * "charge_statistics" updated event counter. Then, check it.
2156          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2157          * if they exceeds softlimit.
2158          */
2159         memcg_check_events(mem, pc->page);
2160 }
2161
2162 /**
2163  * __mem_cgroup_move_account - move account of the page
2164  * @pc: page_cgroup of the page.
2165  * @from: mem_cgroup which the page is moved from.
2166  * @to: mem_cgroup which the page is moved to. @from != @to.
2167  * @uncharge: whether we should call uncharge and css_put against @from.
2168  *
2169  * The caller must confirm following.
2170  * - page is not on LRU (isolate_page() is useful.)
2171  * - the pc is locked, used, and ->mem_cgroup points to @from.
2172  *
2173  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2174  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2175  * true, this function does "uncharge" from old cgroup, but it doesn't if
2176  * @uncharge is false, so a caller should do "uncharge".
2177  */
2178
2179 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2180         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2181 {
2182         VM_BUG_ON(from == to);
2183         VM_BUG_ON(PageLRU(pc->page));
2184         VM_BUG_ON(!page_is_cgroup_locked(pc));
2185         VM_BUG_ON(!PageCgroupUsed(pc));
2186         VM_BUG_ON(pc->mem_cgroup != from);
2187
2188         if (PageCgroupFileMapped(pc)) {
2189                 /* Update mapped_file data for mem_cgroup */
2190                 preempt_disable();
2191                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2192                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2193                 preempt_enable();
2194         }
2195         mem_cgroup_charge_statistics(from, pc, false);
2196         if (uncharge)
2197                 /* This is not "cancel", but cancel_charge does all we need. */
2198                 mem_cgroup_cancel_charge(from, PAGE_SIZE);
2199
2200         /* caller should have done css_get */
2201         pc->mem_cgroup = to;
2202         mem_cgroup_charge_statistics(to, pc, true);
2203         /*
2204          * We charges against "to" which may not have any tasks. Then, "to"
2205          * can be under rmdir(). But in current implementation, caller of
2206          * this function is just force_empty() and move charge, so it's
2207          * garanteed that "to" is never removed. So, we don't check rmdir
2208          * status here.
2209          */
2210 }
2211
2212 /*
2213  * check whether the @pc is valid for moving account and call
2214  * __mem_cgroup_move_account()
2215  */
2216 static int mem_cgroup_move_account(struct page_cgroup *pc,
2217                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2218 {
2219         int ret = -EINVAL;
2220         lock_page_cgroup(pc);
2221         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2222                 __mem_cgroup_move_account(pc, from, to, uncharge);
2223                 ret = 0;
2224         }
2225         unlock_page_cgroup(pc);
2226         /*
2227          * check events
2228          */
2229         memcg_check_events(to, pc->page);
2230         memcg_check_events(from, pc->page);
2231         return ret;
2232 }
2233
2234 /*
2235  * move charges to its parent.
2236  */
2237
2238 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2239                                   struct mem_cgroup *child,
2240                                   gfp_t gfp_mask)
2241 {
2242         struct page *page = pc->page;
2243         struct cgroup *cg = child->css.cgroup;
2244         struct cgroup *pcg = cg->parent;
2245         struct mem_cgroup *parent;
2246         int ret;
2247
2248         /* Is ROOT ? */
2249         if (!pcg)
2250                 return -EINVAL;
2251
2252         ret = -EBUSY;
2253         if (!get_page_unless_zero(page))
2254                 goto out;
2255         if (isolate_lru_page(page))
2256                 goto put;
2257
2258         parent = mem_cgroup_from_cont(pcg);
2259         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false,
2260                                       PAGE_SIZE);
2261         if (ret || !parent)
2262                 goto put_back;
2263
2264         ret = mem_cgroup_move_account(pc, child, parent, true);
2265         if (ret)
2266                 mem_cgroup_cancel_charge(parent, PAGE_SIZE);
2267 put_back:
2268         putback_lru_page(page);
2269 put:
2270         put_page(page);
2271 out:
2272         return ret;
2273 }
2274
2275 /*
2276  * Charge the memory controller for page usage.
2277  * Return
2278  * 0 if the charge was successful
2279  * < 0 if the cgroup is over its limit
2280  */
2281 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2282                                 gfp_t gfp_mask, enum charge_type ctype)
2283 {
2284         struct mem_cgroup *mem = NULL;
2285         struct page_cgroup *pc;
2286         int ret;
2287         int page_size = PAGE_SIZE;
2288
2289         if (PageTransHuge(page))
2290                 page_size <<= compound_order(page);
2291
2292         pc = lookup_page_cgroup(page);
2293         /* can happen at boot */
2294         if (unlikely(!pc))
2295                 return 0;
2296         prefetchw(pc);
2297
2298         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page_size);
2299         if (ret || !mem)
2300                 return ret;
2301
2302         __mem_cgroup_commit_charge(mem, pc, ctype, page_size);
2303         return 0;
2304 }
2305
2306 int mem_cgroup_newpage_charge(struct page *page,
2307                               struct mm_struct *mm, gfp_t gfp_mask)
2308 {
2309         if (mem_cgroup_disabled())
2310                 return 0;
2311         /*
2312          * If already mapped, we don't have to account.
2313          * If page cache, page->mapping has address_space.
2314          * But page->mapping may have out-of-use anon_vma pointer,
2315          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2316          * is NULL.
2317          */
2318         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2319                 return 0;
2320         if (unlikely(!mm))
2321                 mm = &init_mm;
2322         return mem_cgroup_charge_common(page, mm, gfp_mask,
2323                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2324 }
2325
2326 static void
2327 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2328                                         enum charge_type ctype);
2329
2330 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2331                                 gfp_t gfp_mask)
2332 {
2333         int ret;
2334
2335         if (mem_cgroup_disabled())
2336                 return 0;
2337         if (PageCompound(page))
2338                 return 0;
2339         /*
2340          * Corner case handling. This is called from add_to_page_cache()
2341          * in usual. But some FS (shmem) precharges this page before calling it
2342          * and call add_to_page_cache() with GFP_NOWAIT.
2343          *
2344          * For GFP_NOWAIT case, the page may be pre-charged before calling
2345          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2346          * charge twice. (It works but has to pay a bit larger cost.)
2347          * And when the page is SwapCache, it should take swap information
2348          * into account. This is under lock_page() now.
2349          */
2350         if (!(gfp_mask & __GFP_WAIT)) {
2351                 struct page_cgroup *pc;
2352
2353                 pc = lookup_page_cgroup(page);
2354                 if (!pc)
2355                         return 0;
2356                 lock_page_cgroup(pc);
2357                 if (PageCgroupUsed(pc)) {
2358                         unlock_page_cgroup(pc);
2359                         return 0;
2360                 }
2361                 unlock_page_cgroup(pc);
2362         }
2363
2364         if (unlikely(!mm))
2365                 mm = &init_mm;
2366
2367         if (page_is_file_cache(page))
2368                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2369                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2370
2371         /* shmem */
2372         if (PageSwapCache(page)) {
2373                 struct mem_cgroup *mem = NULL;
2374
2375                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2376                 if (!ret)
2377                         __mem_cgroup_commit_charge_swapin(page, mem,
2378                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2379         } else
2380                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2381                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2382
2383         return ret;
2384 }
2385
2386 /*
2387  * While swap-in, try_charge -> commit or cancel, the page is locked.
2388  * And when try_charge() successfully returns, one refcnt to memcg without
2389  * struct page_cgroup is acquired. This refcnt will be consumed by
2390  * "commit()" or removed by "cancel()"
2391  */
2392 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2393                                  struct page *page,
2394                                  gfp_t mask, struct mem_cgroup **ptr)
2395 {
2396         struct mem_cgroup *mem;
2397         int ret;
2398
2399         if (mem_cgroup_disabled())
2400                 return 0;
2401
2402         if (!do_swap_account)
2403                 goto charge_cur_mm;
2404         /*
2405          * A racing thread's fault, or swapoff, may have already updated
2406          * the pte, and even removed page from swap cache: in those cases
2407          * do_swap_page()'s pte_same() test will fail; but there's also a
2408          * KSM case which does need to charge the page.
2409          */
2410         if (!PageSwapCache(page))
2411                 goto charge_cur_mm;
2412         mem = try_get_mem_cgroup_from_page(page);
2413         if (!mem)
2414                 goto charge_cur_mm;
2415         *ptr = mem;
2416         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2417         css_put(&mem->css);
2418         return ret;
2419 charge_cur_mm:
2420         if (unlikely(!mm))
2421                 mm = &init_mm;
2422         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2423 }
2424
2425 static void
2426 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2427                                         enum charge_type ctype)
2428 {
2429         struct page_cgroup *pc;
2430
2431         if (mem_cgroup_disabled())
2432                 return;
2433         if (!ptr)
2434                 return;
2435         cgroup_exclude_rmdir(&ptr->css);
2436         pc = lookup_page_cgroup(page);
2437         mem_cgroup_lru_del_before_commit_swapcache(page);
2438         __mem_cgroup_commit_charge(ptr, pc, ctype, PAGE_SIZE);
2439         mem_cgroup_lru_add_after_commit_swapcache(page);
2440         /*
2441          * Now swap is on-memory. This means this page may be
2442          * counted both as mem and swap....double count.
2443          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2444          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2445          * may call delete_from_swap_cache() before reach here.
2446          */
2447         if (do_swap_account && PageSwapCache(page)) {
2448                 swp_entry_t ent = {.val = page_private(page)};
2449                 unsigned short id;
2450                 struct mem_cgroup *memcg;
2451
2452                 id = swap_cgroup_record(ent, 0);
2453                 rcu_read_lock();
2454                 memcg = mem_cgroup_lookup(id);
2455                 if (memcg) {
2456                         /*
2457                          * This recorded memcg can be obsolete one. So, avoid
2458                          * calling css_tryget
2459                          */
2460                         if (!mem_cgroup_is_root(memcg))
2461                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2462                         mem_cgroup_swap_statistics(memcg, false);
2463                         mem_cgroup_put(memcg);
2464                 }
2465                 rcu_read_unlock();
2466         }
2467         /*
2468          * At swapin, we may charge account against cgroup which has no tasks.
2469          * So, rmdir()->pre_destroy() can be called while we do this charge.
2470          * In that case, we need to call pre_destroy() again. check it here.
2471          */
2472         cgroup_release_and_wakeup_rmdir(&ptr->css);
2473 }
2474
2475 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2476 {
2477         __mem_cgroup_commit_charge_swapin(page, ptr,
2478                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2479 }
2480
2481 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2482 {
2483         if (mem_cgroup_disabled())
2484                 return;
2485         if (!mem)
2486                 return;
2487         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2488 }
2489
2490 static void
2491 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2492               int page_size)
2493 {
2494         struct memcg_batch_info *batch = NULL;
2495         bool uncharge_memsw = true;
2496         /* If swapout, usage of swap doesn't decrease */
2497         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2498                 uncharge_memsw = false;
2499
2500         batch = &current->memcg_batch;
2501         /*
2502          * In usual, we do css_get() when we remember memcg pointer.
