afde618f98952fa74f1324fda45487aeda4cef55
[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 /*
78  * Statistics for memory cgroup.
79  */
80 enum mem_cgroup_stat_index {
81         /*
82          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
83          */
84         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
85         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
86         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
87         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
88         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
89         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
90         MEM_CGROUP_STAT_NSTATS,
91 };
92
93 enum mem_cgroup_events_index {
94         MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
95         MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
96         MEM_CGROUP_EVENTS_COUNT,        /* # of pages paged in/out */
97         MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
98         MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
99         MEM_CGROUP_EVENTS_NSTATS,
100 };
101 /*
102  * Per memcg event counter is incremented at every pagein/pageout. With THP,
103  * it will be incremated by the number of pages. This counter is used for
104  * for trigger some periodic events. This is straightforward and better
105  * than using jiffies etc. to handle periodic memcg event.
106  */
107 enum mem_cgroup_events_target {
108         MEM_CGROUP_TARGET_THRESH,
109         MEM_CGROUP_TARGET_SOFTLIMIT,
110         MEM_CGROUP_TARGET_NUMAINFO,
111         MEM_CGROUP_NTARGETS,
112 };
113 #define THRESHOLDS_EVENTS_TARGET (128)
114 #define SOFTLIMIT_EVENTS_TARGET (1024)
115 #define NUMAINFO_EVENTS_TARGET  (1024)
116
117 struct mem_cgroup_stat_cpu {
118         long count[MEM_CGROUP_STAT_NSTATS];
119         unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
120         unsigned long targets[MEM_CGROUP_NTARGETS];
121 };
122
123 /*
124  * per-zone information in memory controller.
125  */
126 struct mem_cgroup_per_zone {
127         /*
128          * spin_lock to protect the per cgroup LRU
129          */
130         struct list_head        lists[NR_LRU_LISTS];
131         unsigned long           count[NR_LRU_LISTS];
132
133         struct zone_reclaim_stat reclaim_stat;
134         struct rb_node          tree_node;      /* RB tree node */
135         unsigned long long      usage_in_excess;/* Set to the value by which */
136                                                 /* the soft limit is exceeded*/
137         bool                    on_tree;
138         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
139                                                 /* use container_of        */
140 };
141 /* Macro for accessing counter */
142 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
143
144 struct mem_cgroup_per_node {
145         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
146 };
147
148 struct mem_cgroup_lru_info {
149         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
150 };
151
152 /*
153  * Cgroups above their limits are maintained in a RB-Tree, independent of
154  * their hierarchy representation
155  */
156
157 struct mem_cgroup_tree_per_zone {
158         struct rb_root rb_root;
159         spinlock_t lock;
160 };
161
162 struct mem_cgroup_tree_per_node {
163         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
164 };
165
166 struct mem_cgroup_tree {
167         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
168 };
169
170 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
171
172 struct mem_cgroup_threshold {
173         struct eventfd_ctx *eventfd;
174         u64 threshold;
175 };
176
177 /* For threshold */
178 struct mem_cgroup_threshold_ary {
179         /* An array index points to threshold just below usage. */
180         int current_threshold;
181         /* Size of entries[] */
182         unsigned int size;
183         /* Array of thresholds */
184         struct mem_cgroup_threshold entries[0];
185 };
186
187 struct mem_cgroup_thresholds {
188         /* Primary thresholds array */
189         struct mem_cgroup_threshold_ary *primary;
190         /*
191          * Spare threshold array.
192          * This is needed to make mem_cgroup_unregister_event() "never fail".
193          * It must be able to store at least primary->size - 1 entries.
194          */
195         struct mem_cgroup_threshold_ary *spare;
196 };
197
198 /* for OOM */
199 struct mem_cgroup_eventfd_list {
200         struct list_head list;
201         struct eventfd_ctx *eventfd;
202 };
203
204 static void mem_cgroup_threshold(struct mem_cgroup *mem);
205 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
206
207 /*
208  * The memory controller data structure. The memory controller controls both
209  * page cache and RSS per cgroup. We would eventually like to provide
210  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
211  * to help the administrator determine what knobs to tune.
212  *
213  * TODO: Add a water mark for the memory controller. Reclaim will begin when
214  * we hit the water mark. May be even add a low water mark, such that
215  * no reclaim occurs from a cgroup at it's low water mark, this is
216  * a feature that will be implemented much later in the future.
217  */
218 struct mem_cgroup {
219         struct cgroup_subsys_state css;
220         /*
221          * the counter to account for memory usage
222          */
223         struct res_counter res;
224         /*
225          * the counter to account for mem+swap usage.
226          */
227         struct res_counter memsw;
228         /*
229          * Per cgroup active and inactive list, similar to the
230          * per zone LRU lists.
231          */
232         struct mem_cgroup_lru_info info;
233         /*
234          * While reclaiming in a hierarchy, we cache the last child we
235          * reclaimed from.
236          */
237         int last_scanned_child;
238         int last_scanned_node;
239 #if MAX_NUMNODES > 1
240         nodemask_t      scan_nodes;
241         atomic_t        numainfo_events;
242         atomic_t        numainfo_updating;
243 #endif
244         /*
245          * Should the accounting and control be hierarchical, per subtree?
246          */
247         bool use_hierarchy;
248
249         bool            oom_lock;
250         atomic_t        under_oom;
251
252         atomic_t        refcnt;
253
254         int     swappiness;
255         /* OOM-Killer disable */
256         int             oom_kill_disable;
257
258         /* set when res.limit == memsw.limit */
259         bool            memsw_is_minimum;
260
261         /* protect arrays of thresholds */
262         struct mutex thresholds_lock;
263
264         /* thresholds for memory usage. RCU-protected */
265         struct mem_cgroup_thresholds thresholds;
266
267         /* thresholds for mem+swap usage. RCU-protected */
268         struct mem_cgroup_thresholds memsw_thresholds;
269
270         /* For oom notifier event fd */
271         struct list_head oom_notify;
272
273         /*
274          * Should we move charges of a task when a task is moved into this
275          * mem_cgroup ? And what type of charges should we move ?
276          */
277         unsigned long   move_charge_at_immigrate;
278         /*
279          * percpu counter.
280          */
281         struct mem_cgroup_stat_cpu *stat;
282         /*
283          * used when a cpu is offlined or other synchronizations
284          * See mem_cgroup_read_stat().
285          */
286         struct mem_cgroup_stat_cpu nocpu_base;
287         spinlock_t pcp_counter_lock;
288 };
289
290 /* Stuffs for move charges at task migration. */
291 /*
292  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
293  * left-shifted bitmap of these types.
294  */
295 enum move_type {
296         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
297         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
298         NR_MOVE_TYPE,
299 };
300
301 /* "mc" and its members are protected by cgroup_mutex */
302 static struct move_charge_struct {
303         spinlock_t        lock; /* for from, to */
304         struct mem_cgroup *from;
305         struct mem_cgroup *to;
306         unsigned long precharge;
307         unsigned long moved_charge;
308         unsigned long moved_swap;
309         struct task_struct *moving_task;        /* a task moving charges */
310         wait_queue_head_t waitq;                /* a waitq for other context */
311 } mc = {
312         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
313         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
314 };
315
316 static bool move_anon(void)
317 {
318         return test_bit(MOVE_CHARGE_TYPE_ANON,
319                                         &mc.to->move_charge_at_immigrate);
320 }
321
322 static bool move_file(void)
323 {
324         return test_bit(MOVE_CHARGE_TYPE_FILE,
325                                         &mc.to->move_charge_at_immigrate);
326 }
327
328 /*
329  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
330  * limit reclaim to prevent infinite loops, if they ever occur.
331  */
332 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
333 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
334
335 enum charge_type {
336         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
337         MEM_CGROUP_CHARGE_TYPE_MAPPED,
338         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
339         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
340         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
341         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
342         NR_CHARGE_TYPE,
343 };
344
345 /* for encoding cft->private value on file */
346 #define _MEM                    (0)
347 #define _MEMSWAP                (1)
348 #define _OOM_TYPE               (2)
349 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
350 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
351 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
352 /* Used for OOM nofiier */
353 #define OOM_CONTROL             (0)
354
355 /*
356  * Reclaim flags for mem_cgroup_hierarchical_reclaim
357  */
358 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
359 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
360 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
361 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
362 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
363 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
364
365 static void mem_cgroup_get(struct mem_cgroup *mem);
366 static void mem_cgroup_put(struct mem_cgroup *mem);
367 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
368 static void drain_all_stock_async(struct mem_cgroup *mem);
369
370 static struct mem_cgroup_per_zone *
371 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
372 {
373         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
374 }
375
376 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
377 {
378         return &mem->css;
379 }
380
381 static struct mem_cgroup_per_zone *
382 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
383 {
384         int nid = page_to_nid(page);
385         int zid = page_zonenum(page);
386
387         return mem_cgroup_zoneinfo(mem, nid, zid);
388 }
389
390 static struct mem_cgroup_tree_per_zone *
391 soft_limit_tree_node_zone(int nid, int zid)
392 {
393         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
394 }
395
396 static struct mem_cgroup_tree_per_zone *
397 soft_limit_tree_from_page(struct page *page)
398 {
399         int nid = page_to_nid(page);
400         int zid = page_zonenum(page);
401
402         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
403 }
404
405 static void
406 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
407                                 struct mem_cgroup_per_zone *mz,
408                                 struct mem_cgroup_tree_per_zone *mctz,
409                                 unsigned long long new_usage_in_excess)
410 {
411         struct rb_node **p = &mctz->rb_root.rb_node;
412         struct rb_node *parent = NULL;
413         struct mem_cgroup_per_zone *mz_node;
414
415         if (mz->on_tree)
416                 return;
417
418         mz->usage_in_excess = new_usage_in_excess;
419         if (!mz->usage_in_excess)
420                 return;
421         while (*p) {
422                 parent = *p;
423                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
424                                         tree_node);
425                 if (mz->usage_in_excess < mz_node->usage_in_excess)
426                         p = &(*p)->rb_left;
427                 /*
428                  * We can't avoid mem cgroups that are over their soft
429                  * limit by the same amount
430                  */
431                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
432                         p = &(*p)->rb_right;
433         }
434         rb_link_node(&mz->tree_node, parent, p);
435         rb_insert_color(&mz->tree_node, &mctz->rb_root);
436         mz->on_tree = true;
437 }
438
439 static void
440 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
441                                 struct mem_cgroup_per_zone *mz,
442                                 struct mem_cgroup_tree_per_zone *mctz)
443 {
444         if (!mz->on_tree)
445                 return;
446         rb_erase(&mz->tree_node, &mctz->rb_root);
447         mz->on_tree = false;
448 }
449
450 static void
451 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
452                                 struct mem_cgroup_per_zone *mz,
453                                 struct mem_cgroup_tree_per_zone *mctz)
454 {
455         spin_lock(&mctz->lock);
456         __mem_cgroup_remove_exceeded(mem, mz, mctz);
457         spin_unlock(&mctz->lock);
458 }
459
460
461 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
462 {
463         unsigned long long excess;
464         struct mem_cgroup_per_zone *mz;
465         struct mem_cgroup_tree_per_zone *mctz;
466         int nid = page_to_nid(page);
467         int zid = page_zonenum(page);
468         mctz = soft_limit_tree_from_page(page);
469
470         /*
471          * Necessary to update all ancestors when hierarchy is used.
472          * because their event counter is not touched.
473          */
474         for (; mem; mem = parent_mem_cgroup(mem)) {
475                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
476                 excess = res_counter_soft_limit_excess(&mem->res);
477                 /*
478                  * We have to update the tree if mz is on RB-tree or
479                  * mem is over its softlimit.
480                  */
481                 if (excess || mz->on_tree) {
482                         spin_lock(&mctz->lock);
483                         /* if on-tree, remove it */
484                         if (mz->on_tree)
485                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
486                         /*
487                          * Insert again. mz->usage_in_excess will be updated.
488                          * If excess is 0, no tree ops.
489                          */
490                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
491                         spin_unlock(&mctz->lock);
492                 }
493         }
494 }
495
496 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
497 {
498         int node, zone;
499         struct mem_cgroup_per_zone *mz;
500         struct mem_cgroup_tree_per_zone *mctz;
501
502         for_each_node_state(node, N_POSSIBLE) {
503                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
504                         mz = mem_cgroup_zoneinfo(mem, node, zone);
505                         mctz = soft_limit_tree_node_zone(node, zone);
506                         mem_cgroup_remove_exceeded(mem, mz, mctz);
507                 }
508         }
509 }
510
511 static struct mem_cgroup_per_zone *
512 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
513 {
514         struct rb_node *rightmost = NULL;
515         struct mem_cgroup_per_zone *mz;
516
517 retry:
518         mz = NULL;
519         rightmost = rb_last(&mctz->rb_root);
520         if (!rightmost)
521                 goto done;              /* Nothing to reclaim from */
522
523         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
524         /*
525          * Remove the node now but someone else can add it back,
526          * we will to add it back at the end of reclaim to its correct
527          * position in the tree.
528          */
529         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
530         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
531                 !css_tryget(&mz->mem->css))
532                 goto retry;
533 done:
534         return mz;
535 }
536
537 static struct mem_cgroup_per_zone *
538 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
539 {
540         struct mem_cgroup_per_zone *mz;
541
542         spin_lock(&mctz->lock);
543         mz = __mem_cgroup_largest_soft_limit_node(mctz);
544         spin_unlock(&mctz->lock);
545         return mz;
546 }
547
548 /*
549  * Implementation Note: reading percpu statistics for memcg.
550  *
551  * Both of vmstat[] and percpu_counter has threshold and do periodic
552  * synchronization to implement "quick" read. There are trade-off between
553  * reading cost and precision of value. Then, we may have a chance to implement
554  * a periodic synchronizion of counter in memcg's counter.
555  *
556  * But this _read() function is used for user interface now. The user accounts
557  * memory usage by memory cgroup and he _always_ requires exact value because
558  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
559  * have to visit all online cpus and make sum. So, for now, unnecessary
560  * synchronization is not implemented. (just implemented for cpu hotplug)
561  *
562  * If there are kernel internal actions which can make use of some not-exact
563  * value, and reading all cpu value can be performance bottleneck in some
564  * common workload, threashold and synchonization as vmstat[] should be
565  * implemented.