2503          * But in this case, we keep res->usage until end of a series of
2504          * uncharges. Then, it's ok to ignore memcg's refcnt.
2505          */
2506         if (!batch->memcg)
2507                 batch->memcg = mem;
2508         /*
2509          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2510          * In those cases, all pages freed continously can be expected to be in
2511          * the same cgroup and we have chance to coalesce uncharges.
2512          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2513          * because we want to do uncharge as soon as possible.
2514          */
2515
2516         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2517                 goto direct_uncharge;
2518
2519         if (page_size != PAGE_SIZE)
2520                 goto direct_uncharge;
2521
2522         /*
2523          * In typical case, batch->memcg == mem. This means we can
2524          * merge a series of uncharges to an uncharge of res_counter.
2525          * If not, we uncharge res_counter ony by one.
2526          */
2527         if (batch->memcg != mem)
2528                 goto direct_uncharge;
2529         /* remember freed charge and uncharge it later */
2530         batch->bytes += PAGE_SIZE;
2531         if (uncharge_memsw)
2532                 batch->memsw_bytes += PAGE_SIZE;
2533         return;
2534 direct_uncharge:
2535         res_counter_uncharge(&mem->res, page_size);
2536         if (uncharge_memsw)
2537                 res_counter_uncharge(&mem->memsw, page_size);
2538         if (unlikely(batch->memcg != mem))
2539                 memcg_oom_recover(mem);
2540         return;
2541 }
2542
2543 /*
2544  * uncharge if !page_mapped(page)
2545  */
2546 static struct mem_cgroup *
2547 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2548 {
2549         int i;
2550         int count;
2551         struct page_cgroup *pc;
2552         struct mem_cgroup *mem = NULL;
2553         int page_size = PAGE_SIZE;
2554
2555         if (mem_cgroup_disabled())
2556                 return NULL;
2557
2558         if (PageSwapCache(page))
2559                 return NULL;
2560
2561         if (PageTransHuge(page))
2562                 page_size <<= compound_order(page);
2563
2564         count = page_size >> PAGE_SHIFT;
2565         /*
2566          * Check if our page_cgroup is valid
2567          */
2568         pc = lookup_page_cgroup(page);
2569         if (unlikely(!pc || !PageCgroupUsed(pc)))
2570                 return NULL;
2571
2572         lock_page_cgroup(pc);
2573
2574         mem = pc->mem_cgroup;
2575
2576         if (!PageCgroupUsed(pc))
2577                 goto unlock_out;
2578
2579         switch (ctype) {
2580         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2581         case MEM_CGROUP_CHARGE_TYPE_DROP:
2582                 /* See mem_cgroup_prepare_migration() */
2583                 if (page_mapped(page) || PageCgroupMigration(pc))
2584                         goto unlock_out;
2585                 break;
2586         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2587                 if (!PageAnon(page)) {  /* Shared memory */
2588                         if (page->mapping && !page_is_file_cache(page))
2589                                 goto unlock_out;
2590                 } else if (page_mapped(page)) /* Anon */
2591                                 goto unlock_out;
2592                 break;
2593         default:
2594                 break;
2595         }
2596
2597         for (i = 0; i < count; i++)
2598                 mem_cgroup_charge_statistics(mem, pc + i, false);
2599
2600         ClearPageCgroupUsed(pc);
2601         /*
2602          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2603          * freed from LRU. This is safe because uncharged page is expected not
2604          * to be reused (freed soon). Exception is SwapCache, it's handled by
2605          * special functions.
2606          */
2607
2608         unlock_page_cgroup(pc);
2609         /*
2610          * even after unlock, we have mem->res.usage here and this memcg
2611          * will never be freed.
2612          */
2613         memcg_check_events(mem, page);
2614         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2615                 mem_cgroup_swap_statistics(mem, true);
2616                 mem_cgroup_get(mem);
2617         }
2618         if (!mem_cgroup_is_root(mem))
2619                 __do_uncharge(mem, ctype, page_size);
2620
2621         return mem;
2622
2623 unlock_out:
2624         unlock_page_cgroup(pc);
2625         return NULL;
2626 }
2627
2628 void mem_cgroup_uncharge_page(struct page *page)
2629 {
2630         /* early check. */
2631         if (page_mapped(page))
2632                 return;
2633         if (page->mapping && !PageAnon(page))
2634                 return;
2635         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2636 }
2637
2638 void mem_cgroup_uncharge_cache_page(struct page *page)
2639 {
2640         VM_BUG_ON(page_mapped(page));
2641         VM_BUG_ON(page->mapping);
2642         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2643 }
2644
2645 /*
2646  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2647  * In that cases, pages are freed continuously and we can expect pages
2648  * are in the same memcg. All these calls itself limits the number of
2649  * pages freed at once, then uncharge_start/end() is called properly.
2650  * This may be called prural(2) times in a context,
2651  */
2652
2653 void mem_cgroup_uncharge_start(void)
2654 {
2655         current->memcg_batch.do_batch++;
2656         /* We can do nest. */
2657         if (current->memcg_batch.do_batch == 1) {
2658                 current->memcg_batch.memcg = NULL;
2659                 current->memcg_batch.bytes = 0;
2660                 current->memcg_batch.memsw_bytes = 0;
2661         }
2662 }
2663
2664 void mem_cgroup_uncharge_end(void)
2665 {
2666         struct memcg_batch_info *batch = &current->memcg_batch;
2667
2668         if (!batch->do_batch)
2669                 return;
2670
2671         batch->do_batch--;
2672         if (batch->do_batch) /* If stacked, do nothing. */
2673                 return;
2674
2675         if (!batch->memcg)
2676                 return;
2677         /*
2678          * This "batch->memcg" is valid without any css_get/put etc...
2679          * bacause we hide charges behind us.
2680          */
2681         if (batch->bytes)
2682                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2683         if (batch->memsw_bytes)
2684                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2685         memcg_oom_recover(batch->memcg);
2686         /* forget this pointer (for sanity check) */
2687         batch->memcg = NULL;
2688 }
2689
2690 #ifdef CONFIG_SWAP
2691 /*
2692  * called after __delete_from_swap_cache() and drop "page" account.
2693  * memcg information is recorded to swap_cgroup of "ent"
2694  */
2695 void
2696 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2697 {
2698         struct mem_cgroup *memcg;
2699         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2700
2701         if (!swapout) /* this was a swap cache but the swap is unused ! */
2702                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2703
2704         memcg = __mem_cgroup_uncharge_common(page, ctype);
2705
2706         /*
2707          * record memcg information,  if swapout && memcg != NULL,
2708          * mem_cgroup_get() was called in uncharge().
2709          */
2710         if (do_swap_account && swapout && memcg)
2711                 swap_cgroup_record(ent, css_id(&memcg->css));
2712 }
2713 #endif
2714
2715 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2716 /*
2717  * called from swap_entry_free(). remove record in swap_cgroup and
2718  * uncharge "memsw" account.
2719  */
2720 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2721 {
2722         struct mem_cgroup *memcg;
2723         unsigned short id;
2724
2725         if (!do_swap_account)
2726                 return;
2727
2728         id = swap_cgroup_record(ent, 0);
2729         rcu_read_lock();
2730         memcg = mem_cgroup_lookup(id);
2731         if (memcg) {
2732                 /*
2733                  * We uncharge this because swap is freed.
2734                  * This memcg can be obsolete one. We avoid calling css_tryget
2735                  */
2736                 if (!mem_cgroup_is_root(memcg))
2737                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2738                 mem_cgroup_swap_statistics(memcg, false);
2739                 mem_cgroup_put(memcg);
2740         }
2741         rcu_read_unlock();
2742 }
2743
2744 /**
2745  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2746  * @entry: swap entry to be moved
2747  * @from:  mem_cgroup which the entry is moved from
2748  * @to:  mem_cgroup which the entry is moved to
2749  * @need_fixup: whether we should fixup res_counters and refcounts.
2750  *
2751  * It succeeds only when the swap_cgroup's record for this entry is the same
2752  * as the mem_cgroup's id of @from.
2753  *
2754  * Returns 0 on success, -EINVAL on failure.
2755  *
2756  * The caller must have charged to @to, IOW, called res_counter_charge() about
2757  * both res and memsw, and called css_get().
2758  */
2759 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2760                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2761 {
2762         unsigned short old_id, new_id;
2763
2764         old_id = css_id(&from->css);
2765         new_id = css_id(&to->css);
2766
2767         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2768                 mem_cgroup_swap_statistics(from, false);
2769                 mem_cgroup_swap_statistics(to, true);
2770                 /*
2771                  * This function is only called from task migration context now.
2772                  * It postpones res_counter and refcount handling till the end
2773                  * of task migration(mem_cgroup_clear_mc()) for performance
2774                  * improvement. But we cannot postpone mem_cgroup_get(to)
2775                  * because if the process that has been moved to @to does
2776                  * swap-in, the refcount of @to might be decreased to 0.
2777                  */
2778                 mem_cgroup_get(to);
2779                 if (need_fixup) {
2780                         if (!mem_cgroup_is_root(from))
2781                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2782                         mem_cgroup_put(from);
2783                         /*
2784                          * we charged both to->res and to->memsw, so we should
2785                          * uncharge to->res.
2786                          */
2787                         if (!mem_cgroup_is_root(to))
2788                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2789                 }
2790                 return 0;
2791         }
2792         return -EINVAL;
2793 }
2794 #else
2795 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2796                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2797 {
2798         return -EINVAL;
2799 }
2800 #endif
2801
2802 /*
2803  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2804  * page belongs to.
2805  */
2806 int mem_cgroup_prepare_migration(struct page *page,
2807         struct page *newpage, struct mem_cgroup **ptr)
2808 {
2809         struct page_cgroup *pc;
2810         struct mem_cgroup *mem = NULL;
2811         enum charge_type ctype;
2812         int ret = 0;
2813
2814         VM_BUG_ON(PageTransHuge(page));
2815         if (mem_cgroup_disabled())
2816                 return 0;
2817
2818         pc = lookup_page_cgroup(page);
2819         lock_page_cgroup(pc);
2820         if (PageCgroupUsed(pc)) {
2821                 mem = pc->mem_cgroup;
2822                 css_get(&mem->css);
2823                 /*
2824                  * At migrating an anonymous page, its mapcount goes down
2825                  * to 0 and uncharge() will be called. But, even if it's fully
2826                  * unmapped, migration may fail and this page has to be
2827                  * charged again. We set MIGRATION flag here and delay uncharge
2828                  * until end_migration() is called
2829                  *
2830                  * Corner Case Thinking
2831                  * A)
2832                  * When the old page was mapped as Anon and it's unmap-and-freed
2833                  * while migration was ongoing.
2834                  * If unmap finds the old page, uncharge() of it will be delayed
2835                  * until end_migration(). If unmap finds a new page, it's
2836                  * uncharged when it make mapcount to be 1->0. If unmap code
2837                  * finds swap_migration_entry, the new page will not be mapped
2838                  * and end_migration() will find it(mapcount==0).
2839                  *
2840                  * B)
2841                  * When the old page was mapped but migraion fails, the kernel
2842                  * remaps it. A charge for it is kept by MIGRATION flag even
2843                  * if mapcount goes down to 0. We can do remap successfully
2844                  * without charging it again.
2845                  *
2846                  * C)
2847                  * The "old" page is under lock_page() until the end of
2848                  * migration, so, the old page itself will not be swapped-out.
2849                  * If the new page is swapped out before end_migraton, our
2850                  * hook to usual swap-out path will catch the event.
2851                  */
2852                 if (PageAnon(page))
2853                         SetPageCgroupMigration(pc);
2854         }
2855         unlock_page_cgroup(pc);
2856         /*
2857          * If the page is not charged at this point,
2858          * we return here.
2859          */
2860         if (!mem)
2861                 return 0;
2862
2863         *ptr = mem;
2864         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false, PAGE_SIZE);
2865         css_put(&mem->css);/* drop extra refcnt */
2866         if (ret || *ptr == NULL) {
2867                 if (PageAnon(page)) {
2868                         lock_page_cgroup(pc);
2869                         ClearPageCgroupMigration(pc);
2870                         unlock_page_cgroup(pc);
2871                         /*
2872                          * The old page may be fully unmapped while we kept it.
2873                          */
2874                         mem_cgroup_uncharge_page(page);
2875                 }
2876                 return -ENOMEM;
2877         }
2878         /*
2879          * We charge new page before it's used/mapped. So, even if unlock_page()
2880          * is called before end_migration, we can catch all events on this new
2881          * page. In the case new page is migrated but not remapped, new page's
2882          * mapcount will be finally 0 and we call uncharge in end_migration().
2883          */
2884         pc = lookup_page_cgroup(newpage);
2885         if (PageAnon(page))
2886                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2887         else if (page_is_file_cache(page))
2888                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2889         else
2890                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2891         __mem_cgroup_commit_charge(mem, pc, ctype, PAGE_SIZE);
2892         return ret;
2893 }
2894
2895 /* remove redundant charge if migration failed*/
2896 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2897         struct page *oldpage, struct page *newpage)
2898 {
2899         struct page *used, *unused;
2900         struct page_cgroup *pc;
2901
2902         if (!mem)
2903                 return;
2904         /* blocks rmdir() */
2905         cgroup_exclude_rmdir(&mem->css);
2906         /* at migration success, oldpage->mapping is NULL. */
2907         if (oldpage->mapping) {
2908                 used = oldpage;
2909                 unused = newpage;
2910         } else {
2911                 used = newpage;
2912                 unused = oldpage;
2913         }
2914         /*
2915          * We disallowed uncharge of pages under migration because mapcount
2916          * of the page goes down to zero, temporarly.
2917          * Clear the flag and check the page should be charged.
2918          */
2919         pc = lookup_page_cgroup(oldpage);
2920         lock_page_cgroup(pc);
2921         ClearPageCgroupMigration(pc);
2922         unlock_page_cgroup(pc);
2923
2924         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2925
2926         /*
2927          * If a page is a file cache, radix-tree replacement is very atomic
2928          * and we can skip this check. When it was an Anon page, its mapcount
2929          * goes down to 0. But because we added MIGRATION flage, it's not
2930          * uncharged yet. There are several case but page->mapcount check
2931          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2932          * check. (see prepare_charge() also)
2933          */
2934         if (PageAnon(used))
2935                 mem_cgroup_uncharge_page(used);
2936         /*
2937          * At migration, we may charge account against cgroup which has no
2938          * tasks.
2939          * So, rmdir()->pre_destroy() can be called while we do this charge.
2940          * In that case, we need to call pre_destroy() again. check it here.
2941          */
2942         cgroup_release_and_wakeup_rmdir(&mem->css);
2943 }
2944
2945 /*
2946  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2947  * Calling hierarchical_reclaim is not enough because we should update
2948  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2949  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2950  * not from the memcg which this page would be charged to.
2951  * try_charge_swapin does all of these works properly.
2952  */
2953 int mem_cgroup_shmem_charge_fallback(struct page *page,
2954                             struct mm_struct *mm,
2955                             gfp_t gfp_mask)
2956 {
2957         struct mem_cgroup *mem = NULL;
2958         int ret;
2959
2960         if (mem_cgroup_disabled())
2961                 return 0;
2962
2963         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2964         if (!ret)
2965                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2966
2967         return ret;
2968 }
2969
2970 static DEFINE_MUTEX(set_limit_mutex);
2971
2972 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2973                                 unsigned long long val)
2974 {
2975         int retry_count;
2976         u64 memswlimit, memlimit;
2977         int ret = 0;
2978         int children = mem_cgroup_count_children(memcg);
2979         u64 curusage, oldusage;
2980         int enlarge;
2981
2982         /*
2983          * For keeping hierarchical_reclaim simple, how long we should retry
2984          * is depends on callers. We set our retry-count to be function
2985          * of # of children which we should visit in this loop.
2986          */
2987         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2988
2989         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2990
2991         enlarge = 0;
2992         while (retry_count) {
2993                 if (signal_pending(current)) {
2994                         ret = -EINTR;
2995                         break;
2996                 }
2997                 /*
2998                  * Rather than hide all in some function, I do this in
2999                  * open coded manner. You see what this really does.
3000                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3001                  */
3002                 mutex_lock(&set_limit_mutex);
3003                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3004                 if (memswlimit < val) {
3005                         ret = -EINVAL;
3006                         mutex_unlock(&set_limit_mutex);
3007                         break;
3008                 }
3009
3010                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3011                 if (memlimit < val)
3012                         enlarge = 1;
3013
3014                 ret = res_counter_set_limit(&memcg->res, val);
3015                 if (!ret) {
3016                         if (memswlimit == val)
3017                                 memcg->memsw_is_minimum = true;
3018                         else
3019                                 memcg->memsw_is_minimum = false;
3020                 }
3021                 mutex_unlock(&set_limit_mutex);
3022
3023                 if (!ret)
3024                         break;
3025
3026                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3027                                                 MEM_CGROUP_RECLAIM_SHRINK);
3028                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3029                 /* Usage is reduced ? */
3030                 if (curusage >= oldusage)
3031                         retry_count--;
3032                 else
3033                         oldusage = curusage;
3034         }
3035         if (!ret && enlarge)
3036                 memcg_oom_recover(memcg);
3037
3038         return ret;
3039 }
3040
3041 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3042                                         unsigned long long val)
3043 {
3044         int retry_count;
3045         u64 memlimit, memswlimit, oldusage, curusage;
3046         int children = mem_cgroup_count_children(memcg);
3047         int ret = -EBUSY;
3048         int enlarge = 0;
3049
3050         /* see mem_cgroup_resize_res_limit */
3051         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3052         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3053         while (retry_count) {
3054                 if (signal_pending(current)) {
3055                         ret = -EINTR;
3056                         break;
3057                 }
3058                 /*
3059                  * Rather than hide all in some function, I do this in
3060                  * open coded manner. You see what this really does.
3061                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3062                  */
3063                 mutex_lock(&set_limit_mutex);
3064                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3065                 if (memlimit > val) {
3066                         ret = -EINVAL;
3067                         mutex_unlock(&set_limit_mutex);
3068                         break;
3069                 }
3070                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3071                 if (memswlimit < val)
3072                         enlarge = 1;
3073                 ret = res_counter_set_limit(&memcg->memsw, val);
3074                 if (!ret) {
3075                         if (memlimit == val)
3076                                 memcg->memsw_is_minimum = true;
3077                         else
3078                                 memcg->memsw_is_minimum = false;
3079                 }
3080                 mutex_unlock(&set_limit_mutex);
3081
3082                 if (!ret)
3083                         break;
3084
3085                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3086                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3087                                                 MEM_CGROUP_RECLAIM_SHRINK);
3088                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3089                 /* Usage is reduced ? */
3090                 if (curusage >= oldusage)
3091                         retry_count--;
3092                 else
3093                         oldusage = curusage;
3094         }
3095         if (!ret && enlarge)
3096                 memcg_oom_recover(memcg);
3097         return ret;
3098 }
3099
3100 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3101                                             gfp_t gfp_mask)
3102 {
3103         unsigned long nr_reclaimed = 0;
3104         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3105         unsigned long reclaimed;
3106         int loop = 0;
3107         struct mem_cgroup_tree_per_zone *mctz;
3108         unsigned long long excess;
3109
3110         if (order > 0)
3111                 return 0;
3112
3113         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3114         /*
3115          * This loop can run a while, specially if mem_cgroup's continuously
3116          * keep exceeding their soft limit and putting the system under
3117          * pressure
3118          */
3119         do {
3120                 if (next_mz)
3121                         mz = next_mz;
3122                 else
3123                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3124                 if (!mz)
3125                         break;
3126
3127                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3128                                                 gfp_mask,
3129                                                 MEM_CGROUP_RECLAIM_SOFT);
3130                 nr_reclaimed += reclaimed;
3131                 spin_lock(&mctz->lock);
3132
3133                 /*
3134                  * If we failed to reclaim anything from this memory cgroup
3135                  * it is time to move on to the next cgroup
3136                  */
3137                 next_mz = NULL;
3138                 if (!reclaimed) {
3139                         do {
3140                                 /*
3141                                  * Loop until we find yet another one.
3142                                  *
3143                                  * By the time we get the soft_limit lock
3144                                  * again, someone might have aded the
3145                                  * group back on the RB tree. Iterate to
3146                                  * make sure we get a different mem.
3147                                  * mem_cgroup_largest_soft_limit_node returns
3148                                  * NULL if no other cgroup is present on
3149                                  * the tree
3150                                  */
3151                                 next_mz =
3152                                 __mem_cgroup_largest_soft_limit_node(mctz);
3153                                 if (next_mz == mz) {
3154                                         css_put(&next_mz->mem->css);
3155                                         next_mz = NULL;
3156                                 } else /* next_mz == NULL or other memcg */
3157                                         break;
3158                         } while (1);
3159                 }
3160                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3161                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3162                 /*
3163                  * One school of thought says that we should not add
3164                  * back the node to the tree if reclaim returns 0.
3165                  * But our reclaim could return 0, simply because due
3166                  * to priority we are exposing a smaller subset of
3167                  * memory to reclaim from. Consider this as a longer
3168                  * term TODO.