566  */
567 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
568                                  enum mem_cgroup_stat_index idx)
569 {
570         long val = 0;
571         int cpu;
572
573         get_online_cpus();
574         for_each_online_cpu(cpu)
575                 val += per_cpu(mem->stat->count[idx], cpu);
576 #ifdef CONFIG_HOTPLUG_CPU
577         spin_lock(&mem->pcp_counter_lock);
578         val += mem->nocpu_base.count[idx];
579         spin_unlock(&mem->pcp_counter_lock);
580 #endif
581         put_online_cpus();
582         return val;
583 }
584
585 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
586                                          bool charge)
587 {
588         int val = (charge) ? 1 : -1;
589         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
590 }
591
592 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
593 {
594         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
595 }
596
597 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
598 {
599         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
600 }
601
602 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
603                                             enum mem_cgroup_events_index idx)
604 {
605         unsigned long val = 0;
606         int cpu;
607
608         for_each_online_cpu(cpu)
609                 val += per_cpu(mem->stat->events[idx], cpu);
610 #ifdef CONFIG_HOTPLUG_CPU
611         spin_lock(&mem->pcp_counter_lock);
612         val += mem->nocpu_base.events[idx];
613         spin_unlock(&mem->pcp_counter_lock);
614 #endif
615         return val;
616 }
617
618 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
619                                          bool file, int nr_pages)
620 {
621         preempt_disable();
622
623         if (file)
624                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
625         else
626                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
627
628         /* pagein of a big page is an event. So, ignore page size */
629         if (nr_pages > 0)
630                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
631         else {
632                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
633                 nr_pages = -nr_pages; /* for event */
634         }
635
636         __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
637
638         preempt_enable();
639 }
640
641 unsigned long
642 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
643                         unsigned int lru_mask)
644 {
645         struct mem_cgroup_per_zone *mz;
646         enum lru_list l;
647         unsigned long ret = 0;
648
649         mz = mem_cgroup_zoneinfo(mem, nid, zid);
650
651         for_each_lru(l) {
652                 if (BIT(l) & lru_mask)
653                         ret += MEM_CGROUP_ZSTAT(mz, l);
654         }
655         return ret;
656 }
657
658 static unsigned long
659 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
660                         int nid, unsigned int lru_mask)
661 {
662         u64 total = 0;
663         int zid;
664
665         for (zid = 0; zid < MAX_NR_ZONES; zid++)
666                 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
667
668         return total;
669 }
670
671 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
672                         unsigned int lru_mask)
673 {
674         int nid;
675         u64 total = 0;
676
677         for_each_node_state(nid, N_HIGH_MEMORY)
678                 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
679         return total;
680 }
681
682 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
683 {
684         unsigned long val, next;
685
686         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
687         next = this_cpu_read(mem->stat->targets[target]);
688         /* from time_after() in jiffies.h */
689         return ((long)next - (long)val < 0);
690 }
691
692 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
693 {
694         unsigned long val, next;
695
696         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
697
698         switch (target) {
699         case MEM_CGROUP_TARGET_THRESH:
700                 next = val + THRESHOLDS_EVENTS_TARGET;
701                 break;
702         case MEM_CGROUP_TARGET_SOFTLIMIT:
703                 next = val + SOFTLIMIT_EVENTS_TARGET;
704                 break;
705         case MEM_CGROUP_TARGET_NUMAINFO:
706                 next = val + NUMAINFO_EVENTS_TARGET;
707                 break;
708         default:
709                 return;
710         }
711
712         this_cpu_write(mem->stat->targets[target], next);
713 }
714
715 /*
716  * Check events in order.
717  *
718  */
719 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
720 {
721         /* threshold event is triggered in finer grain than soft limit */
722         if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
723                 mem_cgroup_threshold(mem);
724                 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
725                 if (unlikely(__memcg_event_check(mem,
726                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
727                         mem_cgroup_update_tree(mem, page);
728                         __mem_cgroup_target_update(mem,
729                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
730                 }
731 #if MAX_NUMNODES > 1
732                 if (unlikely(__memcg_event_check(mem,
733                         MEM_CGROUP_TARGET_NUMAINFO))) {
734                         atomic_inc(&mem->numainfo_events);
735                         __mem_cgroup_target_update(mem,
736                                 MEM_CGROUP_TARGET_NUMAINFO);
737                 }
738 #endif
739         }
740 }
741
742 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
743 {
744         return container_of(cgroup_subsys_state(cont,
745                                 mem_cgroup_subsys_id), struct mem_cgroup,
746                                 css);
747 }
748
749 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
750 {
751         /*
752          * mm_update_next_owner() may clear mm->owner to NULL
753          * if it races with swapoff, page migration, etc.
754          * So this can be called with p == NULL.
755          */
756         if (unlikely(!p))
757                 return NULL;
758
759         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
760                                 struct mem_cgroup, css);
761 }
762
763 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
764 {
765         struct mem_cgroup *mem = NULL;
766
767         if (!mm)
768                 return NULL;
769         /*
770          * Because we have no locks, mm->owner's may be being moved to other
771          * cgroup. We use css_tryget() here even if this looks
772          * pessimistic (rather than adding locks here).
773          */
774         rcu_read_lock();
775         do {
776                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
777                 if (unlikely(!mem))
778                         break;
779         } while (!css_tryget(&mem->css));
780         rcu_read_unlock();
781         return mem;
782 }
783
784 /* The caller has to guarantee "mem" exists before calling this */
785 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
786 {
787         struct cgroup_subsys_state *css;
788         int found;
789
790         if (!mem) /* ROOT cgroup has the smallest ID */
791                 return root_mem_cgroup; /*css_put/get against root is ignored*/
792         if (!mem->use_hierarchy) {
793                 if (css_tryget(&mem->css))
794                         return mem;
795                 return NULL;
796         }
797         rcu_read_lock();
798         /*
799          * searching a memory cgroup which has the smallest ID under given
800          * ROOT cgroup. (ID >= 1)
801          */
802         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
803         if (css && css_tryget(css))
804                 mem = container_of(css, struct mem_cgroup, css);
805         else
806                 mem = NULL;
807         rcu_read_unlock();
808         return mem;
809 }
810
811 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
812                                         struct mem_cgroup *root,
813                                         bool cond)
814 {
815         int nextid = css_id(&iter->css) + 1;
816         int found;
817         int hierarchy_used;
818         struct cgroup_subsys_state *css;
819
820         hierarchy_used = iter->use_hierarchy;
821
822         css_put(&iter->css);
823         /* If no ROOT, walk all, ignore hierarchy */
824         if (!cond || (root && !hierarchy_used))
825                 return NULL;
826
827         if (!root)
828                 root = root_mem_cgroup;
829
830         do {
831                 iter = NULL;
832                 rcu_read_lock();
833
834                 css = css_get_next(&mem_cgroup_subsys, nextid,
835                                 &root->css, &found);
836                 if (css && css_tryget(css))
837                         iter = container_of(css, struct mem_cgroup, css);
838                 rcu_read_unlock();
839                 /* If css is NULL, no more cgroups will be found */
840                 nextid = found + 1;
841         } while (css && !iter);
842
843         return iter;
844 }
845 /*
846  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
847  * be careful that "break" loop is not allowed. We have reference count.
848  * Instead of that modify "cond" to be false and "continue" to exit the loop.
849  */
850 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
851         for (iter = mem_cgroup_start_loop(root);\
852              iter != NULL;\
853              iter = mem_cgroup_get_next(iter, root, cond))
854
855 #define for_each_mem_cgroup_tree(iter, root) \
856         for_each_mem_cgroup_tree_cond(iter, root, true)
857
858 #define for_each_mem_cgroup_all(iter) \
859         for_each_mem_cgroup_tree_cond(iter, NULL, true)
860
861
862 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
863 {
864         return (mem == root_mem_cgroup);
865 }
866
867 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
868 {
869         struct mem_cgroup *mem;
870
871         if (!mm)
872                 return;
873
874         rcu_read_lock();
875         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
876         if (unlikely(!mem))
877                 goto out;
878
879         switch (idx) {
880         case PGMAJFAULT:
881                 mem_cgroup_pgmajfault(mem, 1);
882                 break;
883         case PGFAULT:
884                 mem_cgroup_pgfault(mem, 1);
885                 break;
886         default:
887                 BUG();
888         }
889 out:
890         rcu_read_unlock();
891 }
892 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
893
894 /*
895  * Following LRU functions are allowed to be used without PCG_LOCK.
896  * Operations are called by routine of global LRU independently from memcg.
897  * What we have to take care of here is validness of pc->mem_cgroup.
898  *
899  * Changes to pc->mem_cgroup happens when
900  * 1. charge
901  * 2. moving account
902  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
903  * It is added to LRU before charge.
904  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
905  * When moving account, the page is not on LRU. It's isolated.
906  */
907
908 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
909 {
910         struct page_cgroup *pc;
911         struct mem_cgroup_per_zone *mz;
912
913         if (mem_cgroup_disabled())
914                 return;
915         pc = lookup_page_cgroup(page);
916         /* can happen while we handle swapcache. */
917         if (!TestClearPageCgroupAcctLRU(pc))
918                 return;
919         VM_BUG_ON(!pc->mem_cgroup);
920         /*
921          * We don't check PCG_USED bit. It's cleared when the "page" is finally
922          * removed from global LRU.
923          */
924         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
925         /* huge page split is done under lru_lock. so, we have no races. */
926         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
927         if (mem_cgroup_is_root(pc->mem_cgroup))
928                 return;
929         VM_BUG_ON(list_empty(&pc->lru));
930         list_del_init(&pc->lru);
931 }
932
933 void mem_cgroup_del_lru(struct page *page)
934 {
935         mem_cgroup_del_lru_list(page, page_lru(page));
936 }
937
938 /*
939  * Writeback is about to end against a page which has been marked for immediate
940  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
941  * inactive list.
942  */
943 void mem_cgroup_rotate_reclaimable_page(struct page *page)
944 {
945         struct mem_cgroup_per_zone *mz;
946         struct page_cgroup *pc;
947         enum lru_list lru = page_lru(page);
948
949         if (mem_cgroup_disabled())
950                 return;
951
952         pc = lookup_page_cgroup(page);
953         /* unused or root page is not rotated. */
954         if (!PageCgroupUsed(pc))
955                 return;
956         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
957         smp_rmb();
958         if (mem_cgroup_is_root(pc->mem_cgroup))
959                 return;
960         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
961         list_move_tail(&pc->lru, &mz->lists[lru]);
962 }
963
964 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
965 {
966         struct mem_cgroup_per_zone *mz;
967         struct page_cgroup *pc;
968
969         if (mem_cgroup_disabled())
970                 return;
971
972         pc = lookup_page_cgroup(page);
973         /* unused or root page is not rotated. */
974         if (!PageCgroupUsed(pc))
975                 return;
976         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
977         smp_rmb();
978         if (mem_cgroup_is_root(pc->mem_cgroup))
979                 return;
980         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
981         list_move(&pc->lru, &mz->lists[lru]);
982 }
983
984 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
985 {
986         struct page_cgroup *pc;
987         struct mem_cgroup_per_zone *mz;
988
989         if (mem_cgroup_disabled())
990                 return;
991         pc = lookup_page_cgroup(page);
992         VM_BUG_ON(PageCgroupAcctLRU(pc));
993         if (!PageCgroupUsed(pc))
994                 return;
995         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
996         smp_rmb();
997         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
998         /* huge page split is done under lru_lock. so, we have no races. */
999         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1000         SetPageCgroupAcctLRU(pc);
1001         if (mem_cgroup_is_root(pc->mem_cgroup))
1002                 return;
1003         list_add(&pc->lru, &mz->lists[lru]);
1004 }
1005
1006 /*
1007  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1008  * while it's linked to lru because the page may be reused after it's fully
1009  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1010  * It's done under lock_page and expected that zone->lru_lock isnever held.
1011  */
1012 static void mem_cgroup_lru_del_before_commit(struct page *page)
1013 {
1014         unsigned long flags;
1015         struct zone *zone = page_zone(page);
1016         struct page_cgroup *pc = lookup_page_cgroup(page);
1017
1018         /*
1019          * Doing this check without taking ->lru_lock seems wrong but this
1020          * is safe. Because if page_cgroup's USED bit is unset, the page
1021          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1022          * set, the commit after this will fail, anyway.
1023          * This all charge/uncharge is done under some mutual execustion.
1024          * So, we don't need to taking care of changes in USED bit.
1025          */
1026         if (likely(!PageLRU(page)))
1027                 return;
1028
1029         spin_lock_irqsave(&zone->lru_lock, flags);
1030         /*
1031          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1032          * is guarded by lock_page() because the page is SwapCache.
1033          */
1034         if (!PageCgroupUsed(pc))
1035                 mem_cgroup_del_lru_list(page, page_lru(page));
1036         spin_unlock_irqrestore(&zone->lru_lock, flags);
1037 }
1038
1039 static void mem_cgroup_lru_add_after_commit(struct page *page)
1040 {
1041         unsigned long flags;
1042         struct zone *zone = page_zone(page);
1043         struct page_cgroup *pc = lookup_page_cgroup(page);
1044
1045         /* taking care of that the page is added to LRU while we commit it */
1046         if (likely(!PageLRU(page)))
1047                 return;
1048         spin_lock_irqsave(&zone->lru_lock, flags);
1049         /* link when the page is linked to LRU but page_cgroup isn't */
1050         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1051                 mem_cgroup_add_lru_list(page, page_lru(page));
1052         spin_unlock_irqrestore(&zone->lru_lock, flags);
1053 }
1054
1055
1056 void mem_cgroup_move_lists(struct page *page,
1057                            enum lru_list from, enum lru_list to)
1058 {
1059         if (mem_cgroup_disabled())
1060                 return;
1061         mem_cgroup_del_lru_list(page, from);
1062         mem_cgroup_add_lru_list(page, to);
1063 }
1064
1065 /*
1066  * Checks whether given mem is same or in the root_mem's
1067  * hierarchy subtree
1068  */
1069 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_mem,
1070                 struct mem_cgroup *mem)
1071 {
1072         if (root_mem != mem) {
1073                 return (root_mem->use_hierarchy &&
1074                         css_is_ancestor(&mem->css, &root_mem->css));
1075         }
1076
1077         return true;
1078 }
1079
1080 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1081 {
1082         int ret;
1083         struct mem_cgroup *curr = NULL;
1084         struct task_struct *p;
1085
1086         p = find_lock_task_mm(task);
1087         if (!p)
1088                 return 0;
1089         curr = try_get_mem_cgroup_from_mm(p->mm);
1090         task_unlock(p);
1091         if (!curr)
1092                 return 0;
1093         /*
1094          * We should check use_hierarchy of "mem" not "curr". Because checking
1095          * use_hierarchy of "curr" here make this function true if hierarchy is
1096          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1097          * hierarchy(even if use_hierarchy is disabled in "mem").