3169                  */
3170                 /* If excess == 0, no tree ops */
3171                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3172                 spin_unlock(&mctz->lock);
3173                 css_put(&mz->mem->css);
3174                 loop++;
3175                 /*
3176                  * Could not reclaim anything and there are no more
3177                  * mem cgroups to try or we seem to be looping without
3178                  * reclaiming anything.
3179                  */
3180                 if (!nr_reclaimed &&
3181                         (next_mz == NULL ||
3182                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3183                         break;
3184         } while (!nr_reclaimed);
3185         if (next_mz)
3186                 css_put(&next_mz->mem->css);
3187         return nr_reclaimed;
3188 }
3189
3190 /*
3191  * This routine traverse page_cgroup in given list and drop them all.
3192  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3193  */
3194 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3195                                 int node, int zid, enum lru_list lru)
3196 {
3197         struct zone *zone;
3198         struct mem_cgroup_per_zone *mz;
3199         struct page_cgroup *pc, *busy;
3200         unsigned long flags, loop;
3201         struct list_head *list;
3202         int ret = 0;
3203
3204         zone = &NODE_DATA(node)->node_zones[zid];
3205         mz = mem_cgroup_zoneinfo(mem, node, zid);
3206         list = &mz->lists[lru];
3207
3208         loop = MEM_CGROUP_ZSTAT(mz, lru);
3209         /* give some margin against EBUSY etc...*/
3210         loop += 256;
3211         busy = NULL;
3212         while (loop--) {
3213                 ret = 0;
3214                 spin_lock_irqsave(&zone->lru_lock, flags);
3215                 if (list_empty(list)) {
3216                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3217                         break;
3218                 }
3219                 pc = list_entry(list->prev, struct page_cgroup, lru);
3220                 if (busy == pc) {
3221                         list_move(&pc->lru, list);
3222                         busy = NULL;
3223                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3224                         continue;
3225                 }
3226                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3227
3228                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3229                 if (ret == -ENOMEM)
3230                         break;
3231
3232                 if (ret == -EBUSY || ret == -EINVAL) {
3233                         /* found lock contention or "pc" is obsolete. */
3234                         busy = pc;
3235                         cond_resched();
3236                 } else
3237                         busy = NULL;
3238         }
3239
3240         if (!ret && !list_empty(list))
3241                 return -EBUSY;
3242         return ret;
3243 }
3244
3245 /*
3246  * make mem_cgroup's charge to be 0 if there is no task.
3247  * This enables deleting this mem_cgroup.
3248  */
3249 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3250 {
3251         int ret;
3252         int node, zid, shrink;
3253         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3254         struct cgroup *cgrp = mem->css.cgroup;
3255
3256         css_get(&mem->css);
3257
3258         shrink = 0;
3259         /* should free all ? */
3260         if (free_all)
3261                 goto try_to_free;
3262 move_account:
3263         do {
3264                 ret = -EBUSY;
3265                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3266                         goto out;
3267                 ret = -EINTR;
3268                 if (signal_pending(current))
3269                         goto out;
3270                 /* This is for making all *used* pages to be on LRU. */
3271                 lru_add_drain_all();
3272                 drain_all_stock_sync();
3273                 ret = 0;
3274                 mem_cgroup_start_move(mem);
3275                 for_each_node_state(node, N_HIGH_MEMORY) {
3276                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3277                                 enum lru_list l;
3278                                 for_each_lru(l) {
3279                                         ret = mem_cgroup_force_empty_list(mem,
3280                                                         node, zid, l);
3281                                         if (ret)
3282                                                 break;
3283                                 }
3284                         }
3285                         if (ret)
3286                                 break;
3287                 }
3288                 mem_cgroup_end_move(mem);
3289                 memcg_oom_recover(mem);
3290                 /* it seems parent cgroup doesn't have enough mem */
3291                 if (ret == -ENOMEM)
3292                         goto try_to_free;
3293                 cond_resched();
3294         /* "ret" should also be checked to ensure all lists are empty. */
3295         } while (mem->res.usage > 0 || ret);
3296 out:
3297         css_put(&mem->css);
3298         return ret;
3299
3300 try_to_free:
3301         /* returns EBUSY if there is a task or if we come here twice. */
3302         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3303                 ret = -EBUSY;
3304                 goto out;
3305         }
3306         /* we call try-to-free pages for make this cgroup empty */
3307         lru_add_drain_all();
3308         /* try to free all pages in this cgroup */
3309         shrink = 1;
3310         while (nr_retries && mem->res.usage > 0) {
3311                 int progress;
3312
3313                 if (signal_pending(current)) {
3314                         ret = -EINTR;
3315                         goto out;
3316                 }
3317                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3318                                                 false, get_swappiness(mem));
3319                 if (!progress) {
3320                         nr_retries--;
3321                         /* maybe some writeback is necessary */
3322                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3323                 }
3324
3325         }
3326         lru_add_drain();
3327         /* try move_account...there may be some *locked* pages. */
3328         goto move_account;
3329 }
3330
3331 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3332 {
3333         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3334 }
3335
3336
3337 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3338 {
3339         return mem_cgroup_from_cont(cont)->use_hierarchy;
3340 }
3341
3342 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3343                                         u64 val)
3344 {
3345         int retval = 0;
3346         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3347         struct cgroup *parent = cont->parent;
3348         struct mem_cgroup *parent_mem = NULL;
3349
3350         if (parent)
3351                 parent_mem = mem_cgroup_from_cont(parent);
3352
3353         cgroup_lock();
3354         /*
3355          * If parent's use_hierarchy is set, we can't make any modifications
3356          * in the child subtrees. If it is unset, then the change can
3357          * occur, provided the current cgroup has no children.
3358          *
3359          * For the root cgroup, parent_mem is NULL, we allow value to be
3360          * set if there are no children.
3361          */
3362         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3363                                 (val == 1 || val == 0)) {
3364                 if (list_empty(&cont->children))
3365                         mem->use_hierarchy = val;
3366                 else
3367                         retval = -EBUSY;
3368         } else
3369                 retval = -EINVAL;
3370         cgroup_unlock();
3371
3372         return retval;
3373 }
3374
3375
3376 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3377                                 enum mem_cgroup_stat_index idx)
3378 {
3379         struct mem_cgroup *iter;
3380         s64 val = 0;
3381
3382         /* each per cpu's value can be minus.Then, use s64 */
3383         for_each_mem_cgroup_tree(iter, mem)
3384                 val += mem_cgroup_read_stat(iter, idx);
3385
3386         if (val < 0) /* race ? */
3387                 val = 0;
3388         return val;
3389 }
3390
3391 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3392 {
3393         u64 val;
3394
3395         if (!mem_cgroup_is_root(mem)) {
3396                 if (!swap)
3397                         return res_counter_read_u64(&mem->res, RES_USAGE);
3398                 else
3399                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3400         }
3401
3402         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3403         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3404
3405         if (swap)
3406                 val += mem_cgroup_get_recursive_idx_stat(mem,
3407                                 MEM_CGROUP_STAT_SWAPOUT);
3408
3409         return val << PAGE_SHIFT;
3410 }
3411
3412 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3413 {
3414         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3415         u64 val;
3416         int type, name;
3417
3418         type = MEMFILE_TYPE(cft->private);
3419         name = MEMFILE_ATTR(cft->private);
3420         switch (type) {
3421         case _MEM:
3422                 if (name == RES_USAGE)
3423                         val = mem_cgroup_usage(mem, false);
3424                 else
3425                         val = res_counter_read_u64(&mem->res, name);
3426                 break;
3427         case _MEMSWAP:
3428                 if (name == RES_USAGE)
3429                         val = mem_cgroup_usage(mem, true);
3430                 else
3431                         val = res_counter_read_u64(&mem->memsw, name);
3432                 break;
3433         default:
3434                 BUG();
3435                 break;
3436         }
3437         return val;
3438 }
3439 /*
3440  * The user of this function is...
3441  * RES_LIMIT.
3442  */
3443 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3444                             const char *buffer)
3445 {
3446         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3447         int type, name;
3448         unsigned long long val;
3449         int ret;
3450
3451         type = MEMFILE_TYPE(cft->private);
3452         name = MEMFILE_ATTR(cft->private);
3453         switch (name) {
3454         case RES_LIMIT:
3455                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3456                         ret = -EINVAL;
3457                         break;
3458                 }
3459                 /* This function does all necessary parse...reuse it */
3460                 ret = res_counter_memparse_write_strategy(buffer, &val);
3461                 if (ret)
3462                         break;
3463                 if (type == _MEM)
3464                         ret = mem_cgroup_resize_limit(memcg, val);
3465                 else
3466                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3467                 break;
3468         case RES_SOFT_LIMIT:
3469                 ret = res_counter_memparse_write_strategy(buffer, &val);
3470                 if (ret)
3471                         break;
3472                 /*
3473                  * For memsw, soft limits are hard to implement in terms
3474                  * of semantics, for now, we support soft limits for
3475                  * control without swap
3476                  */
3477                 if (type == _MEM)
3478                         ret = res_counter_set_soft_limit(&memcg->res, val);
3479                 else
3480                         ret = -EINVAL;
3481                 break;
3482         default:
3483                 ret = -EINVAL; /* should be BUG() ? */
3484                 break;
3485         }
3486         return ret;
3487 }
3488
3489 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3490                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3491 {
3492         struct cgroup *cgroup;
3493         unsigned long long min_limit, min_memsw_limit, tmp;
3494
3495         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3496         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3497         cgroup = memcg->css.cgroup;
3498         if (!memcg->use_hierarchy)
3499                 goto out;
3500
3501         while (cgroup->parent) {
3502                 cgroup = cgroup->parent;
3503                 memcg = mem_cgroup_from_cont(cgroup);
3504                 if (!memcg->use_hierarchy)
3505                         break;
3506                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3507                 min_limit = min(min_limit, tmp);
3508                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3509                 min_memsw_limit = min(min_memsw_limit, tmp);
3510         }
3511 out:
3512         *mem_limit = min_limit;
3513         *memsw_limit = min_memsw_limit;
3514         return;
3515 }
3516
3517 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3518 {
3519         struct mem_cgroup *mem;
3520         int type, name;
3521
3522         mem = mem_cgroup_from_cont(cont);
3523         type = MEMFILE_TYPE(event);
3524         name = MEMFILE_ATTR(event);
3525         switch (name) {
3526         case RES_MAX_USAGE:
3527                 if (type == _MEM)
3528                         res_counter_reset_max(&mem->res);
3529                 else
3530                         res_counter_reset_max(&mem->memsw);
3531                 break;
3532         case RES_FAILCNT:
3533                 if (type == _MEM)
3534                         res_counter_reset_failcnt(&mem->res);
3535                 else
3536                         res_counter_reset_failcnt(&mem->memsw);
3537                 break;
3538         }
3539
3540         return 0;
3541 }
3542
3543 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3544                                         struct cftype *cft)
3545 {
3546         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3547 }
3548
3549 #ifdef CONFIG_MMU
3550 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3551                                         struct cftype *cft, u64 val)
3552 {
3553         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3554
3555         if (val >= (1 << NR_MOVE_TYPE))
3556                 return -EINVAL;
3557         /*
3558          * We check this value several times in both in can_attach() and
3559          * attach(), so we need cgroup lock to prevent this value from being
3560          * inconsistent.