1098          */
1099         ret = mem_cgroup_same_or_subtree(mem, curr);
1100         css_put(&curr->css);
1101         return ret;
1102 }
1103
1104 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1105 {
1106         unsigned long active;
1107         unsigned long inactive;
1108         unsigned long gb;
1109         unsigned long inactive_ratio;
1110
1111         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1112         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1113
1114         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1115         if (gb)
1116                 inactive_ratio = int_sqrt(10 * gb);
1117         else
1118                 inactive_ratio = 1;
1119
1120         if (present_pages) {
1121                 present_pages[0] = inactive;
1122                 present_pages[1] = active;
1123         }
1124
1125         return inactive_ratio;
1126 }
1127
1128 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1129 {
1130         unsigned long active;
1131         unsigned long inactive;
1132         unsigned long present_pages[2];
1133         unsigned long inactive_ratio;
1134
1135         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1136
1137         inactive = present_pages[0];
1138         active = present_pages[1];
1139
1140         if (inactive * inactive_ratio < active)
1141                 return 1;
1142
1143         return 0;
1144 }
1145
1146 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1147 {
1148         unsigned long active;
1149         unsigned long inactive;
1150
1151         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1152         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1153
1154         return (active > inactive);
1155 }
1156
1157 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1158                                                       struct zone *zone)
1159 {
1160         int nid = zone_to_nid(zone);
1161         int zid = zone_idx(zone);
1162         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1163
1164         return &mz->reclaim_stat;
1165 }
1166
1167 struct zone_reclaim_stat *
1168 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1169 {
1170         struct page_cgroup *pc;
1171         struct mem_cgroup_per_zone *mz;
1172
1173         if (mem_cgroup_disabled())
1174                 return NULL;
1175
1176         pc = lookup_page_cgroup(page);
1177         if (!PageCgroupUsed(pc))
1178                 return NULL;
1179         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1180         smp_rmb();
1181         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1182         return &mz->reclaim_stat;
1183 }
1184
1185 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1186                                         struct list_head *dst,
1187                                         unsigned long *scanned, int order,
1188                                         int mode, struct zone *z,
1189                                         struct mem_cgroup *mem_cont,
1190                                         int active, int file)
1191 {
1192         unsigned long nr_taken = 0;
1193         struct page *page;
1194         unsigned long scan;
1195         LIST_HEAD(pc_list);
1196         struct list_head *src;
1197         struct page_cgroup *pc, *tmp;
1198         int nid = zone_to_nid(z);
1199         int zid = zone_idx(z);
1200         struct mem_cgroup_per_zone *mz;
1201         int lru = LRU_FILE * file + active;
1202         int ret;
1203
1204         BUG_ON(!mem_cont);
1205         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1206         src = &mz->lists[lru];
1207
1208         scan = 0;
1209         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1210                 if (scan >= nr_to_scan)
1211                         break;
1212
1213                 if (unlikely(!PageCgroupUsed(pc)))
1214                         continue;
1215
1216                 page = lookup_cgroup_page(pc);
1217
1218                 if (unlikely(!PageLRU(page)))
1219                         continue;
1220
1221                 scan++;
1222                 ret = __isolate_lru_page(page, mode, file);
1223                 switch (ret) {
1224                 case 0:
1225                         list_move(&page->lru, dst);
1226                         mem_cgroup_del_lru(page);
1227                         nr_taken += hpage_nr_pages(page);
1228                         break;
1229                 case -EBUSY:
1230                         /* we don't affect global LRU but rotate in our LRU */
1231                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1232                         break;
1233                 default:
1234                         break;
1235                 }
1236         }
1237
1238         *scanned = scan;
1239
1240         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1241                                       0, 0, 0, mode);
1242
1243         return nr_taken;
1244 }
1245
1246 #define mem_cgroup_from_res_counter(counter, member)    \
1247         container_of(counter, struct mem_cgroup, member)
1248
1249 /**
1250  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1251  * @mem: the memory cgroup
1252  *
1253  * Returns the maximum amount of memory @mem can be charged with, in
1254  * pages.
1255  */
1256 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1257 {
1258         unsigned long long margin;
1259
1260         margin = res_counter_margin(&mem->res);
1261         if (do_swap_account)
1262                 margin = min(margin, res_counter_margin(&mem->memsw));
1263         return margin >> PAGE_SHIFT;
1264 }
1265
1266 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1267 {
1268         struct cgroup *cgrp = memcg->css.cgroup;
1269
1270         /* root ? */
1271         if (cgrp->parent == NULL)
1272                 return vm_swappiness;
1273
1274         return memcg->swappiness;
1275 }
1276
1277 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1278 {
1279         int cpu;
1280
1281         get_online_cpus();
1282         spin_lock(&mem->pcp_counter_lock);
1283         for_each_online_cpu(cpu)
1284                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1285         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1286         spin_unlock(&mem->pcp_counter_lock);
1287         put_online_cpus();
1288
1289         synchronize_rcu();
1290 }
1291
1292 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1293 {
1294         int cpu;
1295
1296         if (!mem)
1297                 return;
1298         get_online_cpus();
1299         spin_lock(&mem->pcp_counter_lock);
1300         for_each_online_cpu(cpu)
1301                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1302         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1303         spin_unlock(&mem->pcp_counter_lock);
1304         put_online_cpus();
1305 }
1306 /*
1307  * 2 routines for checking "mem" is under move_account() or not.
1308  *
1309  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1310  *                        for avoiding race in accounting. If true,
1311  *                        pc->mem_cgroup may be overwritten.
1312  *
1313  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1314  *                        under hierarchy of moving cgroups. This is for
1315  *                        waiting at hith-memory prressure caused by "move".
1316  */
1317
1318 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1319 {
1320         VM_BUG_ON(!rcu_read_lock_held());
1321         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1322 }
1323
1324 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1325 {
1326         struct mem_cgroup *from;
1327         struct mem_cgroup *to;
1328         bool ret = false;
1329         /*
1330          * Unlike task_move routines, we access mc.to, mc.from not under
1331          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1332          */
1333         spin_lock(&mc.lock);
1334         from = mc.from;
1335         to = mc.to;
1336         if (!from)
1337                 goto unlock;
1338
1339         ret = mem_cgroup_same_or_subtree(mem, from)
1340                 || mem_cgroup_same_or_subtree(mem, to);
1341 unlock:
1342         spin_unlock(&mc.lock);
1343         return ret;
1344 }
1345
1346 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1347 {
1348         if (mc.moving_task && current != mc.moving_task) {
1349                 if (mem_cgroup_under_move(mem)) {
1350                         DEFINE_WAIT(wait);
1351                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1352                         /* moving charge context might have finished. */
1353                         if (mc.moving_task)
1354                                 schedule();
1355                         finish_wait(&mc.waitq, &wait);
1356                         return true;
1357                 }
1358         }
1359         return false;
1360 }
1361
1362 /**
1363  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1364  * @memcg: The memory cgroup that went over limit
1365  * @p: Task that is going to be killed
1366  *
1367  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1368  * enabled
1369  */
1370 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1371 {
1372         struct cgroup *task_cgrp;
1373         struct cgroup *mem_cgrp;
1374         /*
1375          * Need a buffer in BSS, can't rely on allocations. The code relies
1376          * on the assumption that OOM is serialized for memory controller.
1377          * If this assumption is broken, revisit this code.
1378          */
1379         static char memcg_name[PATH_MAX];
1380         int ret;
1381
1382         if (!memcg || !p)
1383                 return;
1384
1385
1386         rcu_read_lock();
1387
1388         mem_cgrp = memcg->css.cgroup;
1389         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1390
1391         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1392         if (ret < 0) {
1393                 /*
1394                  * Unfortunately, we are unable to convert to a useful name
1395                  * But we'll still print out the usage information
1396                  */
1397                 rcu_read_unlock();
1398                 goto done;
1399         }
1400         rcu_read_unlock();
1401
1402         printk(KERN_INFO "Task in %s killed", memcg_name);
1403
1404         rcu_read_lock();
1405         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1406         if (ret < 0) {
1407                 rcu_read_unlock();
1408                 goto done;
1409         }
1410         rcu_read_unlock();
1411
1412         /*
1413          * Continues from above, so we don't need an KERN_ level
1414          */
1415         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1416 done:
1417
1418         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1419                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1420                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1421                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1422         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1423                 "failcnt %llu\n",
1424                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1425                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1426                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1427 }
1428
1429 /*
1430  * This function returns the number of memcg under hierarchy tree. Returns
1431  * 1(self count) if no children.
1432  */
1433 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1434 {
1435         int num = 0;
1436         struct mem_cgroup *iter;
1437
1438         for_each_mem_cgroup_tree(iter, mem)
1439                 num++;
1440         return num;
1441 }
1442
1443 /*
1444  * Return the memory (and swap, if configured) limit for a memcg.
1445  */
1446 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1447 {
1448         u64 limit;
1449         u64 memsw;
1450
1451         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1452         limit += total_swap_pages << PAGE_SHIFT;
1453
1454         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1455         /*
1456          * If memsw is finite and limits the amount of swap space available
1457          * to this memcg, return that limit.
1458          */
1459         return min(limit, memsw);
1460 }
1461
1462 /*
1463  * Visit the first child (need not be the first child as per the ordering
1464  * of the cgroup list, since we track last_scanned_child) of @mem and use
1465  * that to reclaim free pages from.
1466  */
1467 static struct mem_cgroup *
1468 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1469 {
1470         struct mem_cgroup *ret = NULL;
1471         struct cgroup_subsys_state *css;
1472         int nextid, found;
1473
1474         if (!root_mem->use_hierarchy) {
1475                 css_get(&root_mem->css);
1476                 ret = root_mem;
1477         }
1478
1479         while (!ret) {
1480                 rcu_read_lock();
1481                 nextid = root_mem->last_scanned_child + 1;
1482                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1483                                    &found);
1484                 if (css && css_tryget(css))
1485                         ret = container_of(css, struct mem_cgroup, css);
1486
1487                 rcu_read_unlock();
1488                 /* Updates scanning parameter */
1489                 if (!css) {
1490                         /* this means start scan from ID:1 */
1491                         root_mem->last_scanned_child = 0;
1492                 } else
1493                         root_mem->last_scanned_child = found;
1494         }
1495
1496         return ret;
1497 }
1498
1499 /**
1500  * test_mem_cgroup_node_reclaimable
1501  * @mem: the target memcg
1502  * @nid: the node ID to be checked.
1503  * @noswap : specify true here if the user wants flle only information.
1504  *
1505  * This function returns whether the specified memcg contains any
1506  * reclaimable pages on a node. Returns true if there are any reclaimable
1507  * pages in the node.
1508  */
1509 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1510                 int nid, bool noswap)
1511 {
1512         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1513                 return true;
1514         if (noswap || !total_swap_pages)
1515                 return false;
1516         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1517                 return true;
1518         return false;
1519
1520 }
1521 #if MAX_NUMNODES > 1
1522
1523 /*
1524  * Always updating the nodemask is not very good - even if we have an empty
1525  * list or the wrong list here, we can start from some node and traverse all
1526  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1527  *
1528  */
1529 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1530 {
1531         int nid;
1532         /*
1533          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1534          * pagein/pageout changes since the last update.
1535          */
1536         if (!atomic_read(&mem->numainfo_events))
1537                 return;
1538         if (atomic_inc_return(&mem->numainfo_updating) > 1)
1539                 return;
1540
1541         /* make a nodemask where this memcg uses memory from */
1542         mem->scan_nodes = node_states[N_HIGH_MEMORY];
1543
1544         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1545
1546                 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1547                         node_clear(nid, mem->scan_nodes);
1548         }
1549
1550         atomic_set(&mem->numainfo_events, 0);
1551         atomic_set(&mem->numainfo_updating, 0);
1552 }
1553
1554 /*
1555  * Selecting a node where we start reclaim from. Because what we need is just
1556  * reducing usage counter, start from anywhere is O,K. Considering
1557  * memory reclaim from current node, there are pros. and cons.
1558  *
1559  * Freeing memory from current node means freeing memory from a node which
1560  * we'll use or we've used. So, it may make LRU bad. And if several threads
1561  * hit limits, it will see a contention on a node. But freeing from remote
1562  * node means more costs for memory reclaim because of memory latency.
1563  *
1564  * Now, we use round-robin. Better algorithm is welcomed.
1565  */
1566 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1567 {
1568         int node;
1569
1570         mem_cgroup_may_update_nodemask(mem);
1571         node = mem->last_scanned_node;
1572
1573         node = next_node(node, mem->scan_nodes);
1574         if (node == MAX_NUMNODES)
1575                 node = first_node(mem->scan_nodes);
1576         /*
1577          * We call this when we hit limit, not when pages are added to LRU.
1578          * No LRU may hold pages because all pages are UNEVICTABLE or
1579          * memcg is too small and all pages are not on LRU. In that case,
1580          * we use curret node.
1581          */
1582         if (unlikely(node == MAX_NUMNODES))
1583                 node = numa_node_id();
1584
1585         mem->last_scanned_node = node;
1586         return node;
1587 }
1588
1589 /*
1590  * Check all nodes whether it contains reclaimable pages or not.
1591  * For quick scan, we make use of scan_nodes. This will allow us to skip
1592  * unused nodes. But scan_nodes is lazily updated and may not cotain
1593  * enough new information. We need to do double check.
1594  */
1595 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1596 {
1597         int nid;
1598
1599         /*
1600          * quick check...making use of scan_node.
1601          * We can skip unused nodes.
1602          */
1603         if (!nodes_empty(mem->scan_nodes)) {
1604                 for (nid = first_node(mem->scan_nodes);
1605                      nid < MAX_NUMNODES;
1606                      nid = next_node(nid, mem->scan_nodes)) {
1607
1608                         if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1609                                 return true;
1610                 }
1611         }
1612         /*
1613          * Check rest of nodes.
1614          */
1615         for_each_node_state(nid, N_HIGH_MEMORY) {
1616                 if (node_isset(nid, mem->scan_nodes))
1617                         continue;
1618                 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1619                         return true;
1620         }
1621         return false;
1622 }
1623
1624 #else
1625 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1626 {
1627         return 0;
1628 }
1629
1630 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1631 {
1632         return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1633 }
1634 #endif
1635
1636 /*
1637  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1638  * we reclaimed from, so that we don't end up penalizing one child extensively
1639  * based on its position in the children list.
1640  *
1641  * root_mem is the original ancestor that we've been reclaim from.
1642  *
1643  * We give up and return to the caller when we visit root_mem twice.
1644  * (other groups can be removed while we're walking....)
1645  *
1646  * If shrink==true, for avoiding to free too much, this returns immedieately.
1647  */
1648 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1649                                                 struct zone *zone,
1650                                                 gfp_t gfp_mask,
1651                                                 unsigned long reclaim_options,
1652                                                 unsigned long *total_scanned)
1653 {
1654         struct mem_cgroup *victim;
1655         int ret, total = 0;
1656         int loop = 0;
1657         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1658         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1659         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1660         unsigned long excess;
1661         unsigned long nr_scanned;
1662
1663         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1664
1665         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1666         if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1667                 noswap = true;
1668
1669         while (1) {
1670                 victim = mem_cgroup_select_victim(root_mem);
1671                 if (victim == root_mem) {
1672                         loop++;
1673                         /*
1674                          * We are not draining per cpu cached charges during
1675                          * soft limit reclaim  because global reclaim doesn't
1676                          * care about charges. It tries to free some memory and
1677                          * charges will not give any.
1678                          */
1679                         if (!check_soft && loop >= 1)
1680                                 drain_all_stock_async(root_mem);
1681                         if (loop >= 2) {
1682                                 /*
1683                                  * If we have not been able to reclaim
1684                                  * anything, it might because there are
1685                                  * no reclaimable pages under this hierarchy
1686                                  */
1687                                 if (!check_soft || !total) {
1688                                         css_put(&victim->css);
1689                                         break;
1690                                 }
1691                                 /*
1692                                  * We want to do more targeted reclaim.
1693                                  * excess >> 2 is not to excessive so as to
1694                                  * reclaim too much, nor too less that we keep
1695                                  * coming back to reclaim from this cgroup
1696                                  */
1697                                 if (total >= (excess >> 2) ||
1698                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1699                                         css_put(&victim->css);
1700                                         break;
1701                                 }
1702                         }
1703                 }
1704                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1705                         /* this cgroup's local usage == 0 */
1706                         css_put(&victim->css);
1707                         continue;
1708                 }
1709                 /* we use swappiness of local cgroup */
1710                 if (check_soft) {
1711                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1712                                 noswap, zone, &nr_scanned);
1713                         *total_scanned += nr_scanned;
1714                 } else
1715                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1716                                                 noswap);
1717                 css_put(&victim->css);
1718                 /*
1719                  * At shrinking usage, we can't check we should stop here or
1720                  * reclaim more. It's depends on callers. last_scanned_child
1721                  * will work enough for keeping fairness under tree.
1722                  */
1723                 if (shrink)
1724                         return ret;
1725                 total += ret;
1726                 if (check_soft) {
1727                         if (!res_counter_soft_limit_excess(&root_mem->res))
1728                                 return total;
1729                 } else if (mem_cgroup_margin(root_mem))
1730                         return total;
1731         }
1732         return total;
1733 }
1734
1735 /*
1736  * Check OOM-Killer is already running under our hierarchy.