3561          */
3562         cgroup_lock();
3563         mem->move_charge_at_immigrate = val;
3564         cgroup_unlock();
3565
3566         return 0;
3567 }
3568 #else
3569 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3570                                         struct cftype *cft, u64 val)
3571 {
3572         return -ENOSYS;
3573 }
3574 #endif
3575
3576
3577 /* For read statistics */
3578 enum {
3579         MCS_CACHE,
3580         MCS_RSS,
3581         MCS_FILE_MAPPED,
3582         MCS_PGPGIN,
3583         MCS_PGPGOUT,
3584         MCS_SWAP,
3585         MCS_INACTIVE_ANON,
3586         MCS_ACTIVE_ANON,
3587         MCS_INACTIVE_FILE,
3588         MCS_ACTIVE_FILE,
3589         MCS_UNEVICTABLE,
3590         NR_MCS_STAT,
3591 };
3592
3593 struct mcs_total_stat {
3594         s64 stat[NR_MCS_STAT];
3595 };
3596
3597 struct {
3598         char *local_name;
3599         char *total_name;
3600 } memcg_stat_strings[NR_MCS_STAT] = {
3601         {"cache", "total_cache"},
3602         {"rss", "total_rss"},
3603         {"mapped_file", "total_mapped_file"},
3604         {"pgpgin", "total_pgpgin"},
3605         {"pgpgout", "total_pgpgout"},
3606         {"swap", "total_swap"},
3607         {"inactive_anon", "total_inactive_anon"},
3608         {"active_anon", "total_active_anon"},
3609         {"inactive_file", "total_inactive_file"},
3610         {"active_file", "total_active_file"},
3611         {"unevictable", "total_unevictable"}
3612 };
3613
3614
3615 static void
3616 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3617 {
3618         s64 val;
3619
3620         /* per cpu stat */
3621         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3622         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3623         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3624         s->stat[MCS_RSS] += val * PAGE_SIZE;
3625         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3626         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3627         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3628         s->stat[MCS_PGPGIN] += val;
3629         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3630         s->stat[MCS_PGPGOUT] += val;
3631         if (do_swap_account) {
3632                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3633                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3634         }
3635
3636         /* per zone stat */
3637         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3638         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3639         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3640         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3641         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3642         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3643         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3644         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3645         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3646         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3647 }
3648
3649 static void
3650 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3651 {
3652         struct mem_cgroup *iter;
3653
3654         for_each_mem_cgroup_tree(iter, mem)
3655                 mem_cgroup_get_local_stat(iter, s);
3656 }
3657
3658 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3659                                  struct cgroup_map_cb *cb)
3660 {
3661         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3662         struct mcs_total_stat mystat;
3663         int i;
3664
3665         memset(&mystat, 0, sizeof(mystat));
3666         mem_cgroup_get_local_stat(mem_cont, &mystat);
3667
3668         for (i = 0; i < NR_MCS_STAT; i++) {
3669                 if (i == MCS_SWAP && !do_swap_account)
3670                         continue;
3671                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3672         }
3673
3674         /* Hierarchical information */
3675         {
3676                 unsigned long long limit, memsw_limit;
3677                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3678                 cb->fill(cb, "hierarchical_memory_limit", limit);
3679                 if (do_swap_account)
3680                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3681         }
3682
3683         memset(&mystat, 0, sizeof(mystat));
3684         mem_cgroup_get_total_stat(mem_cont, &mystat);
3685         for (i = 0; i < NR_MCS_STAT; i++) {
3686                 if (i == MCS_SWAP && !do_swap_account)
3687                         continue;
3688                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3689         }
3690
3691 #ifdef CONFIG_DEBUG_VM
3692         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3693
3694         {
3695                 int nid, zid;
3696                 struct mem_cgroup_per_zone *mz;
3697                 unsigned long recent_rotated[2] = {0, 0};
3698                 unsigned long recent_scanned[2] = {0, 0};
3699
3700                 for_each_online_node(nid)
3701                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3702                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3703
3704                                 recent_rotated[0] +=
3705                                         mz->reclaim_stat.recent_rotated[0];
3706                                 recent_rotated[1] +=
3707                                         mz->reclaim_stat.recent_rotated[1];
3708                                 recent_scanned[0] +=
3709                                         mz->reclaim_stat.recent_scanned[0];
3710                                 recent_scanned[1] +=
3711                                         mz->reclaim_stat.recent_scanned[1];
3712                         }
3713                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3714                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3715                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3716                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3717         }
3718 #endif
3719
3720         return 0;
3721 }
3722
3723 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3724 {
3725         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3726
3727         return get_swappiness(memcg);
3728 }
3729
3730 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3731                                        u64 val)
3732 {
3733         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3734         struct mem_cgroup *parent;
3735
3736         if (val > 100)
3737                 return -EINVAL;
3738
3739         if (cgrp->parent == NULL)
3740                 return -EINVAL;
3741
3742         parent = mem_cgroup_from_cont(cgrp->parent);
3743
3744         cgroup_lock();
3745
3746         /* If under hierarchy, only empty-root can set this value */
3747         if ((parent->use_hierarchy) ||
3748             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3749                 cgroup_unlock();
3750                 return -EINVAL;
3751         }
3752
3753         spin_lock(&memcg->reclaim_param_lock);
3754         memcg->swappiness = val;
3755         spin_unlock(&memcg->reclaim_param_lock);
3756
3757         cgroup_unlock();
3758
3759         return 0;
3760 }
3761
3762 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3763 {
3764         struct mem_cgroup_threshold_ary *t;
3765         u64 usage;
3766         int i;
3767
3768         rcu_read_lock();
3769         if (!swap)
3770                 t = rcu_dereference(memcg->thresholds.primary);
3771         else
3772                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3773
3774         if (!t)
3775                 goto unlock;
3776
3777         usage = mem_cgroup_usage(memcg, swap);
3778
3779         /*
3780          * current_threshold points to threshold just below usage.
3781          * If it's not true, a threshold was crossed after last
3782          * call of __mem_cgroup_threshold().
3783          */
3784         i = t->current_threshold;
3785
3786         /*
3787          * Iterate backward over array of thresholds starting from
3788          * current_threshold and check if a threshold is crossed.
3789          * If none of thresholds below usage is crossed, we read
3790          * only one element of the array here.
3791          */
3792         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3793                 eventfd_signal(t->entries[i].eventfd, 1);
3794
3795         /* i = current_threshold + 1 */
3796         i++;
3797
3798         /*
3799          * Iterate forward over array of thresholds starting from
3800          * current_threshold+1 and check if a threshold is crossed.
3801          * If none of thresholds above usage is crossed, we read
3802          * only one element of the array here.
3803          */
3804         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3805                 eventfd_signal(t->entries[i].eventfd, 1);
3806
3807         /* Update current_threshold */
3808         t->current_threshold = i - 1;
3809 unlock:
3810         rcu_read_unlock();
3811 }
3812
3813 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3814 {
3815         while (memcg) {
3816                 __mem_cgroup_threshold(memcg, false);
3817                 if (do_swap_account)
3818                         __mem_cgroup_threshold(memcg, true);
3819
3820                 memcg = parent_mem_cgroup(memcg);
3821         }
3822 }
3823
3824 static int compare_thresholds(const void *a, const void *b)
3825 {
3826         const struct mem_cgroup_threshold *_a = a;
3827         const struct mem_cgroup_threshold *_b = b;
3828
3829         return _a->threshold - _b->threshold;
3830 }
3831
3832 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3833 {
3834         struct mem_cgroup_eventfd_list *ev;
3835
3836         list_for_each_entry(ev, &mem->oom_notify, list)
3837                 eventfd_signal(ev->eventfd, 1);
3838         return 0;
3839 }
3840
3841 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3842 {
3843         struct mem_cgroup *iter;
3844
3845         for_each_mem_cgroup_tree(iter, mem)
3846                 mem_cgroup_oom_notify_cb(iter);
3847 }
3848
3849 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3850         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3851 {
3852         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3853         struct mem_cgroup_thresholds *thresholds;
3854         struct mem_cgroup_threshold_ary *new;
3855         int type = MEMFILE_TYPE(cft->private);
3856         u64 threshold, usage;
3857         int i, size, ret;
3858
3859         ret = res_counter_memparse_write_strategy(args, &threshold);
3860         if (ret)
3861                 return ret;
3862
3863         mutex_lock(&memcg->thresholds_lock);
3864
3865         if (type == _MEM)
3866                 thresholds = &memcg->thresholds;
3867         else if (type == _MEMSWAP)
3868                 thresholds = &memcg->memsw_thresholds;
3869         else
3870                 BUG();
3871
3872         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3873
3874         /* Check if a threshold crossed before adding a new one */
3875         if (thresholds->primary)
3876                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3877
3878         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3879
3880         /* Allocate memory for new array of thresholds */
3881         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3882                         GFP_KERNEL);
3883         if (!new) {
3884                 ret = -ENOMEM;
3885                 goto unlock;
3886         }
3887         new->size = size;
3888
3889         /* Copy thresholds (if any) to new array */
3890         if (thresholds->primary) {
3891                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3892                                 sizeof(struct mem_cgroup_threshold));
3893         }
3894
3895         /* Add new threshold */
3896         new->entries[size - 1].eventfd = eventfd;
3897         new->entries[size - 1].threshold = threshold;
3898
3899         /* Sort thresholds. Registering of new threshold isn't time-critical */
3900         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3901                         compare_thresholds, NULL);
3902
3903         /* Find current threshold */
3904         new->current_threshold = -1;
3905         for (i = 0; i < size; i++) {
3906                 if (new->entries[i].threshold < usage) {
3907                         /*
3908                          * new->current_threshold will not be used until
3909                          * rcu_assign_pointer(), so it's safe to increment
3910                          * it here.