1737  * If someone is running, return false.
1738  * Has to be called with memcg_oom_lock
1739  */
1740 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1741 {
1742         struct mem_cgroup *iter, *failed = NULL;
1743         bool cond = true;
1744
1745         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1746                 if (iter->oom_lock) {
1747                         /*
1748                          * this subtree of our hierarchy is already locked
1749                          * so we cannot give a lock.
1750                          */
1751                         failed = iter;
1752                         cond = false;
1753                 } else
1754                         iter->oom_lock = true;
1755         }
1756
1757         if (!failed)
1758                 return true;
1759
1760         /*
1761          * OK, we failed to lock the whole subtree so we have to clean up
1762          * what we set up to the failing subtree
1763          */
1764         cond = true;
1765         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1766                 if (iter == failed) {
1767                         cond = false;
1768                         continue;
1769                 }
1770                 iter->oom_lock = false;
1771         }
1772         return false;
1773 }
1774
1775 /*
1776  * Has to be called with memcg_oom_lock
1777  */
1778 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1779 {
1780         struct mem_cgroup *iter;
1781
1782         for_each_mem_cgroup_tree(iter, mem)
1783                 iter->oom_lock = false;
1784         return 0;
1785 }
1786
1787 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1788 {
1789         struct mem_cgroup *iter;
1790
1791         for_each_mem_cgroup_tree(iter, mem)
1792                 atomic_inc(&iter->under_oom);
1793 }
1794
1795 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1796 {
1797         struct mem_cgroup *iter;
1798
1799         /*
1800          * When a new child is created while the hierarchy is under oom,
1801          * mem_cgroup_oom_lock() may not be called. We have to use
1802          * atomic_add_unless() here.
1803          */
1804         for_each_mem_cgroup_tree(iter, mem)
1805                 atomic_add_unless(&iter->under_oom, -1, 0);
1806 }
1807
1808 static DEFINE_SPINLOCK(memcg_oom_lock);
1809 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1810
1811 struct oom_wait_info {
1812         struct mem_cgroup *mem;
1813         wait_queue_t    wait;
1814 };
1815
1816 static int memcg_oom_wake_function(wait_queue_t *wait,
1817         unsigned mode, int sync, void *arg)
1818 {
1819         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg,
1820                           *oom_wait_mem;
1821         struct oom_wait_info *oom_wait_info;
1822
1823         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1824         oom_wait_mem = oom_wait_info->mem;
1825
1826         /*
1827          * Both of oom_wait_info->mem and wake_mem are stable under us.
1828          * Then we can use css_is_ancestor without taking care of RCU.
1829          */
1830         if (!mem_cgroup_same_or_subtree(oom_wait_mem, wake_mem)
1831                         && !mem_cgroup_same_or_subtree(wake_mem, oom_wait_mem))
1832                 return 0;
1833         return autoremove_wake_function(wait, mode, sync, arg);
1834 }
1835
1836 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1837 {
1838         /* for filtering, pass "mem" as argument. */
1839         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1840 }
1841
1842 static void memcg_oom_recover(struct mem_cgroup *mem)
1843 {
1844         if (mem && atomic_read(&mem->under_oom))
1845                 memcg_wakeup_oom(mem);
1846 }
1847
1848 /*
1849  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1850  */
1851 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1852 {
1853         struct oom_wait_info owait;
1854         bool locked, need_to_kill;
1855
1856         owait.mem = mem;
1857         owait.wait.flags = 0;
1858         owait.wait.func = memcg_oom_wake_function;
1859         owait.wait.private = current;
1860         INIT_LIST_HEAD(&owait.wait.task_list);
1861         need_to_kill = true;
1862         mem_cgroup_mark_under_oom(mem);
1863
1864         /* At first, try to OOM lock hierarchy under mem.*/
1865         spin_lock(&memcg_oom_lock);
1866         locked = mem_cgroup_oom_lock(mem);
1867         /*
1868          * Even if signal_pending(), we can't quit charge() loop without
1869          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1870          * under OOM is always welcomed, use TASK_KILLABLE here.
1871          */
1872         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1873         if (!locked || mem->oom_kill_disable)
1874                 need_to_kill = false;
1875         if (locked)
1876                 mem_cgroup_oom_notify(mem);
1877         spin_unlock(&memcg_oom_lock);
1878
1879         if (need_to_kill) {
1880                 finish_wait(&memcg_oom_waitq, &owait.wait);
1881                 mem_cgroup_out_of_memory(mem, mask);
1882         } else {
1883                 schedule();
1884                 finish_wait(&memcg_oom_waitq, &owait.wait);
1885         }
1886         spin_lock(&memcg_oom_lock);
1887         if (locked)
1888                 mem_cgroup_oom_unlock(mem);
1889         memcg_wakeup_oom(mem);
1890         spin_unlock(&memcg_oom_lock);
1891
1892         mem_cgroup_unmark_under_oom(mem);
1893
1894         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1895                 return false;
1896         /* Give chance to dying process */
1897         schedule_timeout(1);
1898         return true;
1899 }
1900
1901 /*
1902  * Currently used to update mapped file statistics, but the routine can be
1903  * generalized to update other statistics as well.
1904  *
1905  * Notes: Race condition
1906  *
1907  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1908  * it tends to be costly. But considering some conditions, we doesn't need
1909  * to do so _always_.
1910  *
1911  * Considering "charge", lock_page_cgroup() is not required because all
1912  * file-stat operations happen after a page is attached to radix-tree. There
1913  * are no race with "charge".
1914  *
1915  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1916  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1917  * if there are race with "uncharge". Statistics itself is properly handled
1918  * by flags.
1919  *
1920  * Considering "move", this is an only case we see a race. To make the race
1921  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1922  * possibility of race condition. If there is, we take a lock.
1923  */
1924
1925 void mem_cgroup_update_page_stat(struct page *page,
1926                                  enum mem_cgroup_page_stat_item idx, int val)
1927 {
1928         struct mem_cgroup *mem;
1929         struct page_cgroup *pc = lookup_page_cgroup(page);
1930         bool need_unlock = false;
1931         unsigned long uninitialized_var(flags);
1932
1933         if (unlikely(!pc))
1934                 return;
1935
1936         rcu_read_lock();
1937         mem = pc->mem_cgroup;
1938         if (unlikely(!mem || !PageCgroupUsed(pc)))
1939                 goto out;
1940         /* pc->mem_cgroup is unstable ? */
1941         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1942                 /* take a lock against to access pc->mem_cgroup */
1943                 move_lock_page_cgroup(pc, &flags);
1944                 need_unlock = true;
1945                 mem = pc->mem_cgroup;
1946                 if (!mem || !PageCgroupUsed(pc))
1947                         goto out;
1948         }
1949
1950         switch (idx) {
1951         case MEMCG_NR_FILE_MAPPED:
1952                 if (val > 0)
1953                         SetPageCgroupFileMapped(pc);
1954                 else if (!page_mapped(page))
1955                         ClearPageCgroupFileMapped(pc);
1956                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1957                 break;
1958         default:
1959                 BUG();
1960         }
1961
1962         this_cpu_add(mem->stat->count[idx], val);
1963
1964 out:
1965         if (unlikely(need_unlock))
1966                 move_unlock_page_cgroup(pc, &flags);
1967         rcu_read_unlock();
1968         return;
1969 }
1970 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1971
1972 /*
1973  * size of first charge trial. "32" comes from vmscan.c's magic value.
1974  * TODO: maybe necessary to use big numbers in big irons.
1975  */
1976 #define CHARGE_BATCH    32U
1977 struct memcg_stock_pcp {
1978         struct mem_cgroup *cached; /* this never be root cgroup */
1979         unsigned int nr_pages;
1980         struct work_struct work;
1981         unsigned long flags;
1982 #define FLUSHING_CACHED_CHARGE  (0)
1983 };
1984 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1985 static DEFINE_MUTEX(percpu_charge_mutex);
1986
1987 /*
1988  * Try to consume stocked charge on this cpu. If success, one page is consumed
1989  * from local stock and true is returned. If the stock is 0 or charges from a
1990  * cgroup which is not current target, returns false. This stock will be
1991  * refilled.
1992  */
1993 static bool consume_stock(struct mem_cgroup *mem)
1994 {
1995         struct memcg_stock_pcp *stock;
1996         bool ret = true;
1997
1998         stock = &get_cpu_var(memcg_stock);
1999         if (mem == stock->cached && stock->nr_pages)
2000                 stock->nr_pages--;
2001         else /* need to call res_counter_charge */
2002                 ret = false;
2003         put_cpu_var(memcg_stock);
2004         return ret;
2005 }
2006
2007 /*
2008  * Returns stocks cached in percpu to res_counter and reset cached information.
2009  */
2010 static void drain_stock(struct memcg_stock_pcp *stock)
2011 {
2012         struct mem_cgroup *old = stock->cached;
2013
2014         if (stock->nr_pages) {
2015                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2016
2017                 res_counter_uncharge(&old->res, bytes);
2018                 if (do_swap_account)
2019                         res_counter_uncharge(&old->memsw, bytes);
2020                 stock->nr_pages = 0;
2021         }
2022         stock->cached = NULL;
2023 }
2024
2025 /*
2026  * This must be called under preempt disabled or must be called by
2027  * a thread which is pinned to local cpu.
2028  */
2029 static void drain_local_stock(struct work_struct *dummy)
2030 {
2031         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2032         drain_stock(stock);
2033         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2034 }
2035
2036 /*
2037  * Cache charges(val) which is from res_counter, to local per_cpu area.
2038  * This will be consumed by consume_stock() function, later.
2039  */
2040 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2041 {
2042         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2043
2044         if (stock->cached != mem) { /* reset if necessary */
2045                 drain_stock(stock);
2046                 stock->cached = mem;
2047         }
2048         stock->nr_pages += nr_pages;
2049         put_cpu_var(memcg_stock);
2050 }
2051
2052 /*
2053  * Drains all per-CPU charge caches for given root_mem resp. subtree
2054  * of the hierarchy under it. sync flag says whether we should block
2055  * until the work is done.
2056  */
2057 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2058 {
2059         int cpu, curcpu;
2060
2061         /* Notify other cpus that system-wide "drain" is running */
2062         get_online_cpus();
2063         curcpu = get_cpu();
2064         for_each_online_cpu(cpu) {
2065                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2066                 struct mem_cgroup *mem;
2067
2068                 mem = stock->cached;
2069                 if (!mem || !stock->nr_pages)
2070                         continue;
2071                 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2072                         continue;
2073                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2074                         if (cpu == curcpu)
2075                                 drain_local_stock(&stock->work);
2076                         else
2077                                 schedule_work_on(cpu, &stock->work);
2078                 }
2079         }
2080         put_cpu();
2081
2082         if (!sync)
2083                 goto out;
2084
2085         for_each_online_cpu(cpu) {
2086                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2087                 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2088                         flush_work(&stock->work);
2089         }
2090 out:
2091         put_online_cpus();
2092 }
2093
2094 /*
2095  * Tries to drain stocked charges in other cpus. This function is asynchronous
2096  * and just put a work per cpu for draining localy on each cpu. Caller can
2097  * expects some charges will be back to res_counter later but cannot wait for
2098  * it.
2099  */
2100 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2101 {
2102         /*
2103          * If someone calls draining, avoid adding more kworker runs.
2104          */
2105         if (!mutex_trylock(&percpu_charge_mutex))
2106                 return;
2107         drain_all_stock(root_mem, false);
2108         mutex_unlock(&percpu_charge_mutex);
2109 }
2110
2111 /* This is a synchronous drain interface. */
2112 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2113 {
2114         /* called when force_empty is called */
2115         mutex_lock(&percpu_charge_mutex);
2116         drain_all_stock(root_mem, true);
2117         mutex_unlock(&percpu_charge_mutex);
2118 }
2119
2120 /*
2121  * This function drains percpu counter value from DEAD cpu and
2122  * move it to local cpu. Note that this function can be preempted.
2123  */
2124 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2125 {
2126         int i;
2127
2128         spin_lock(&mem->pcp_counter_lock);
2129         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2130                 long x = per_cpu(mem->stat->count[i], cpu);
2131
2132                 per_cpu(mem->stat->count[i], cpu) = 0;
2133                 mem->nocpu_base.count[i] += x;
2134         }
2135         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2136                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2137
2138                 per_cpu(mem->stat->events[i], cpu) = 0;
2139                 mem->nocpu_base.events[i] += x;
2140         }
2141         /* need to clear ON_MOVE value, works as a kind of lock. */
2142         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2143         spin_unlock(&mem->pcp_counter_lock);
2144 }
2145
2146 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2147 {
2148         int idx = MEM_CGROUP_ON_MOVE;
2149
2150         spin_lock(&mem->pcp_counter_lock);
2151         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2152         spin_unlock(&mem->pcp_counter_lock);
2153 }
2154
2155 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2156                                         unsigned long action,
2157                                         void *hcpu)
2158 {
2159         int cpu = (unsigned long)hcpu;
2160         struct memcg_stock_pcp *stock;
2161         struct mem_cgroup *iter;
2162
2163         if ((action == CPU_ONLINE)) {
2164                 for_each_mem_cgroup_all(iter)
2165                         synchronize_mem_cgroup_on_move(iter, cpu);
2166                 return NOTIFY_OK;
2167         }
2168
2169         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2170                 return NOTIFY_OK;
2171
2172         for_each_mem_cgroup_all(iter)
2173                 mem_cgroup_drain_pcp_counter(iter, cpu);
2174
2175         stock = &per_cpu(memcg_stock, cpu);
2176         drain_stock(stock);
2177         return NOTIFY_OK;
2178 }
2179
2180
2181 /* See __mem_cgroup_try_charge() for details */
2182 enum {
2183         CHARGE_OK,              /* success */
2184         CHARGE_RETRY,           /* need to retry but retry is not bad */
2185         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2186         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2187         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2188 };
2189
2190 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2191                                 unsigned int nr_pages, bool oom_check)
2192 {
2193         unsigned long csize = nr_pages * PAGE_SIZE;
2194         struct mem_cgroup *mem_over_limit;
2195         struct res_counter *fail_res;
2196         unsigned long flags = 0;
2197         int ret;
2198
2199         ret = res_counter_charge(&mem->res, csize, &fail_res);
2200
2201         if (likely(!ret)) {
2202                 if (!do_swap_account)
2203                         return CHARGE_OK;
2204                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2205                 if (likely(!ret))
2206                         return CHARGE_OK;
2207
2208                 res_counter_uncharge(&mem->res, csize);
2209                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2210                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2211         } else
2212                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2213         /*
2214          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2215          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2216          *
2217          * Never reclaim on behalf of optional batching, retry with a
2218          * single page instead.
2219          */
2220         if (nr_pages == CHARGE_BATCH)
2221                 return CHARGE_RETRY;
2222
2223         if (!(gfp_mask & __GFP_WAIT))
2224                 return CHARGE_WOULDBLOCK;
2225
2226         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2227                                               gfp_mask, flags, NULL);
2228         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2229                 return CHARGE_RETRY;
2230         /*
2231          * Even though the limit is exceeded at this point, reclaim
2232          * may have been able to free some pages.  Retry the charge
2233          * before killing the task.
2234          *
2235          * Only for regular pages, though: huge pages are rather
2236          * unlikely to succeed so close to the limit, and we fall back
2237          * to regular pages anyway in case of failure.