3911                          */
3912                         ++new->current_threshold;
3913                 }
3914         }
3915
3916         /* Free old spare buffer and save old primary buffer as spare */
3917         kfree(thresholds->spare);
3918         thresholds->spare = thresholds->primary;
3919
3920         rcu_assign_pointer(thresholds->primary, new);
3921
3922         /* To be sure that nobody uses thresholds */
3923         synchronize_rcu();
3924
3925 unlock:
3926         mutex_unlock(&memcg->thresholds_lock);
3927
3928         return ret;
3929 }
3930
3931 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3932         struct cftype *cft, struct eventfd_ctx *eventfd)
3933 {
3934         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3935         struct mem_cgroup_thresholds *thresholds;
3936         struct mem_cgroup_threshold_ary *new;
3937         int type = MEMFILE_TYPE(cft->private);
3938         u64 usage;
3939         int i, j, size;
3940
3941         mutex_lock(&memcg->thresholds_lock);
3942         if (type == _MEM)
3943                 thresholds = &memcg->thresholds;
3944         else if (type == _MEMSWAP)
3945                 thresholds = &memcg->memsw_thresholds;
3946         else
3947                 BUG();
3948
3949         /*
3950          * Something went wrong if we trying to unregister a threshold
3951          * if we don't have thresholds
3952          */
3953         BUG_ON(!thresholds);
3954
3955         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3956
3957         /* Check if a threshold crossed before removing */
3958         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3959
3960         /* Calculate new number of threshold */
3961         size = 0;
3962         for (i = 0; i < thresholds->primary->size; i++) {
3963                 if (thresholds->primary->entries[i].eventfd != eventfd)
3964                         size++;
3965         }
3966
3967         new = thresholds->spare;
3968
3969         /* Set thresholds array to NULL if we don't have thresholds */
3970         if (!size) {
3971                 kfree(new);
3972                 new = NULL;
3973                 goto swap_buffers;
3974         }
3975
3976         new->size = size;
3977
3978         /* Copy thresholds and find current threshold */
3979         new->current_threshold = -1;
3980         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3981                 if (thresholds->primary->entries[i].eventfd == eventfd)
3982                         continue;
3983
3984                 new->entries[j] = thresholds->primary->entries[i];
3985                 if (new->entries[j].threshold < usage) {
3986                         /*
3987                          * new->current_threshold will not be used
3988                          * until rcu_assign_pointer(), so it's safe to increment
3989                          * it here.
3990                          */
3991                         ++new->current_threshold;
3992                 }
3993                 j++;
3994         }
3995
3996 swap_buffers:
3997         /* Swap primary and spare array */
3998         thresholds->spare = thresholds->primary;
3999         rcu_assign_pointer(thresholds->primary, new);
4000
4001         /* To be sure that nobody uses thresholds */
4002         synchronize_rcu();
4003
4004         mutex_unlock(&memcg->thresholds_lock);
4005 }
4006
4007 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4008         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4009 {
4010         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4011         struct mem_cgroup_eventfd_list *event;
4012         int type = MEMFILE_TYPE(cft->private);
4013
4014         BUG_ON(type != _OOM_TYPE);
4015         event = kmalloc(sizeof(*event), GFP_KERNEL);
4016         if (!event)
4017                 return -ENOMEM;
4018
4019         mutex_lock(&memcg_oom_mutex);
4020
4021         event->eventfd = eventfd;
4022         list_add(&event->list, &memcg->oom_notify);
4023
4024         /* already in OOM ? */
4025         if (atomic_read(&memcg->oom_lock))
4026                 eventfd_signal(eventfd, 1);
4027         mutex_unlock(&memcg_oom_mutex);
4028
4029         return 0;
4030 }
4031
4032 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4033         struct cftype *cft, struct eventfd_ctx *eventfd)
4034 {
4035         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4036         struct mem_cgroup_eventfd_list *ev, *tmp;
4037         int type = MEMFILE_TYPE(cft->private);
4038
4039         BUG_ON(type != _OOM_TYPE);
4040
4041         mutex_lock(&memcg_oom_mutex);
4042
4043         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4044                 if (ev->eventfd == eventfd) {
4045                         list_del(&ev->list);
4046                         kfree(ev);
4047                 }
4048         }
4049
4050         mutex_unlock(&memcg_oom_mutex);
4051 }
4052
4053 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4054         struct cftype *cft,  struct cgroup_map_cb *cb)
4055 {
4056         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4057
4058         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4059
4060         if (atomic_read(&mem->oom_lock))
4061                 cb->fill(cb, "under_oom", 1);
4062         else
4063                 cb->fill(cb, "under_oom", 0);
4064         return 0;
4065 }
4066
4067 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4068         struct cftype *cft, u64 val)
4069 {
4070         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4071         struct mem_cgroup *parent;
4072
4073         /* cannot set to root cgroup and only 0 and 1 are allowed */
4074         if (!cgrp->parent || !((val == 0) || (val == 1)))
4075                 return -EINVAL;
4076
4077         parent = mem_cgroup_from_cont(cgrp->parent);
4078
4079         cgroup_lock();
4080         /* oom-kill-disable is a flag for subhierarchy. */
4081         if ((parent->use_hierarchy) ||
4082             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4083                 cgroup_unlock();
4084                 return -EINVAL;
4085         }
4086         mem->oom_kill_disable = val;
4087         if (!val)
4088                 memcg_oom_recover(mem);
4089         cgroup_unlock();
4090         return 0;
4091 }
4092
4093 static struct cftype mem_cgroup_files[] = {
4094         {
4095                 .name = "usage_in_bytes",
4096                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4097                 .read_u64 = mem_cgroup_read,
4098                 .register_event = mem_cgroup_usage_register_event,
4099                 .unregister_event = mem_cgroup_usage_unregister_event,
4100         },
4101         {
4102                 .name = "max_usage_in_bytes",
4103                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4104                 .trigger = mem_cgroup_reset,
4105                 .read_u64 = mem_cgroup_read,
4106         },
4107         {
4108                 .name = "limit_in_bytes",
4109                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4110                 .write_string = mem_cgroup_write,
4111                 .read_u64 = mem_cgroup_read,
4112         },
4113         {
4114                 .name = "soft_limit_in_bytes",
4115                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4116                 .write_string = mem_cgroup_write,
4117                 .read_u64 = mem_cgroup_read,
4118         },
4119         {
4120                 .name = "failcnt",
4121                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4122                 .trigger = mem_cgroup_reset,
4123                 .read_u64 = mem_cgroup_read,
4124         },
4125         {
4126                 .name = "stat",
4127                 .read_map = mem_control_stat_show,
4128         },
4129         {
4130                 .name = "force_empty",
4131                 .trigger = mem_cgroup_force_empty_write,
4132         },
4133         {
4134                 .name = "use_hierarchy",
4135                 .write_u64 = mem_cgroup_hierarchy_write,
4136                 .read_u64 = mem_cgroup_hierarchy_read,
4137         },
4138         {
4139                 .name = "swappiness",
4140                 .read_u64 = mem_cgroup_swappiness_read,
4141                 .write_u64 = mem_cgroup_swappiness_write,
4142         },
4143         {
4144                 .name = "move_charge_at_immigrate",
4145                 .read_u64 = mem_cgroup_move_charge_read,
4146                 .write_u64 = mem_cgroup_move_charge_write,
4147         },
4148         {
4149                 .name = "oom_control",
4150                 .read_map = mem_cgroup_oom_control_read,
4151                 .write_u64 = mem_cgroup_oom_control_write,
4152                 .register_event = mem_cgroup_oom_register_event,
4153                 .unregister_event = mem_cgroup_oom_unregister_event,
4154                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4155         },
4156 };
4157
4158 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4159 static struct cftype memsw_cgroup_files[] = {
4160         {
4161                 .name = "memsw.usage_in_bytes",
4162                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4163                 .read_u64 = mem_cgroup_read,
4164                 .register_event = mem_cgroup_usage_register_event,
4165                 .unregister_event = mem_cgroup_usage_unregister_event,
4166         },
4167         {
4168                 .name = "memsw.max_usage_in_bytes",
4169                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4170                 .trigger = mem_cgroup_reset,
4171                 .read_u64 = mem_cgroup_read,
4172         },
4173         {
4174                 .name = "memsw.limit_in_bytes",
4175                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4176                 .write_string = mem_cgroup_write,
4177                 .read_u64 = mem_cgroup_read,
4178         },
4179         {
4180                 .name = "memsw.failcnt",
4181                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4182                 .trigger = mem_cgroup_reset,
4183                 .read_u64 = mem_cgroup_read,
4184         },
4185 };
4186
4187 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4188 {
4189         if (!do_swap_account)
4190                 return 0;
4191         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4192                                 ARRAY_SIZE(memsw_cgroup_files));
4193 };
4194 #else
4195 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4196 {
4197         return 0;
4198 }
4199 #endif
4200
4201 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4202 {
4203         struct mem_cgroup_per_node *pn;
4204         struct mem_cgroup_per_zone *mz;
4205         enum lru_list l;
4206         int zone, tmp = node;
4207         /*
4208          * This routine is called against possible nodes.
4209          * But it's BUG to call kmalloc() against offline node.
4210          *
4211          * TODO: this routine can waste much memory for nodes which will
4212          *       never be onlined. It's better to use memory hotplug callback
4213          *       function.
4214          */
4215         if (!node_state(node, N_NORMAL_MEMORY))
4216                 tmp = -1;
4217         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4218         if (!pn)
4219                 return 1;
4220
4221         mem->info.nodeinfo[node] = pn;
4222         memset(pn, 0, sizeof(*pn));
4223
4224         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4225                 mz = &pn->zoneinfo[zone];
4226                 for_each_lru(l)
4227                         INIT_LIST_HEAD(&mz->lists[l]);
4228                 mz->usage_in_excess = 0;
4229                 mz->on_tree = false;
4230                 mz->mem = mem;
4231         }
4232         return 0;
4233 }
4234
4235 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4236 {
4237         kfree(mem->info.nodeinfo[node]);
4238 }
4239
4240 static struct mem_cgroup *mem_cgroup_alloc(void)
4241 {
4242         struct mem_cgroup *mem;
4243         int size = sizeof(struct mem_cgroup);
4244
4245         /* Can be very big if MAX_NUMNODES is very big */
4246         if (size < PAGE_SIZE)
4247                 mem = kmalloc(size, GFP_KERNEL);
4248         else
4249                 mem = vmalloc(size);
4250
4251         if (!mem)
4252                 return NULL;
4253
4254         memset(mem, 0, size);
4255         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4256         if (!mem->stat)
4257                 goto out_free;
4258         spin_lock_init(&mem->pcp_counter_lock);
4259         return mem;
4260
4261 out_free:
4262         if (size < PAGE_SIZE)
4263                 kfree(mem);
4264         else
4265                 vfree(mem);
4266         return NULL;
4267 }
4268
4269 /*
4270  * At destroying mem_cgroup, references from swap_cgroup can remain.
4271  * (scanning all at force_empty is too costly...)
4272  *
4273  * Instead of clearing all references at force_empty, we remember
4274  * the number of reference from swap_cgroup and free mem_cgroup when
4275  * it goes down to 0.
4276  *
4277  * Removal of cgroup itself succeeds regardless of refs from swap.