2238          */
2239         if (nr_pages == 1 && ret)
2240                 return CHARGE_RETRY;
2241
2242         /*
2243          * At task move, charge accounts can be doubly counted. So, it's
2244          * better to wait until the end of task_move if something is going on.
2245          */
2246         if (mem_cgroup_wait_acct_move(mem_over_limit))
2247                 return CHARGE_RETRY;
2248
2249         /* If we don't need to call oom-killer at el, return immediately */
2250         if (!oom_check)
2251                 return CHARGE_NOMEM;
2252         /* check OOM */
2253         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2254                 return CHARGE_OOM_DIE;
2255
2256         return CHARGE_RETRY;
2257 }
2258
2259 /*
2260  * Unlike exported interface, "oom" parameter is added. if oom==true,
2261  * oom-killer can be invoked.
2262  */
2263 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2264                                    gfp_t gfp_mask,
2265                                    unsigned int nr_pages,
2266                                    struct mem_cgroup **memcg,
2267                                    bool oom)
2268 {
2269         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2270         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2271         struct mem_cgroup *mem = NULL;
2272         int ret;
2273
2274         /*
2275          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2276          * in system level. So, allow to go ahead dying process in addition to
2277          * MEMDIE process.
2278          */
2279         if (unlikely(test_thread_flag(TIF_MEMDIE)
2280                      || fatal_signal_pending(current)))
2281                 goto bypass;
2282
2283         /*
2284          * We always charge the cgroup the mm_struct belongs to.
2285          * The mm_struct's mem_cgroup changes on task migration if the
2286          * thread group leader migrates. It's possible that mm is not
2287          * set, if so charge the init_mm (happens for pagecache usage).
2288          */
2289         if (!*memcg && !mm)
2290                 goto bypass;
2291 again:
2292         if (*memcg) { /* css should be a valid one */
2293                 mem = *memcg;
2294                 VM_BUG_ON(css_is_removed(&mem->css));
2295                 if (mem_cgroup_is_root(mem))
2296                         goto done;
2297                 if (nr_pages == 1 && consume_stock(mem))
2298                         goto done;
2299                 css_get(&mem->css);
2300         } else {
2301                 struct task_struct *p;
2302
2303                 rcu_read_lock();
2304                 p = rcu_dereference(mm->owner);
2305                 /*
2306                  * Because we don't have task_lock(), "p" can exit.
2307                  * In that case, "mem" can point to root or p can be NULL with
2308                  * race with swapoff. Then, we have small risk of mis-accouning.
2309                  * But such kind of mis-account by race always happens because
2310                  * we don't have cgroup_mutex(). It's overkill and we allo that
2311                  * small race, here.
2312                  * (*) swapoff at el will charge against mm-struct not against
2313                  * task-struct. So, mm->owner can be NULL.
2314                  */
2315                 mem = mem_cgroup_from_task(p);
2316                 if (!mem || mem_cgroup_is_root(mem)) {
2317                         rcu_read_unlock();
2318                         goto done;
2319                 }
2320                 if (nr_pages == 1 && consume_stock(mem)) {
2321                         /*
2322                          * It seems dagerous to access memcg without css_get().
2323                          * But considering how consume_stok works, it's not
2324                          * necessary. If consume_stock success, some charges
2325                          * from this memcg are cached on this cpu. So, we
2326                          * don't need to call css_get()/css_tryget() before
2327                          * calling consume_stock().
2328                          */
2329                         rcu_read_unlock();
2330                         goto done;
2331                 }
2332                 /* after here, we may be blocked. we need to get refcnt */
2333                 if (!css_tryget(&mem->css)) {
2334                         rcu_read_unlock();
2335                         goto again;
2336                 }
2337                 rcu_read_unlock();
2338         }
2339
2340         do {
2341                 bool oom_check;
2342
2343                 /* If killed, bypass charge */
2344                 if (fatal_signal_pending(current)) {
2345                         css_put(&mem->css);
2346                         goto bypass;
2347                 }
2348
2349                 oom_check = false;
2350                 if (oom && !nr_oom_retries) {
2351                         oom_check = true;
2352                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2353                 }
2354
2355                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2356                 switch (ret) {
2357                 case CHARGE_OK:
2358                         break;
2359                 case CHARGE_RETRY: /* not in OOM situation but retry */
2360                         batch = nr_pages;
2361                         css_put(&mem->css);
2362                         mem = NULL;
2363                         goto again;
2364                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2365                         css_put(&mem->css);
2366                         goto nomem;
2367                 case CHARGE_NOMEM: /* OOM routine works */
2368                         if (!oom) {
2369                                 css_put(&mem->css);
2370                                 goto nomem;
2371                         }
2372                         /* If oom, we never return -ENOMEM */
2373                         nr_oom_retries--;
2374                         break;
2375                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2376                         css_put(&mem->css);
2377                         goto bypass;
2378                 }
2379         } while (ret != CHARGE_OK);
2380
2381         if (batch > nr_pages)
2382                 refill_stock(mem, batch - nr_pages);
2383         css_put(&mem->css);
2384 done:
2385         *memcg = mem;
2386         return 0;
2387 nomem:
2388         *memcg = NULL;
2389         return -ENOMEM;
2390 bypass:
2391         *memcg = NULL;
2392         return 0;
2393 }
2394
2395 /*
2396  * Somemtimes we have to undo a charge we got by try_charge().
2397  * This function is for that and do uncharge, put css's refcnt.
2398  * gotten by try_charge().
2399  */
2400 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2401                                        unsigned int nr_pages)
2402 {
2403         if (!mem_cgroup_is_root(mem)) {
2404                 unsigned long bytes = nr_pages * PAGE_SIZE;
2405
2406                 res_counter_uncharge(&mem->res, bytes);
2407                 if (do_swap_account)
2408                         res_counter_uncharge(&mem->memsw, bytes);
2409         }
2410 }
2411
2412 /*
2413  * A helper function to get mem_cgroup from ID. must be called under
2414  * rcu_read_lock(). The caller must check css_is_removed() or some if
2415  * it's concern. (dropping refcnt from swap can be called against removed
2416  * memcg.)
2417  */
2418 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2419 {
2420         struct cgroup_subsys_state *css;
2421
2422         /* ID 0 is unused ID */
2423         if (!id)
2424                 return NULL;
2425         css = css_lookup(&mem_cgroup_subsys, id);
2426         if (!css)
2427                 return NULL;
2428         return container_of(css, struct mem_cgroup, css);
2429 }
2430
2431 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2432 {
2433         struct mem_cgroup *mem = NULL;
2434         struct page_cgroup *pc;
2435         unsigned short id;
2436         swp_entry_t ent;
2437
2438         VM_BUG_ON(!PageLocked(page));
2439
2440         pc = lookup_page_cgroup(page);
2441         lock_page_cgroup(pc);
2442         if (PageCgroupUsed(pc)) {
2443                 mem = pc->mem_cgroup;
2444                 if (mem && !css_tryget(&mem->css))
2445                         mem = NULL;
2446         } else if (PageSwapCache(page)) {
2447                 ent.val = page_private(page);
2448                 id = lookup_swap_cgroup(ent);
2449                 rcu_read_lock();
2450                 mem = mem_cgroup_lookup(id);
2451                 if (mem && !css_tryget(&mem->css))
2452                         mem = NULL;
2453                 rcu_read_unlock();
2454         }
2455         unlock_page_cgroup(pc);
2456         return mem;
2457 }
2458
2459 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2460                                        struct page *page,
2461                                        unsigned int nr_pages,
2462                                        struct page_cgroup *pc,
2463                                        enum charge_type ctype)
2464 {
2465         lock_page_cgroup(pc);
2466         if (unlikely(PageCgroupUsed(pc))) {
2467                 unlock_page_cgroup(pc);
2468                 __mem_cgroup_cancel_charge(mem, nr_pages);
2469                 return;
2470         }
2471         /*
2472          * we don't need page_cgroup_lock about tail pages, becase they are not
2473          * accessed by any other context at this point.
2474          */
2475         pc->mem_cgroup = mem;
2476         /*
2477          * We access a page_cgroup asynchronously without lock_page_cgroup().
2478          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2479          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2480          * before USED bit, we need memory barrier here.
2481          * See mem_cgroup_add_lru_list(), etc.
2482          */
2483         smp_wmb();
2484         switch (ctype) {
2485         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2486         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2487                 SetPageCgroupCache(pc);
2488                 SetPageCgroupUsed(pc);
2489                 break;
2490         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2491                 ClearPageCgroupCache(pc);
2492                 SetPageCgroupUsed(pc);
2493                 break;
2494         default:
2495                 break;
2496         }
2497
2498         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2499         unlock_page_cgroup(pc);
2500         /*
2501          * "charge_statistics" updated event counter. Then, check it.
2502          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2503          * if they exceeds softlimit.
2504          */
2505         memcg_check_events(mem, page);
2506 }
2507
2508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2509
2510 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2511                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2512 /*
2513  * Because tail pages are not marked as "used", set it. We're under
2514  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2515  */
2516 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2517 {
2518         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2519         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2520         unsigned long flags;
2521
2522         if (mem_cgroup_disabled())
2523                 return;
2524         /*
2525          * We have no races with charge/uncharge but will have races with
2526          * page state accounting.
2527          */
2528         move_lock_page_cgroup(head_pc, &flags);
2529
2530         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2531         smp_wmb(); /* see __commit_charge() */
2532         if (PageCgroupAcctLRU(head_pc)) {
2533                 enum lru_list lru;
2534                 struct mem_cgroup_per_zone *mz;
2535
2536                 /*
2537                  * LRU flags cannot be copied because we need to add tail
2538                  *.page to LRU by generic call and our hook will be called.
2539                  * We hold lru_lock, then, reduce counter directly.
2540                  */
2541                 lru = page_lru(head);
2542                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2543                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2544         }
2545         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2546         move_unlock_page_cgroup(head_pc, &flags);
2547 }
2548 #endif
2549
2550 /**
2551  * mem_cgroup_move_account - move account of the page
2552  * @page: the page
2553  * @nr_pages: number of regular pages (>1 for huge pages)
2554  * @pc: page_cgroup of the page.
2555  * @from: mem_cgroup which the page is moved from.
2556  * @to: mem_cgroup which the page is moved to. @from != @to.
2557  * @uncharge: whether we should call uncharge and css_put against @from.
2558  *
2559  * The caller must confirm following.
2560  * - page is not on LRU (isolate_page() is useful.)
2561  * - compound_lock is held when nr_pages > 1
2562  *
2563  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2564  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2565  * true, this function does "uncharge" from old cgroup, but it doesn't if
2566  * @uncharge is false, so a caller should do "uncharge".
2567  */
2568 static int mem_cgroup_move_account(struct page *page,
2569                                    unsigned int nr_pages,
2570                                    struct page_cgroup *pc,
2571                                    struct mem_cgroup *from,
2572                                    struct mem_cgroup *to,
2573                                    bool uncharge)
2574 {
2575         unsigned long flags;
2576         int ret;
2577
2578         VM_BUG_ON(from == to);
2579         VM_BUG_ON(PageLRU(page));
2580         /*
2581          * The page is isolated from LRU. So, collapse function
2582          * will not handle this page. But page splitting can happen.
2583          * Do this check under compound_page_lock(). The caller should
2584          * hold it.
2585          */
2586         ret = -EBUSY;
2587         if (nr_pages > 1 && !PageTransHuge(page))
2588                 goto out;
2589
2590         lock_page_cgroup(pc);
2591
2592         ret = -EINVAL;
2593         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2594                 goto unlock;
2595
2596         move_lock_page_cgroup(pc, &flags);
2597
2598         if (PageCgroupFileMapped(pc)) {
2599                 /* Update mapped_file data for mem_cgroup */
2600                 preempt_disable();
2601                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2602                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2603                 preempt_enable();
2604         }
2605         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2606         if (uncharge)
2607                 /* This is not "cancel", but cancel_charge does all we need. */
2608                 __mem_cgroup_cancel_charge(from, nr_pages);
2609
2610         /* caller should have done css_get */
2611         pc->mem_cgroup = to;
2612         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2613         /*
2614          * We charges against "to" which may not have any tasks. Then, "to"
2615          * can be under rmdir(). But in current implementation, caller of
2616          * this function is just force_empty() and move charge, so it's
2617          * guaranteed that "to" is never removed. So, we don't check rmdir
2618          * status here.
2619          */
2620         move_unlock_page_cgroup(pc, &flags);
2621         ret = 0;
2622 unlock:
2623         unlock_page_cgroup(pc);
2624         /*
2625          * check events
2626          */
2627         memcg_check_events(to, page);
2628         memcg_check_events(from, page);
2629 out:
2630         return ret;
2631 }
2632
2633 /*
2634  * move charges to its parent.
2635  */
2636
2637 static int mem_cgroup_move_parent(struct page *page,
2638                                   struct page_cgroup *pc,
2639                                   struct mem_cgroup *child,
2640                                   gfp_t gfp_mask)
2641 {
2642         struct cgroup *cg = child->css.cgroup;
2643         struct cgroup *pcg = cg->parent;
2644         struct mem_cgroup *parent;
2645         unsigned int nr_pages;
2646         unsigned long uninitialized_var(flags);
2647         int ret;
2648
2649         /* Is ROOT ? */
2650         if (!pcg)
2651                 return -EINVAL;
2652
2653         ret = -EBUSY;
2654         if (!get_page_unless_zero(page))
2655                 goto out;
2656         if (isolate_lru_page(page))
2657                 goto put;
2658
2659         nr_pages = hpage_nr_pages(page);
2660
2661         parent = mem_cgroup_from_cont(pcg);
2662         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2663         if (ret || !parent)
2664                 goto put_back;
2665
2666         if (nr_pages > 1)
2667                 flags = compound_lock_irqsave(page);
2668
2669         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2670         if (ret)
2671                 __mem_cgroup_cancel_charge(parent, nr_pages);
2672
2673         if (nr_pages > 1)
2674                 compound_unlock_irqrestore(page, flags);
2675 put_back:
2676         putback_lru_page(page);
2677 put:
2678         put_page(page);
2679 out:
2680         return ret;
2681 }
2682
2683 /*
2684  * Charge the memory controller for page usage.
2685  * Return
2686  * 0 if the charge was successful
2687  * < 0 if the cgroup is over its limit
2688  */
2689 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2690                                 gfp_t gfp_mask, enum charge_type ctype)
2691 {
2692         struct mem_cgroup *mem = NULL;
2693         unsigned int nr_pages = 1;
2694         struct page_cgroup *pc;
2695         bool oom = true;
2696         int ret;
2697
2698         if (PageTransHuge(page)) {
2699                 nr_pages <<= compound_order(page);
2700                 VM_BUG_ON(!PageTransHuge(page));
2701                 /*
2702                  * Never OOM-kill a process for a huge page.  The
2703                  * fault handler will fall back to regular pages.
2704                  */
2705                 oom = false;
2706         }
2707
2708         pc = lookup_page_cgroup(page);
2709         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2710
2711         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2712         if (ret || !mem)
2713                 return ret;
2714
2715         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2716         return 0;
2717 }
2718
2719 int mem_cgroup_newpage_charge(struct page *page,
2720                               struct mm_struct *mm, gfp_t gfp_mask)
2721 {
2722         if (mem_cgroup_disabled())
2723                 return 0;
2724         /*
2725          * If already mapped, we don't have to account.
2726          * If page cache, page->mapping has address_space.
2727          * But page->mapping may have out-of-use anon_vma pointer,
2728          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2729          * is NULL.
2730          */
2731         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2732                 return 0;
2733         if (unlikely(!mm))
2734                 mm = &init_mm;
2735         return mem_cgroup_charge_common(page, mm, gfp_mask,
2736                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2737 }
2738
2739 static void
2740 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2741                                         enum charge_type ctype);
2742
2743 static void
2744 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2745                                         enum charge_type ctype)
2746 {
2747         struct page_cgroup *pc = lookup_page_cgroup(page);
2748         /*
2749          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2750          * is already on LRU. It means the page may on some other page_cgroup's
2751          * LRU. Take care of it.
2752          */
2753         mem_cgroup_lru_del_before_commit(page);
2754         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2755         mem_cgroup_lru_add_after_commit(page);
2756         return;
2757 }
2758
2759 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2760                                 gfp_t gfp_mask)
2761 {
2762         struct mem_cgroup *mem = NULL;
2763         int ret;
2764
2765         if (mem_cgroup_disabled())
2766                 return 0;
2767         if (PageCompound(page))
2768                 return 0;
2769
2770         if (unlikely(!mm))
2771                 mm = &init_mm;
2772
2773         if (page_is_file_cache(page)) {
2774                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2775                 if (ret || !mem)
2776                         return ret;
2777
2778                 /*
2779                  * FUSE reuses pages without going through the final
2780                  * put that would remove them from the LRU list, make
2781                  * sure that they get relinked properly.
2782                  */
2783                 __mem_cgroup_commit_charge_lrucare(page, mem,
2784                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2785                 return ret;
2786         }
2787         /* shmem */
2788         if (PageSwapCache(page)) {
2789                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2790                 if (!ret)
2791                         __mem_cgroup_commit_charge_swapin(page, mem,
2792                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2793         } else
2794                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2795                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2796
2797         return ret;
2798 }
2799
2800 /*
2801  * While swap-in, try_charge -> commit or cancel, the page is locked.
2802  * And when try_charge() successfully returns, one refcnt to memcg without
2803  * struct page_cgroup is acquired. This refcnt will be consumed by
2804  * "commit()" or removed by "cancel()"
2805  */
2806 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2807                                  struct page *page,
2808                                  gfp_t mask, struct mem_cgroup **ptr)
2809 {
2810         struct mem_cgroup *mem;
2811         int ret;
2812
2813         *ptr = NULL;
2814
2815         if (mem_cgroup_disabled())
2816                 return 0;
2817
2818         if (!do_swap_account)
2819                 goto charge_cur_mm;
2820         /*
2821          * A racing thread's fault, or swapoff, may have already updated
2822          * the pte, and even removed page from swap cache: in those cases
2823          * do_swap_page()'s pte_same() test will fail; but there's also a
2824          * KSM case which does need to charge the page.
2825          */
2826         if (!PageSwapCache(page))
2827                 goto charge_cur_mm;
2828         mem = try_get_mem_cgroup_from_page(page);
2829         if (!mem)
2830                 goto charge_cur_mm;
2831         *ptr = mem;
2832         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2833         css_put(&mem->css);
2834         return ret;
2835 charge_cur_mm:
2836         if (unlikely(!mm))
2837                 mm = &init_mm;
2838         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2839 }
2840
2841 static void
2842 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2843                                         enum charge_type ctype)
2844 {
2845         if (mem_cgroup_disabled())
2846                 return;
2847         if (!ptr)
2848                 return;
2849         cgroup_exclude_rmdir(&ptr->css);
2850
2851         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2852         /*
2853          * Now swap is on-memory. This means this page may be
2854          * counted both as mem and swap....double count.
2855          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2856          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2857          * may call delete_from_swap_cache() before reach here.
2858          */
2859         if (do_swap_account && PageSwapCache(page)) {
2860                 swp_entry_t ent = {.val = page_private(page)};
2861                 unsigned short id;
2862                 struct mem_cgroup *memcg;
2863
2864                 id = swap_cgroup_record(ent, 0);
2865                 rcu_read_lock();
2866                 memcg = mem_cgroup_lookup(id);
2867                 if (memcg) {
2868                         /*
2869                          * This recorded memcg can be obsolete one. So, avoid
2870                          * calling css_tryget
2871                          */
2872                         if (!mem_cgroup_is_root(memcg))
2873                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2874                         mem_cgroup_swap_statistics(memcg, false);
2875                         mem_cgroup_put(memcg);
2876                 }
2877                 rcu_read_unlock();
2878         }
2879         /*
2880          * At swapin, we may charge account against cgroup which has no tasks.
2881          * So, rmdir()->pre_destroy() can be called while we do this charge.
2882          * In that case, we need to call pre_destroy() again. check it here.
2883          */
2884         cgroup_release_and_wakeup_rmdir(&ptr->css);
2885 }
2886
2887 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2888 {
2889         __mem_cgroup_commit_charge_swapin(page, ptr,
2890                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2891 }
2892
2893 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2894 {
2895         if (mem_cgroup_disabled())
2896                 return;
2897         if (!mem)
2898                 return;
2899         __mem_cgroup_cancel_charge(mem, 1);
2900 }
2901
2902 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2903                                    unsigned int nr_pages,
2904                                    const enum charge_type ctype)
2905 {
2906         struct memcg_batch_info *batch = NULL;
2907         bool uncharge_memsw = true;
2908
2909         /* If swapout, usage of swap doesn't decrease */
2910         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2911                 uncharge_memsw = false;
2912
2913         batch = &current->memcg_batch;
2914         /*
2915          * In usual, we do css_get() when we remember memcg pointer.
2916          * But in this case, we keep res->usage until end of a series of
2917          * uncharges. Then, it's ok to ignore memcg's refcnt.
2918          */
2919         if (!batch->memcg)
2920                 batch->memcg = mem;
2921         /*
2922          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2923          * In those cases, all pages freed continuously can be expected to be in
2924          * the same cgroup and we have chance to coalesce uncharges.
2925          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2926          * because we want to do uncharge as soon as possible.
2927          */
2928
2929         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2930                 goto direct_uncharge;
2931
2932         if (nr_pages > 1)
2933                 goto direct_uncharge;
2934
2935         /*
2936          * In typical case, batch->memcg == mem. This means we can
2937          * merge a series of uncharges to an uncharge of res_counter.
2938          * If not, we uncharge res_counter ony by one.
2939          */
2940         if (batch->memcg != mem)
2941                 goto direct_uncharge;
2942         /* remember freed charge and uncharge it later */
2943         batch->nr_pages++;
2944         if (uncharge_memsw)
2945                 batch->memsw_nr_pages++;
2946         return;
2947 direct_uncharge:
2948         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2949         if (uncharge_memsw)
2950                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2951         if (unlikely(batch->memcg != mem))
2952                 memcg_oom_recover(mem);
2953         return;
2954 }
2955
2956 /*
2957  * uncharge if !page_mapped(page)
2958  */
2959 static struct mem_cgroup *
2960 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2961 {
2962         struct mem_cgroup *mem = NULL;
2963         unsigned int nr_pages = 1;
2964         struct page_cgroup *pc;
2965
2966         if (mem_cgroup_disabled())
2967                 return NULL;
2968
2969         if (PageSwapCache(page))
2970                 return NULL;
2971
2972         if (PageTransHuge(page)) {
2973                 nr_pages <<= compound_order(page);
2974                 VM_BUG_ON(!PageTransHuge(page));
2975         }
2976         /*
2977          * Check if our page_cgroup is valid
2978          */
2979         pc = lookup_page_cgroup(page);
2980         if (unlikely(!pc || !PageCgroupUsed(pc)))
2981                 return NULL;
2982
2983         lock_page_cgroup(pc);
2984
2985         mem = pc->mem_cgroup;
2986
2987         if (!PageCgroupUsed(pc))
2988                 goto unlock_out;
2989
2990         switch (ctype) {
2991         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2992         case MEM_CGROUP_CHARGE_TYPE_DROP:
2993                 /* See mem_cgroup_prepare_migration() */
2994                 if (page_mapped(page) || PageCgroupMigration(pc))
2995                         goto unlock_out;
2996                 break;
2997         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2998                 if (!PageAnon(page)) {  /* Shared memory */
2999                         if (page->mapping && !page_is_file_cache(page))
3000                                 goto unlock_out;
3001                 } else if (page_mapped(page)) /* Anon */
3002                                 goto unlock_out;
3003                 break;
3004         default:
3005                 break;
3006         }
3007
3008         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3009
3010         ClearPageCgroupUsed(pc);
3011         /*
3012          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3013          * freed from LRU. This is safe because uncharged page is expected not
3014          * to be reused (freed soon). Exception is SwapCache, it's handled by
3015          * special functions.
3016          */
3017
3018         unlock_page_cgroup(pc);
3019         /*
3020          * even after unlock, we have mem->res.usage here and this memcg
3021          * will never be freed.
3022          */
3023         memcg_check_events(mem, page);
3024         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3025                 mem_cgroup_swap_statistics(mem, true);
3026                 mem_cgroup_get(mem);
3027         }
3028         if (!mem_cgroup_is_root(mem))
3029                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3030
3031         return mem;
3032
3033 unlock_out:
3034         unlock_page_cgroup(pc);
3035         return NULL;
3036 }
3037
3038 void mem_cgroup_uncharge_page(struct page *page)
3039 {
3040         /* early check. */
3041         if (page_mapped(page))
3042                 return;
3043         if (page->mapping && !PageAnon(page))
3044                 return;
3045         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3046 }
3047
3048 void mem_cgroup_uncharge_cache_page(struct page *page)
3049 {
3050         VM_BUG_ON(page_mapped(page));
3051         VM_BUG_ON(page->mapping);
3052         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3053 }
3054
3055 /*
3056  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3057  * In that cases, pages are freed continuously and we can expect pages
3058  * are in the same memcg. All these calls itself limits the number of
3059  * pages freed at once, then uncharge_start/end() is called properly.
3060  * This may be called prural(2) times in a context,
3061  */
3062
3063 void mem_cgroup_uncharge_start(void)
3064 {
3065         current->memcg_batch.do_batch++;
3066         /* We can do nest. */
3067         if (current->memcg_batch.do_batch == 1) {
3068                 current->memcg_batch.memcg = NULL;
3069                 current->memcg_batch.nr_pages = 0;
3070                 current->memcg_batch.memsw_nr_pages = 0;
3071         }
3072 }
3073
3074 void mem_cgroup_uncharge_end(void)
3075 {
3076         struct memcg_batch_info *batch = &current->memcg_batch;
3077
3078         if (!batch->do_batch)
3079                 return;
3080
3081         batch->do_batch--;
3082         if (batch->do_batch) /* If stacked, do nothing. */
3083                 return;
3084
3085         if (!batch->memcg)
3086                 return;
3087         /*
3088          * This "batch->memcg" is valid without any css_get/put etc...
3089          * bacause we hide charges behind us.
3090          */
3091         if (batch->nr_pages)
3092                 res_counter_uncharge(&batch->memcg->res,
3093                                      batch->nr_pages * PAGE_SIZE);
3094         if (batch->memsw_nr_pages)
3095                 res_counter_uncharge(&batch->memcg->memsw,
3096                                      batch->memsw_nr_pages * PAGE_SIZE);
3097         memcg_oom_recover(batch->memcg);
3098         /* forget this pointer (for sanity check) */
3099         batch->memcg = NULL;
3100 }
3101
3102 #ifdef CONFIG_SWAP
3103 /*
3104  * called after __delete_from_swap_cache() and drop "page" account.
3105  * memcg information is recorded to swap_cgroup of "ent"
3106  */
3107 void
3108 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3109 {
3110         struct mem_cgroup *memcg;
3111         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3112
3113         if (!swapout) /* this was a swap cache but the swap is unused ! */
3114                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3115
3116         memcg = __mem_cgroup_uncharge_common(page, ctype);
3117
3118         /*
3119          * record memcg information,  if swapout && memcg != NULL,
3120          * mem_cgroup_get() was called in uncharge().
3121          */
3122         if (do_swap_account && swapout && memcg)
3123                 swap_cgroup_record(ent, css_id(&memcg->css));
3124 }
3125 #endif
3126
3127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3128 /*
3129  * called from swap_entry_free(). remove record in swap_cgroup and
3130  * uncharge "memsw" account.
3131  */
3132 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3133 {
3134         struct mem_cgroup *memcg;
3135         unsigned short id;
3136
3137         if (!do_swap_account)
3138                 return;
3139
3140         id = swap_cgroup_record(ent, 0);
3141         rcu_read_lock();
3142         memcg = mem_cgroup_lookup(id);
3143         if (memcg) {
3144                 /*
3145                  * We uncharge this because swap is freed.
3146                  * This memcg can be obsolete one. We avoid calling css_tryget
3147                  */
3148                 if (!mem_cgroup_is_root(memcg))
3149                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3150                 mem_cgroup_swap_statistics(memcg, false);
3151                 mem_cgroup_put(memcg);
3152         }
3153         rcu_read_unlock();
3154 }
3155
3156 /**
3157  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3158  * @entry: swap entry to be moved
3159  * @from:  mem_cgroup which the entry is moved from
3160  * @to:  mem_cgroup which the entry is moved to
3161  * @need_fixup: whether we should fixup res_counters and refcounts.
3162  *
3163  * It succeeds only when the swap_cgroup's record for this entry is the same
3164  * as the mem_cgroup's id of @from.
3165  *
3166  * Returns 0 on success, -EINVAL on failure.
3167  *
3168  * The caller must have charged to @to, IOW, called res_counter_charge() about
3169  * both res and memsw, and called css_get().
3170  */
3171 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3172                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3173 {
3174         unsigned short old_id, new_id;
3175
3176         old_id = css_id(&from->css);
3177         new_id = css_id(&to->css);
3178
3179         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3180                 mem_cgroup_swap_statistics(from, false);
3181                 mem_cgroup_swap_statistics(to, true);
3182                 /*
3183                  * This function is only called from task migration context now.
3184                  * It postpones res_counter and refcount handling till the end
3185                  * of task migration(mem_cgroup_clear_mc()) for performance
3186                  * improvement. But we cannot postpone mem_cgroup_get(to)
3187                  * because if the process that has been moved to @to does
3188                  * swap-in, the refcount of @to might be decreased to 0.
3189                  */
3190                 mem_cgroup_get(to);
3191                 if (need_fixup) {
3192                         if (!mem_cgroup_is_root(from))
3193                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3194                         mem_cgroup_put(from);
3195                         /*
3196                          * we charged both to->res and to->memsw, so we should
3197                          * uncharge to->res.
3198                          */
3199                         if (!mem_cgroup_is_root(to))
3200                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3201                 }
3202                 return 0;
3203         }
3204         return -EINVAL;
3205 }
3206 #else
3207 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3208                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3209 {
3210         return -EINVAL;
3211 }
3212 #endif
3213
3214 /*
3215  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3216  * page belongs to.
3217  */
3218 int mem_cgroup_prepare_migration(struct page *page,
3219         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3220 {
3221         struct mem_cgroup *mem = NULL;
3222         struct page_cgroup *pc;
3223         enum charge_type ctype;
3224         int ret = 0;
3225
3226         *ptr = NULL;
3227
3228         VM_BUG_ON(PageTransHuge(page));
3229         if (mem_cgroup_disabled())
3230                 return 0;
3231
3232         pc = lookup_page_cgroup(page);
3233         lock_page_cgroup(pc);
3234         if (PageCgroupUsed(pc)) {
3235                 mem = pc->mem_cgroup;
3236                 css_get(&mem->css);
3237                 /*
3238                  * At migrating an anonymous page, its mapcount goes down
3239                  * to 0 and uncharge() will be called. But, even if it's fully
3240                  * unmapped, migration may fail and this page has to be
3241                  * charged again. We set MIGRATION flag here and delay uncharge
3242                  * until end_migration() is called
3243                  *
3244                  * Corner Case Thinking
3245                  * A)
3246                  * When the old page was mapped as Anon and it's unmap-and-freed
3247                  * while migration was ongoing.