4278  */
4279
4280 static void __mem_cgroup_free(struct mem_cgroup *mem)
4281 {
4282         int node;
4283
4284         mem_cgroup_remove_from_trees(mem);
4285         free_css_id(&mem_cgroup_subsys, &mem->css);
4286
4287         for_each_node_state(node, N_POSSIBLE)
4288                 free_mem_cgroup_per_zone_info(mem, node);
4289
4290         free_percpu(mem->stat);
4291         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4292                 kfree(mem);
4293         else
4294                 vfree(mem);
4295 }
4296
4297 static void mem_cgroup_get(struct mem_cgroup *mem)
4298 {
4299         atomic_inc(&mem->refcnt);
4300 }
4301
4302 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4303 {
4304         if (atomic_sub_and_test(count, &mem->refcnt)) {
4305                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4306                 __mem_cgroup_free(mem);
4307                 if (parent)
4308                         mem_cgroup_put(parent);
4309         }
4310 }
4311
4312 static void mem_cgroup_put(struct mem_cgroup *mem)
4313 {
4314         __mem_cgroup_put(mem, 1);
4315 }
4316
4317 /*
4318  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4319  */
4320 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4321 {
4322         if (!mem->res.parent)
4323                 return NULL;
4324         return mem_cgroup_from_res_counter(mem->res.parent, res);
4325 }
4326
4327 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4328 static void __init enable_swap_cgroup(void)
4329 {
4330         if (!mem_cgroup_disabled() && really_do_swap_account)
4331                 do_swap_account = 1;
4332 }
4333 #else
4334 static void __init enable_swap_cgroup(void)
4335 {
4336 }
4337 #endif
4338
4339 static int mem_cgroup_soft_limit_tree_init(void)
4340 {
4341         struct mem_cgroup_tree_per_node *rtpn;
4342         struct mem_cgroup_tree_per_zone *rtpz;
4343         int tmp, node, zone;
4344
4345         for_each_node_state(node, N_POSSIBLE) {
4346                 tmp = node;
4347                 if (!node_state(node, N_NORMAL_MEMORY))
4348                         tmp = -1;
4349                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4350                 if (!rtpn)
4351                         return 1;
4352
4353                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4354
4355                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4356                         rtpz = &rtpn->rb_tree_per_zone[zone];
4357                         rtpz->rb_root = RB_ROOT;
4358                         spin_lock_init(&rtpz->lock);
4359                 }
4360         }
4361         return 0;
4362 }
4363
4364 static struct cgroup_subsys_state * __ref
4365 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4366 {
4367         struct mem_cgroup *mem, *parent;
4368         long error = -ENOMEM;
4369         int node;
4370
4371         mem = mem_cgroup_alloc();
4372         if (!mem)
4373                 return ERR_PTR(error);
4374
4375         for_each_node_state(node, N_POSSIBLE)
4376                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4377                         goto free_out;
4378
4379         /* root ? */
4380         if (cont->parent == NULL) {
4381                 int cpu;
4382                 enable_swap_cgroup();
4383                 parent = NULL;
4384                 root_mem_cgroup = mem;
4385                 if (mem_cgroup_soft_limit_tree_init())
4386                         goto free_out;
4387                 for_each_possible_cpu(cpu) {
4388                         struct memcg_stock_pcp *stock =
4389                                                 &per_cpu(memcg_stock, cpu);
4390                         INIT_WORK(&stock->work, drain_local_stock);
4391                 }
4392                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4393         } else {
4394                 parent = mem_cgroup_from_cont(cont->parent);
4395                 mem->use_hierarchy = parent->use_hierarchy;
4396                 mem->oom_kill_disable = parent->oom_kill_disable;
4397         }
4398
4399         if (parent && parent->use_hierarchy) {
4400                 res_counter_init(&mem->res, &parent->res);
4401                 res_counter_init(&mem->memsw, &parent->memsw);
4402                 /*
4403                  * We increment refcnt of the parent to ensure that we can
4404                  * safely access it on res_counter_charge/uncharge.
4405                  * This refcnt will be decremented when freeing this
4406                  * mem_cgroup(see mem_cgroup_put).
4407                  */
4408                 mem_cgroup_get(parent);
4409         } else {
4410                 res_counter_init(&mem->res, NULL);
4411                 res_counter_init(&mem->memsw, NULL);
4412         }
4413         mem->last_scanned_child = 0;
4414         spin_lock_init(&mem->reclaim_param_lock);
4415         INIT_LIST_HEAD(&mem->oom_notify);
4416
4417         if (parent)
4418                 mem->swappiness = get_swappiness(parent);
4419         atomic_set(&mem->refcnt, 1);
4420         mem->move_charge_at_immigrate = 0;
4421         mutex_init(&mem->thresholds_lock);
4422         return &mem->css;
4423 free_out:
4424         __mem_cgroup_free(mem);
4425         root_mem_cgroup = NULL;
4426         return ERR_PTR(error);
4427 }
4428
4429 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4430                                         struct cgroup *cont)
4431 {
4432         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4433
4434         return mem_cgroup_force_empty(mem, false);
4435 }
4436
4437 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4438                                 struct cgroup *cont)
4439 {
4440         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4441
4442         mem_cgroup_put(mem);
4443 }
4444
4445 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4446                                 struct cgroup *cont)
4447 {
4448         int ret;
4449
4450         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4451                                 ARRAY_SIZE(mem_cgroup_files));
4452
4453         if (!ret)
4454                 ret = register_memsw_files(cont, ss);
4455         return ret;
4456 }
4457
4458 #ifdef CONFIG_MMU
4459 /* Handlers for move charge at task migration. */
4460 #define PRECHARGE_COUNT_AT_ONCE 256
4461 static int mem_cgroup_do_precharge(unsigned long count)
4462 {
4463         int ret = 0;
4464         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4465         struct mem_cgroup *mem = mc.to;
4466
4467         if (mem_cgroup_is_root(mem)) {
4468                 mc.precharge += count;
4469                 /* we don't need css_get for root */
4470                 return ret;
4471         }
4472         /* try to charge at once */
4473         if (count > 1) {
4474                 struct res_counter *dummy;
4475                 /*
4476                  * "mem" cannot be under rmdir() because we've already checked
4477                  * by cgroup_lock_live_cgroup() that it is not removed and we
4478                  * are still under the same cgroup_mutex. So we can postpone
4479                  * css_get().
4480                  */
4481                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4482                         goto one_by_one;
4483                 if (do_swap_account && res_counter_charge(&mem->memsw,
4484                                                 PAGE_SIZE * count, &dummy)) {
4485                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4486                         goto one_by_one;
4487                 }
4488                 mc.precharge += count;
4489                 return ret;
4490         }
4491 one_by_one:
4492         /* fall back to one by one charge */
4493         while (count--) {
4494                 if (signal_pending(current)) {
4495                         ret = -EINTR;
4496                         break;
4497                 }
4498                 if (!batch_count--) {
4499                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4500                         cond_resched();
4501                 }
4502                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4503                                               PAGE_SIZE);
4504                 if (ret || !mem)
4505                         /* mem_cgroup_clear_mc() will do uncharge later */
4506                         return -ENOMEM;
4507                 mc.precharge++;
4508         }
4509         return ret;
4510 }
4511
4512 /**
4513  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4514  * @vma: the vma the pte to be checked belongs
4515  * @addr: the address corresponding to the pte to be checked
4516  * @ptent: the pte to be checked
4517  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4518  *
4519  * Returns
4520  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4521  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4522  *     move charge. if @target is not NULL, the page is stored in target->page
4523  *     with extra refcnt got(Callers should handle it).
4524  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4525  *     target for charge migration. if @target is not NULL, the entry is stored
4526  *     in target->ent.
4527  *
4528  * Called with pte lock held.
4529  */
4530 union mc_target {
4531         struct page     *page;
4532         swp_entry_t     ent;
4533 };
4534
4535 enum mc_target_type {
4536         MC_TARGET_NONE, /* not used */
4537         MC_TARGET_PAGE,
4538         MC_TARGET_SWAP,
4539 };
4540
4541 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4542                                                 unsigned long addr, pte_t ptent)
4543 {
4544         struct page *page = vm_normal_page(vma, addr, ptent);
4545
4546         if (!page || !page_mapped(page))
4547                 return NULL;
4548         if (PageAnon(page)) {
4549                 /* we don't move shared anon */
4550                 if (!move_anon() || page_mapcount(page) > 2)
4551                         return NULL;
4552         } else if (!move_file())
4553                 /* we ignore mapcount for file pages */
4554                 return NULL;
4555         if (!get_page_unless_zero(page))
4556                 return NULL;
4557
4558         return page;
4559 }
4560
4561 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4562                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4563 {
4564         int usage_count;
4565         struct page *page = NULL;
4566         swp_entry_t ent = pte_to_swp_entry(ptent);
4567
4568         if (!move_anon() || non_swap_entry(ent))
4569                 return NULL;
4570         usage_count = mem_cgroup_count_swap_user(ent, &page);
4571         if (usage_count > 1) { /* we don't move shared anon */
4572                 if (page)
4573                         put_page(page);
4574                 return NULL;
4575         }
4576         if (do_swap_account)
4577                 entry->val = ent.val;
4578
4579         return page;
4580 }
4581
4582 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4583                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4584 {
4585         struct page *page = NULL;
4586         struct inode *inode;
4587         struct address_space *mapping;
4588         pgoff_t pgoff;
4589
4590         if (!vma->vm_file) /* anonymous vma */
4591                 return NULL;
4592         if (!move_file())
4593                 return NULL;
4594
4595         inode = vma->vm_file->f_path.dentry->d_inode;
4596         mapping = vma->vm_file->f_mapping;
4597         if (pte_none(ptent))
4598                 pgoff = linear_page_index(vma, addr);
4599         else /* pte_file(ptent) is true */
4600                 pgoff = pte_to_pgoff(ptent);
4601
4602         /* page is moved even if it's not RSS of this task(page-faulted). */
4603         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4604                 page = find_get_page(mapping, pgoff);
4605         } else { /* shmem/tmpfs file. we should take account of swap too. */
4606                 swp_entry_t ent;
4607                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4608                 if (do_swap_account)
4609                         entry->val = ent.val;
4610         }
4611
4612         return page;
4613 }
4614
4615 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4616                 unsigned long addr, pte_t ptent, union mc_target *target)
4617 {
4618         struct page *page = NULL;
4619         struct page_cgroup *pc;
4620         int ret = 0;
4621         swp_entry_t ent = { .val = 0 };
4622
4623         if (pte_present(ptent))
4624                 page = mc_handle_present_pte(vma, addr, ptent);
4625         else if (is_swap_pte(ptent))
4626                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4627         else if (pte_none(ptent) || pte_file(ptent))
4628                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4629
4630         if (!page && !ent.val)
4631                 return 0;
4632         if (page) {
4633                 pc = lookup_page_cgroup(page);
4634                 /*
4635                  * Do only loose check w/o page_cgroup lock.
4636                  * mem_cgroup_move_account() checks the pc is valid or not under
4637                  * the lock.