3248                  * If unmap finds the old page, uncharge() of it will be delayed
3249                  * until end_migration(). If unmap finds a new page, it's
3250                  * uncharged when it make mapcount to be 1->0. If unmap code
3251                  * finds swap_migration_entry, the new page will not be mapped
3252                  * and end_migration() will find it(mapcount==0).
3253                  *
3254                  * B)
3255                  * When the old page was mapped but migraion fails, the kernel
3256                  * remaps it. A charge for it is kept by MIGRATION flag even
3257                  * if mapcount goes down to 0. We can do remap successfully
3258                  * without charging it again.
3259                  *
3260                  * C)
3261                  * The "old" page is under lock_page() until the end of
3262                  * migration, so, the old page itself will not be swapped-out.
3263                  * If the new page is swapped out before end_migraton, our
3264                  * hook to usual swap-out path will catch the event.
3265                  */
3266                 if (PageAnon(page))
3267                         SetPageCgroupMigration(pc);
3268         }
3269         unlock_page_cgroup(pc);
3270         /*
3271          * If the page is not charged at this point,
3272          * we return here.
3273          */
3274         if (!mem)
3275                 return 0;
3276
3277         *ptr = mem;
3278         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3279         css_put(&mem->css);/* drop extra refcnt */
3280         if (ret || *ptr == NULL) {
3281                 if (PageAnon(page)) {
3282                         lock_page_cgroup(pc);
3283                         ClearPageCgroupMigration(pc);
3284                         unlock_page_cgroup(pc);
3285                         /*
3286                          * The old page may be fully unmapped while we kept it.
3287                          */
3288                         mem_cgroup_uncharge_page(page);
3289                 }
3290                 return -ENOMEM;
3291         }
3292         /*
3293          * We charge new page before it's used/mapped. So, even if unlock_page()
3294          * is called before end_migration, we can catch all events on this new
3295          * page. In the case new page is migrated but not remapped, new page's
3296          * mapcount will be finally 0 and we call uncharge in end_migration().
3297          */
3298         pc = lookup_page_cgroup(newpage);
3299         if (PageAnon(page))
3300                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3301         else if (page_is_file_cache(page))
3302                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3303         else
3304                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3305         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3306         return ret;
3307 }
3308
3309 /* remove redundant charge if migration failed*/
3310 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3311         struct page *oldpage, struct page *newpage, bool migration_ok)
3312 {
3313         struct page *used, *unused;
3314         struct page_cgroup *pc;
3315
3316         if (!mem)
3317                 return;
3318         /* blocks rmdir() */
3319         cgroup_exclude_rmdir(&mem->css);
3320         if (!migration_ok) {
3321                 used = oldpage;
3322                 unused = newpage;
3323         } else {
3324                 used = newpage;
3325                 unused = oldpage;
3326         }
3327         /*
3328          * We disallowed uncharge of pages under migration because mapcount
3329          * of the page goes down to zero, temporarly.
3330          * Clear the flag and check the page should be charged.
3331          */
3332         pc = lookup_page_cgroup(oldpage);
3333         lock_page_cgroup(pc);
3334         ClearPageCgroupMigration(pc);
3335         unlock_page_cgroup(pc);
3336
3337         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3338
3339         /*
3340          * If a page is a file cache, radix-tree replacement is very atomic
3341          * and we can skip this check. When it was an Anon page, its mapcount
3342          * goes down to 0. But because we added MIGRATION flage, it's not
3343          * uncharged yet. There are several case but page->mapcount check
3344          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3345          * check. (see prepare_charge() also)
3346          */
3347         if (PageAnon(used))
3348                 mem_cgroup_uncharge_page(used);
3349         /*
3350          * At migration, we may charge account against cgroup which has no
3351          * tasks.
3352          * So, rmdir()->pre_destroy() can be called while we do this charge.
3353          * In that case, we need to call pre_destroy() again. check it here.
3354          */
3355         cgroup_release_and_wakeup_rmdir(&mem->css);
3356 }
3357
3358 #ifdef CONFIG_DEBUG_VM
3359 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3360 {
3361         struct page_cgroup *pc;
3362
3363         pc = lookup_page_cgroup(page);
3364         if (likely(pc) && PageCgroupUsed(pc))
3365                 return pc;
3366         return NULL;
3367 }
3368
3369 bool mem_cgroup_bad_page_check(struct page *page)
3370 {
3371         if (mem_cgroup_disabled())
3372                 return false;
3373
3374         return lookup_page_cgroup_used(page) != NULL;
3375 }
3376
3377 void mem_cgroup_print_bad_page(struct page *page)
3378 {
3379         struct page_cgroup *pc;
3380
3381         pc = lookup_page_cgroup_used(page);
3382         if (pc) {
3383                 int ret = -1;
3384                 char *path;
3385
3386                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3387                        pc, pc->flags, pc->mem_cgroup);
3388
3389                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3390                 if (path) {
3391                         rcu_read_lock();
3392                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3393                                                         path, PATH_MAX);
3394                         rcu_read_unlock();
3395                 }
3396
3397                 printk(KERN_CONT "(%s)\n",
3398                                 (ret < 0) ? "cannot get the path" : path);
3399                 kfree(path);
3400         }
3401 }
3402 #endif
3403
3404 static DEFINE_MUTEX(set_limit_mutex);
3405
3406 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3407                                 unsigned long long val)
3408 {
3409         int retry_count;
3410         u64 memswlimit, memlimit;
3411         int ret = 0;
3412         int children = mem_cgroup_count_children(memcg);
3413         u64 curusage, oldusage;
3414         int enlarge;
3415
3416         /*
3417          * For keeping hierarchical_reclaim simple, how long we should retry
3418          * is depends on callers. We set our retry-count to be function
3419          * of # of children which we should visit in this loop.
3420          */
3421         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3422
3423         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3424
3425         enlarge = 0;
3426         while (retry_count) {
3427                 if (signal_pending(current)) {
3428                         ret = -EINTR;
3429                         break;
3430                 }
3431                 /*
3432                  * Rather than hide all in some function, I do this in
3433                  * open coded manner. You see what this really does.
3434                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3435                  */
3436                 mutex_lock(&set_limit_mutex);
3437                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3438                 if (memswlimit < val) {
3439                         ret = -EINVAL;
3440                         mutex_unlock(&set_limit_mutex);
3441                         break;
3442                 }
3443
3444                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3445                 if (memlimit < val)
3446                         enlarge = 1;
3447
3448                 ret = res_counter_set_limit(&memcg->res, val);
3449                 if (!ret) {
3450                         if (memswlimit == val)
3451                                 memcg->memsw_is_minimum = true;
3452                         else
3453                                 memcg->memsw_is_minimum = false;
3454                 }
3455                 mutex_unlock(&set_limit_mutex);
3456
3457                 if (!ret)
3458                         break;
3459
3460                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3461                                                 MEM_CGROUP_RECLAIM_SHRINK,
3462                                                 NULL);
3463                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3464                 /* Usage is reduced ? */
3465                 if (curusage >= oldusage)
3466                         retry_count--;
3467                 else
3468                         oldusage = curusage;
3469         }
3470         if (!ret && enlarge)
3471                 memcg_oom_recover(memcg);
3472
3473         return ret;
3474 }
3475
3476 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3477                                         unsigned long long val)
3478 {
3479         int retry_count;
3480         u64 memlimit, memswlimit, oldusage, curusage;
3481         int children = mem_cgroup_count_children(memcg);
3482         int ret = -EBUSY;
3483         int enlarge = 0;
3484
3485         /* see mem_cgroup_resize_res_limit */
3486         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3487         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3488         while (retry_count) {
3489                 if (signal_pending(current)) {
3490                         ret = -EINTR;
3491                         break;
3492                 }
3493                 /*
3494                  * Rather than hide all in some function, I do this in
3495                  * open coded manner. You see what this really does.
3496                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3497                  */
3498                 mutex_lock(&set_limit_mutex);
3499                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3500                 if (memlimit > val) {
3501                         ret = -EINVAL;
3502                         mutex_unlock(&set_limit_mutex);
3503                         break;
3504                 }
3505                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3506                 if (memswlimit < val)
3507                         enlarge = 1;
3508                 ret = res_counter_set_limit(&memcg->memsw, val);
3509                 if (!ret) {
3510                         if (memlimit == val)
3511                                 memcg->memsw_is_minimum = true;
3512                         else
3513                                 memcg->memsw_is_minimum = false;
3514                 }
3515                 mutex_unlock(&set_limit_mutex);
3516
3517                 if (!ret)
3518                         break;
3519
3520                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3521                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3522                                                 MEM_CGROUP_RECLAIM_SHRINK,
3523                                                 NULL);
3524                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3525                 /* Usage is reduced ? */
3526                 if (curusage >= oldusage)
3527                         retry_count--;
3528                 else
3529                         oldusage = curusage;
3530         }
3531         if (!ret && enlarge)
3532                 memcg_oom_recover(memcg);
3533         return ret;
3534 }
3535
3536 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3537                                             gfp_t gfp_mask,
3538                                             unsigned long *total_scanned)
3539 {
3540         unsigned long nr_reclaimed = 0;
3541         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3542         unsigned long reclaimed;
3543         int loop = 0;
3544         struct mem_cgroup_tree_per_zone *mctz;
3545         unsigned long long excess;
3546         unsigned long nr_scanned;
3547
3548         if (order > 0)
3549                 return 0;
3550
3551         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3552         /*
3553          * This loop can run a while, specially if mem_cgroup's continuously
3554          * keep exceeding their soft limit and putting the system under
3555          * pressure
3556          */
3557         do {
3558                 if (next_mz)
3559                         mz = next_mz;
3560                 else
3561                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3562                 if (!mz)
3563                         break;
3564
3565                 nr_scanned = 0;
3566                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3567                                                 gfp_mask,
3568                                                 MEM_CGROUP_RECLAIM_SOFT,
3569                                                 &nr_scanned);
3570                 nr_reclaimed += reclaimed;
3571                 *total_scanned += nr_scanned;
3572                 spin_lock(&mctz->lock);
3573
3574                 /*
3575                  * If we failed to reclaim anything from this memory cgroup
3576                  * it is time to move on to the next cgroup
3577                  */
3578                 next_mz = NULL;
3579                 if (!reclaimed) {
3580                         do {
3581                                 /*
3582                                  * Loop until we find yet another one.
3583                                  *
3584                                  * By the time we get the soft_limit lock
3585                                  * again, someone might have aded the
3586                                  * group back on the RB tree. Iterate to
3587                                  * make sure we get a different mem.
3588                                  * mem_cgroup_largest_soft_limit_node returns
3589                                  * NULL if no other cgroup is present on
3590                                  * the tree
3591                                  */
3592                                 next_mz =
3593                                 __mem_cgroup_largest_soft_limit_node(mctz);
3594                                 if (next_mz == mz)
3595                                         css_put(&next_mz->mem->css);
3596                                 else /* next_mz == NULL or other memcg */
3597                                         break;
3598                         } while (1);
3599                 }
3600                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3601                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3602                 /*
3603                  * One school of thought says that we should not add
3604                  * back the node to the tree if reclaim returns 0.
3605                  * But our reclaim could return 0, simply because due
3606                  * to priority we are exposing a smaller subset of
3607                  * memory to reclaim from. Consider this as a longer
3608                  * term TODO.
3609                  */
3610                 /* If excess == 0, no tree ops */
3611                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3612                 spin_unlock(&mctz->lock);
3613                 css_put(&mz->mem->css);
3614                 loop++;
3615                 /*
3616                  * Could not reclaim anything and there are no more
3617                  * mem cgroups to try or we seem to be looping without
3618                  * reclaiming anything.
3619                  */
3620                 if (!nr_reclaimed &&
3621                         (next_mz == NULL ||
3622                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3623                         break;
3624         } while (!nr_reclaimed);
3625         if (next_mz)
3626                 css_put(&next_mz->mem->css);
3627         return nr_reclaimed;
3628 }
3629
3630 /*
3631  * This routine traverse page_cgroup in given list and drop them all.
3632  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3633  */
3634 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3635                                 int node, int zid, enum lru_list lru)
3636 {
3637         struct zone *zone;
3638         struct mem_cgroup_per_zone *mz;
3639         struct page_cgroup *pc, *busy;
3640         unsigned long flags, loop;
3641         struct list_head *list;
3642         int ret = 0;
3643
3644         zone = &NODE_DATA(node)->node_zones[zid];
3645         mz = mem_cgroup_zoneinfo(mem, node, zid);
3646         list = &mz->lists[lru];
3647
3648         loop = MEM_CGROUP_ZSTAT(mz, lru);
3649         /* give some margin against EBUSY etc...*/
3650         loop += 256;
3651         busy = NULL;
3652         while (loop--) {
3653                 struct page *page;
3654
3655                 ret = 0;
3656                 spin_lock_irqsave(&zone->lru_lock, flags);
3657                 if (list_empty(list)) {
3658                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3659                         break;
3660                 }
3661                 pc = list_entry(list->prev, struct page_cgroup, lru);
3662                 if (busy == pc) {
3663                         list_move(&pc->lru, list);
3664                         busy = NULL;
3665                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3666                         continue;
3667                 }
3668                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3669
3670                 page = lookup_cgroup_page(pc);
3671
3672                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3673                 if (ret == -ENOMEM)
3674                         break;
3675
3676                 if (ret == -EBUSY || ret == -EINVAL) {
3677                         /* found lock contention or "pc" is obsolete. */
3678                         busy = pc;
3679                         cond_resched();
3680                 } else
3681                         busy = NULL;
3682         }
3683
3684         if (!ret && !list_empty(list))
3685                 return -EBUSY;
3686         return ret;
3687 }
3688
3689 /*
3690  * make mem_cgroup's charge to be 0 if there is no task.
3691  * This enables deleting this mem_cgroup.