4638                  */
4639                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4640                         ret = MC_TARGET_PAGE;
4641                         if (target)
4642                                 target->page = page;
4643                 }
4644                 if (!ret || !target)
4645                         put_page(page);
4646         }
4647         /* There is a swap entry and a page doesn't exist or isn't charged */
4648         if (ent.val && !ret &&
4649                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4650                 ret = MC_TARGET_SWAP;
4651                 if (target)
4652                         target->ent = ent;
4653         }
4654         return ret;
4655 }
4656
4657 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4658                                         unsigned long addr, unsigned long end,
4659                                         struct mm_walk *walk)
4660 {
4661         struct vm_area_struct *vma = walk->private;
4662         pte_t *pte;
4663         spinlock_t *ptl;
4664
4665         VM_BUG_ON(pmd_trans_huge(*pmd));
4666         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4667         for (; addr != end; pte++, addr += PAGE_SIZE)
4668                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4669                         mc.precharge++; /* increment precharge temporarily */
4670         pte_unmap_unlock(pte - 1, ptl);
4671         cond_resched();
4672
4673         return 0;
4674 }
4675
4676 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4677 {
4678         unsigned long precharge;
4679         struct vm_area_struct *vma;
4680
4681         /* We've already held the mmap_sem */
4682         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4683                 struct mm_walk mem_cgroup_count_precharge_walk = {
4684                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4685                         .mm = mm,
4686                         .private = vma,
4687                 };
4688                 if (is_vm_hugetlb_page(vma))
4689                         continue;
4690                 walk_page_range(vma->vm_start, vma->vm_end,
4691                                         &mem_cgroup_count_precharge_walk);
4692         }
4693
4694         precharge = mc.precharge;
4695         mc.precharge = 0;
4696
4697         return precharge;
4698 }
4699
4700 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4701 {
4702         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4703 }
4704
4705 static void mem_cgroup_clear_mc(void)
4706 {
4707         struct mem_cgroup *from = mc.from;
4708         struct mem_cgroup *to = mc.to;
4709
4710         /* we must uncharge all the leftover precharges from mc.to */
4711         if (mc.precharge) {
4712                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4713                 mc.precharge = 0;
4714         }
4715         /*
4716          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4717          * we must uncharge here.
4718          */
4719         if (mc.moved_charge) {
4720                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4721                 mc.moved_charge = 0;
4722         }
4723         /* we must fixup refcnts and charges */
4724         if (mc.moved_swap) {
4725                 /* uncharge swap account from the old cgroup */
4726                 if (!mem_cgroup_is_root(mc.from))
4727                         res_counter_uncharge(&mc.from->memsw,
4728                                                 PAGE_SIZE * mc.moved_swap);
4729                 __mem_cgroup_put(mc.from, mc.moved_swap);
4730
4731                 if (!mem_cgroup_is_root(mc.to)) {
4732                         /*
4733                          * we charged both to->res and to->memsw, so we should
4734                          * uncharge to->res.
4735                          */
4736                         res_counter_uncharge(&mc.to->res,
4737                                                 PAGE_SIZE * mc.moved_swap);
4738                 }
4739                 /* we've already done mem_cgroup_get(mc.to) */
4740
4741                 mc.moved_swap = 0;
4742         }
4743         if (mc.mm) {
4744                 up_read(&mc.mm->mmap_sem);
4745                 mmput(mc.mm);
4746         }
4747         spin_lock(&mc.lock);
4748         mc.from = NULL;
4749         mc.to = NULL;
4750         spin_unlock(&mc.lock);
4751         mc.moving_task = NULL;
4752         mc.mm = NULL;
4753         mem_cgroup_end_move(from);
4754         memcg_oom_recover(from);
4755         memcg_oom_recover(to);
4756         wake_up_all(&mc.waitq);
4757 }
4758
4759 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4760                                 struct cgroup *cgroup,
4761                                 struct task_struct *p,
4762                                 bool threadgroup)
4763 {
4764         int ret = 0;
4765         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4766
4767         if (mem->move_charge_at_immigrate) {
4768                 struct mm_struct *mm;
4769                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4770
4771                 VM_BUG_ON(from == mem);
4772
4773                 mm = get_task_mm(p);
4774                 if (!mm)
4775                         return 0;
4776                 /* We move charges only when we move a owner of the mm */
4777                 if (mm->owner == p) {
4778                         /*
4779                          * We do all the move charge works under one mmap_sem to
4780                          * avoid deadlock with down_write(&mmap_sem)
4781                          * -> try_charge() -> if (mc.moving_task) -> sleep.
4782                          */
4783                         down_read(&mm->mmap_sem);
4784
4785                         VM_BUG_ON(mc.from);
4786                         VM_BUG_ON(mc.to);
4787                         VM_BUG_ON(mc.precharge);
4788                         VM_BUG_ON(mc.moved_charge);
4789                         VM_BUG_ON(mc.moved_swap);
4790                         VM_BUG_ON(mc.moving_task);
4791                         VM_BUG_ON(mc.mm);
4792
4793                         mem_cgroup_start_move(from);
4794                         spin_lock(&mc.lock);
4795                         mc.from = from;
4796                         mc.to = mem;
4797                         mc.precharge = 0;
4798                         mc.moved_charge = 0;
4799                         mc.moved_swap = 0;
4800                         spin_unlock(&mc.lock);
4801                         mc.moving_task = current;
4802                         mc.mm = mm;
4803
4804                         ret = mem_cgroup_precharge_mc(mm);
4805                         if (ret)
4806                                 mem_cgroup_clear_mc();
4807                         /* We call up_read() and mmput() in clear_mc(). */
4808                 } else
4809                         mmput(mm);
4810         }
4811         return ret;
4812 }
4813
4814 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4815                                 struct cgroup *cgroup,
4816                                 struct task_struct *p,
4817                                 bool threadgroup)
4818 {
4819         mem_cgroup_clear_mc();
4820 }
4821
4822 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4823                                 unsigned long addr, unsigned long end,
4824                                 struct mm_walk *walk)
4825 {
4826         int ret = 0;
4827         struct vm_area_struct *vma = walk->private;
4828         pte_t *pte;
4829         spinlock_t *ptl;
4830
4831 retry:
4832         VM_BUG_ON(pmd_trans_huge(*pmd));
4833         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4834         for (; addr != end; addr += PAGE_SIZE) {
4835                 pte_t ptent = *(pte++);
4836                 union mc_target target;
4837                 int type;
4838                 struct page *page;
4839                 struct page_cgroup *pc;
4840                 swp_entry_t ent;
4841
4842                 if (!mc.precharge)
4843                         break;
4844
4845                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4846                 switch (type) {
4847                 case MC_TARGET_PAGE:
4848                         page = target.page;
4849                         if (isolate_lru_page(page))
4850                                 goto put;
4851                         pc = lookup_page_cgroup(page);
4852                         if (!mem_cgroup_move_account(pc,
4853                                                 mc.from, mc.to, false)) {
4854                                 mc.precharge--;
4855                                 /* we uncharge from mc.from later. */
4856                                 mc.moved_charge++;
4857                         }
4858                         putback_lru_page(page);
4859 put:                    /* is_target_pte_for_mc() gets the page */
4860                         put_page(page);
4861                         break;
4862                 case MC_TARGET_SWAP:
4863                         ent = target.ent;
4864                         if (!mem_cgroup_move_swap_account(ent,
4865                                                 mc.from, mc.to, false)) {
4866                                 mc.precharge--;
4867                                 /* we fixup refcnts and charges later. */
4868                                 mc.moved_swap++;
4869                         }
4870                         break;
4871                 default:
4872                         break;
4873                 }
4874         }
4875         pte_unmap_unlock(pte - 1, ptl);
4876         cond_resched();
4877
4878         if (addr != end) {
4879                 /*
4880                  * We have consumed all precharges we got in can_attach().
4881                  * We try charge one by one, but don't do any additional
4882                  * charges to mc.to if we have failed in charge once in attach()
4883                  * phase.
4884                  */
4885                 ret = mem_cgroup_do_precharge(1);
4886                 if (!ret)
4887                         goto retry;
4888         }
4889
4890         return ret;
4891 }
4892
4893 static void mem_cgroup_move_charge(struct mm_struct *mm)
4894 {
4895         struct vm_area_struct *vma;
4896
4897         lru_add_drain_all();
4898         /* We've already held the mmap_sem */
4899         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4900                 int ret;
4901                 struct mm_walk mem_cgroup_move_charge_walk = {
4902                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4903                         .mm = mm,
4904                         .private = vma,
4905                 };
4906                 if (is_vm_hugetlb_page(vma))
4907                         continue;
4908                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4909                                                 &mem_cgroup_move_charge_walk);
4910                 if (ret)
4911                         /*
4912                          * means we have consumed all precharges and failed in
4913                          * doing additional charge. Just abandon here.
4914                          */
4915                         break;
4916         }
4917 }
4918
4919 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4920                                 struct cgroup *cont,
4921                                 struct cgroup *old_cont,
4922                                 struct task_struct *p,
4923                                 bool threadgroup)
4924 {
4925         if (!mc.mm)
4926                 /* no need to move charge */
4927                 return;
4928
4929         mem_cgroup_move_charge(mc.mm);
4930         mem_cgroup_clear_mc();
4931 }
4932 #else   /* !CONFIG_MMU */
4933 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4934                                 struct cgroup *cgroup,
4935                                 struct task_struct *p,
4936                                 bool threadgroup)
4937 {
4938         return 0;
4939 }
4940 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4941                                 struct cgroup *cgroup,
4942                                 struct task_struct *p,
4943                                 bool threadgroup)
4944 {
4945 }
4946 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4947                                 struct cgroup *cont,
4948                                 struct cgroup *old_cont,
4949                                 struct task_struct *p,
4950                                 bool threadgroup)
4951 {
4952 }
4953 #endif
4954
4955 struct cgroup_subsys mem_cgroup_subsys = {
4956         .name = "memory",
4957         .subsys_id = mem_cgroup_subsys_id,
4958         .create = mem_cgroup_create,
4959         .pre_destroy = mem_cgroup_pre_destroy,
4960         .destroy = mem_cgroup_destroy,
4961         .populate = mem_cgroup_populate,
4962         .can_attach = mem_cgroup_can_attach,
4963         .cancel_attach = mem_cgroup_cancel_attach,
4964         .attach = mem_cgroup_move_task,
4965         .early_init = 0,
4966         .use_id = 1,
4967 };
4968
4969 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4970 static int __init enable_swap_account(char *s)
4971 {
4972         /* consider enabled if no parameter or 1 is given */
4973         if (!s || !strcmp(s, "1"))
4974                 really_do_swap_account = 1;
4975         else if (!strcmp(s, "0"))
4976                 really_do_swap_account = 0;
4977         return 1;
4978 }
4979 __setup("swapaccount", enable_swap_account);
4980
4981 static int __init disable_swap_account(char *s)
4982 {
4983         enable_swap_account("0");
4984         return 1;
4985 }
4986 __setup("noswapaccount", disable_swap_account);
4987 #endif