3692  */
3693 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3694 {
3695         int ret;
3696         int node, zid, shrink;
3697         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3698         struct cgroup *cgrp = mem->css.cgroup;
3699
3700         css_get(&mem->css);
3701
3702         shrink = 0;
3703         /* should free all ? */
3704         if (free_all)
3705                 goto try_to_free;
3706 move_account:
3707         do {
3708                 ret = -EBUSY;
3709                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3710                         goto out;
3711                 ret = -EINTR;
3712                 if (signal_pending(current))
3713                         goto out;
3714                 /* This is for making all *used* pages to be on LRU. */
3715                 lru_add_drain_all();
3716                 drain_all_stock_sync(mem);
3717                 ret = 0;
3718                 mem_cgroup_start_move(mem);
3719                 for_each_node_state(node, N_HIGH_MEMORY) {
3720                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3721                                 enum lru_list l;
3722                                 for_each_lru(l) {
3723                                         ret = mem_cgroup_force_empty_list(mem,
3724                                                         node, zid, l);
3725                                         if (ret)
3726                                                 break;
3727                                 }
3728                         }
3729                         if (ret)
3730                                 break;
3731                 }
3732                 mem_cgroup_end_move(mem);
3733                 memcg_oom_recover(mem);
3734                 /* it seems parent cgroup doesn't have enough mem */
3735                 if (ret == -ENOMEM)
3736                         goto try_to_free;
3737                 cond_resched();
3738         /* "ret" should also be checked to ensure all lists are empty. */
3739         } while (mem->res.usage > 0 || ret);
3740 out:
3741         css_put(&mem->css);
3742         return ret;
3743
3744 try_to_free:
3745         /* returns EBUSY if there is a task or if we come here twice. */
3746         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3747                 ret = -EBUSY;
3748                 goto out;
3749         }
3750         /* we call try-to-free pages for make this cgroup empty */
3751         lru_add_drain_all();
3752         /* try to free all pages in this cgroup */
3753         shrink = 1;
3754         while (nr_retries && mem->res.usage > 0) {
3755                 int progress;
3756
3757                 if (signal_pending(current)) {
3758                         ret = -EINTR;
3759                         goto out;
3760                 }
3761                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3762                                                 false);
3763                 if (!progress) {
3764                         nr_retries--;
3765                         /* maybe some writeback is necessary */
3766                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3767                 }
3768
3769         }
3770         lru_add_drain();
3771         /* try move_account...there may be some *locked* pages. */
3772         goto move_account;
3773 }
3774
3775 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3776 {
3777         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3778 }
3779
3780
3781 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3782 {
3783         return mem_cgroup_from_cont(cont)->use_hierarchy;
3784 }
3785
3786 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3787                                         u64 val)
3788 {
3789         int retval = 0;
3790         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3791         struct cgroup *parent = cont->parent;
3792         struct mem_cgroup *parent_mem = NULL;
3793
3794         if (parent)
3795                 parent_mem = mem_cgroup_from_cont(parent);
3796
3797         cgroup_lock();
3798         /*
3799          * If parent's use_hierarchy is set, we can't make any modifications
3800          * in the child subtrees. If it is unset, then the change can
3801          * occur, provided the current cgroup has no children.
3802          *
3803          * For the root cgroup, parent_mem is NULL, we allow value to be
3804          * set if there are no children.
3805          */
3806         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3807                                 (val == 1 || val == 0)) {
3808                 if (list_empty(&cont->children))
3809                         mem->use_hierarchy = val;
3810                 else
3811                         retval = -EBUSY;
3812         } else
3813                 retval = -EINVAL;
3814         cgroup_unlock();
3815
3816         return retval;
3817 }
3818
3819
3820 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3821                                                enum mem_cgroup_stat_index idx)
3822 {
3823         struct mem_cgroup *iter;
3824         long val = 0;
3825
3826         /* Per-cpu values can be negative, use a signed accumulator */
3827         for_each_mem_cgroup_tree(iter, mem)
3828                 val += mem_cgroup_read_stat(iter, idx);
3829
3830         if (val < 0) /* race ? */
3831                 val = 0;
3832         return val;
3833 }
3834
3835 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3836 {
3837         u64 val;
3838
3839         if (!mem_cgroup_is_root(mem)) {
3840                 if (!swap)
3841                         return res_counter_read_u64(&mem->res, RES_USAGE);
3842                 else
3843                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3844         }
3845
3846         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3847         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3848
3849         if (swap)
3850                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3851
3852         return val << PAGE_SHIFT;
3853 }
3854
3855 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3856 {
3857         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3858         u64 val;
3859         int type, name;
3860
3861         type = MEMFILE_TYPE(cft->private);
3862         name = MEMFILE_ATTR(cft->private);
3863         switch (type) {
3864         case _MEM:
3865                 if (name == RES_USAGE)
3866                         val = mem_cgroup_usage(mem, false);
3867                 else
3868                         val = res_counter_read_u64(&mem->res, name);
3869                 break;
3870         case _MEMSWAP:
3871                 if (name == RES_USAGE)
3872                         val = mem_cgroup_usage(mem, true);
3873                 else
3874                         val = res_counter_read_u64(&mem->memsw, name);
3875                 break;
3876         default:
3877                 BUG();
3878                 break;
3879         }
3880         return val;
3881 }
3882 /*
3883  * The user of this function is...
3884  * RES_LIMIT.
3885  */
3886 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3887                             const char *buffer)
3888 {
3889         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3890         int type, name;
3891         unsigned long long val;
3892         int ret;
3893
3894         type = MEMFILE_TYPE(cft->private);
3895         name = MEMFILE_ATTR(cft->private);
3896         switch (name) {
3897         case RES_LIMIT:
3898                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3899                         ret = -EINVAL;
3900                         break;
3901                 }
3902                 /* This function does all necessary parse...reuse it */
3903                 ret = res_counter_memparse_write_strategy(buffer, &val);
3904                 if (ret)
3905                         break;
3906                 if (type == _MEM)
3907                         ret = mem_cgroup_resize_limit(memcg, val);
3908                 else
3909                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3910                 break;
3911         case RES_SOFT_LIMIT:
3912                 ret = res_counter_memparse_write_strategy(buffer, &val);
3913                 if (ret)
3914                         break;
3915                 /*
3916                  * For memsw, soft limits are hard to implement in terms
3917                  * of semantics, for now, we support soft limits for
3918                  * control without swap
3919                  */
3920                 if (type == _MEM)
3921                         ret = res_counter_set_soft_limit(&memcg->res, val);
3922                 else
3923                         ret = -EINVAL;
3924                 break;
3925         default:
3926                 ret = -EINVAL; /* should be BUG() ? */
3927                 break;
3928         }
3929         return ret;
3930 }
3931
3932 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3933                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3934 {
3935         struct cgroup *cgroup;
3936         unsigned long long min_limit, min_memsw_limit, tmp;
3937
3938         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3939         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3940         cgroup = memcg->css.cgroup;
3941         if (!memcg->use_hierarchy)
3942                 goto out;
3943
3944         while (cgroup->parent) {
3945                 cgroup = cgroup->parent;
3946                 memcg = mem_cgroup_from_cont(cgroup);
3947                 if (!memcg->use_hierarchy)
3948                         break;
3949                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3950                 min_limit = min(min_limit, tmp);
3951                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3952                 min_memsw_limit = min(min_memsw_limit, tmp);
3953         }
3954 out:
3955         *mem_limit = min_limit;
3956         *memsw_limit = min_memsw_limit;
3957         return;
3958 }
3959
3960 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3961 {
3962         struct mem_cgroup *mem;
3963         int type, name;
3964
3965         mem = mem_cgroup_from_cont(cont);
3966         type = MEMFILE_TYPE(event);
3967         name = MEMFILE_ATTR(event);
3968         switch (name) {
3969         case RES_MAX_USAGE:
3970                 if (type == _MEM)
3971                         res_counter_reset_max(&mem->res);
3972                 else
3973                         res_counter_reset_max(&mem->memsw);
3974                 break;
3975         case RES_FAILCNT:
3976                 if (type == _MEM)
3977                         res_counter_reset_failcnt(&mem->res);
3978                 else
3979                         res_counter_reset_failcnt(&mem->memsw);
3980                 break;
3981         }
3982
3983         return 0;
3984 }
3985
3986 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3987                                         struct cftype *cft)
3988 {
3989         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3990 }
3991
3992 #ifdef CONFIG_MMU
3993 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3994                                         struct cftype *cft, u64 val)
3995 {
3996         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3997
3998         if (val >= (1 << NR_MOVE_TYPE))
3999                 return -EINVAL;
4000         /*
4001          * We check this value several times in both in can_attach() and
4002          * attach(), so we need cgroup lock to prevent this value from being
4003          * inconsistent.
4004          */
4005         cgroup_lock();
4006         mem->move_charge_at_immigrate = val;
4007         cgroup_unlock();
4008
4009         return 0;
4010 }
4011 #else
4012 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4013                                         struct cftype *cft, u64 val)
4014 {
4015         return -ENOSYS;
4016 }
4017 #endif
4018
4019
4020 /* For read statistics */
4021 enum {
4022         MCS_CACHE,
4023         MCS_RSS,
4024         MCS_FILE_MAPPED,
4025         MCS_PGPGIN,
4026         MCS_PGPGOUT,
4027         MCS_SWAP,
4028         MCS_PGFAULT,
4029         MCS_PGMAJFAULT,
4030         MCS_INACTIVE_ANON,
4031         MCS_ACTIVE_ANON,
4032         MCS_INACTIVE_FILE,
4033         MCS_ACTIVE_FILE,
4034         MCS_UNEVICTABLE,
4035         NR_MCS_STAT,
4036 };
4037
4038 struct mcs_total_stat {
4039         s64 stat[NR_MCS_STAT];
4040 };
4041
4042 struct {
4043         char *local_name;
4044         char *total_name;
4045 } memcg_stat_strings[NR_MCS_STAT] = {
4046         {"cache", "total_cache"},
4047         {"rss", "total_rss"},
4048         {"mapped_file", "total_mapped_file"},
4049         {"pgpgin", "total_pgpgin"},
4050         {"pgpgout", "total_pgpgout"},
4051         {"swap", "total_swap"},
4052         {"pgfault", "total_pgfault"},
4053         {"pgmajfault", "total_pgmajfault"},
4054         {"inactive_anon", "total_inactive_anon"},
4055         {"active_anon", "total_active_anon"},
4056         {"inactive_file", "total_inactive_file"},
4057         {"active_file", "total_active_file"},
4058         {"unevictable", "total_unevictable"}
4059 };
4060
4061
4062 static void
4063 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4064 {
4065         s64 val;
4066
4067         /* per cpu stat */
4068         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4069         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4070         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4071         s->stat[MCS_RSS] += val * PAGE_SIZE;
4072         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4073         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4074         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4075         s->stat[MCS_PGPGIN] += val;
4076         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4077         s->stat[MCS_PGPGOUT] += val;
4078         if (do_swap_account) {
4079                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4080                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4081         }
4082         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4083         s->stat[MCS_PGFAULT] += val;
4084         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4085         s->stat[MCS_PGMAJFAULT] += val;
4086
4087         /* per zone stat */
4088         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4089         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4090         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4091         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4092         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4093         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4094         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4095         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4096         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4097         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4098 }
4099
4100 static void
4101 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4102 {
4103         struct mem_cgroup *iter;
4104
4105         for_each_mem_cgroup_tree(iter, mem)
4106                 mem_cgroup_get_local_stat(iter, s);
4107 }
4108
4109 #ifdef CONFIG_NUMA
4110 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4111 {
4112         int nid;
4113         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4114         unsigned long node_nr;
4115         struct cgroup *cont = m->private;
4116         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4117
4118         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4119         seq_printf(m, "total=%lu", total_nr);
4120         for_each_node_state(nid, N_HIGH_MEMORY) {
4121                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4122                 seq_printf(m, " N%d=%lu", nid, node_nr);
4123         }
4124         seq_putc(m, '\n');
4125
4126         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4127         seq_printf(m, "file=%lu", file_nr);
4128         for_each_node_state(nid, N_HIGH_MEMORY) {
4129                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4130                                 LRU_ALL_FILE);
4131                 seq_printf(m, " N%d=%lu", nid, node_nr);
4132         }
4133         seq_putc(m, '\n');
4134
4135         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4136         seq_printf(m, "anon=%lu", anon_nr);
4137         for_each_node_state(nid, N_HIGH_MEMORY) {
4138                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4139                                 LRU_ALL_ANON);
4140                 seq_printf(m, " N%d=%lu", nid, node_nr);
4141         }
4142         seq_putc(m, '\n');
4143
4144         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4145         seq_printf(m, "unevictable=%lu", unevictable_nr);
4146         for_each_node_state(nid, N_HIGH_MEMORY) {
4147                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4148                                 BIT(LRU_UNEVICTABLE));
4149                 seq_printf(m, " N%d=%lu", nid, node_nr);
4150         }
4151         seq_putc(m, '\n');
4152         return 0;
4153 }
4154 #endif /* CONFIG_NUMA */
4155
4156 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4157                                  struct cgroup_map_cb *cb)
4158 {
4159         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4160         struct mcs_total_stat mystat;
4161         int i;
4162
4163         memset(&mystat, 0, sizeof(mystat));
4164         mem_cgroup_get_local_stat(mem_cont, &mystat);
4165
4166
4167         for (i = 0; i < NR_MCS_STAT; i++) {
4168                 if (i == MCS_SWAP && !do_swap_account)
4169                         continue;
4170                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4171         }
4172
4173         /* Hierarchical information */
4174         {
4175                 unsigned long long limit, memsw_limit;
4176                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4177                 cb->fill(cb, "hierarchical_memory_limit", limit);
4178                 if (do_swap_account)
4179                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4180         }
4181
4182         memset(&mystat, 0, sizeof(mystat));
4183         mem_cgroup_get_total_stat(mem_cont, &mystat);
4184         for (i = 0; i < NR_MCS_STAT; i++) {
4185                 if (i == MCS_SWAP && !do_swap_account)
4186                         continue;
4187                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4188         }
4189
4190 #ifdef CONFIG_DEBUG_VM
4191         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4192
4193         {
4194                 int nid, zid;
4195                 struct mem_cgroup_per_zone *mz;
4196                 unsigned long recent_rotated[2] = {0, 0};
4197                 unsigned long recent_scanned[2] = {0, 0};
4198
4199                 for_each_online_node(nid)
4200                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4201                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4202
4203                                 recent_rotated[0] +=
4204                                         mz->reclaim_stat.recent_rotated[0];
4205                                 recent_rotated[1] +=
4206                                         mz->reclaim_stat.recent_rotated[1];
4207                                 recent_scanned[0] +=
4208                                         mz->reclaim_stat.recent_scanned[0];
4209                                 recent_scanned[1] +=
4210                                         mz->reclaim_stat.recent_scanned[1];
4211                         }
4212                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4213                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4214                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4215                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4216         }
4217 #endif
4218
4219         return 0;
4220 }
4221
4222 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4223 {
4224         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4225
4226         return mem_cgroup_swappiness(memcg);
4227 }
4228
4229 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4230                                        u64 val)
4231 {
4232         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4233         struct mem_cgroup *parent;
4234
4235         if (val > 100)
4236                 return -EINVAL;
4237
4238         if (cgrp->parent == NULL)
4239                 return -EINVAL;
4240
4241         parent = mem_cgroup_from_cont(cgrp->parent);
4242
4243         cgroup_lock();
4244
4245         /* If under hierarchy, only empty-root can set this value */
4246         if ((parent->use_hierarchy) ||
4247             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4248                 cgroup_unlock();
4249                 return -EINVAL;
4250         }
4251
4252         memcg->swappiness = val;
4253
4254         cgroup_unlock();
4255
4256         return 0;
4257 }
4258
4259 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4260 {
4261         struct mem_cgroup_threshold_ary *t;
4262         u64 usage;
4263         int i;
4264
4265         rcu_read_lock();
4266         if (!swap)
4267                 t = rcu_dereference(memcg->thresholds.primary);
4268         else
4269                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4270
4271         if (!t)
4272                 goto unlock;
4273
4274         usage = mem_cgroup_usage(memcg, swap);
4275
4276         /*
4277          * current_threshold points to threshold just below usage.
4278          * If it's not true, a threshold was crossed after last
4279          * call of __mem_cgroup_threshold().
4280          */
4281         i = t->current_threshold;
4282
4283         /*
4284          * Iterate backward over array of thresholds starting from
4285          * current_threshold and check if a threshold is crossed.
4286          * If none of thresholds below usage is crossed, we read
4287          * only one element of the array here.
4288          */
4289         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4290                 eventfd_signal(t->entries[i].eventfd, 1);
4291
4292         /* i = current_threshold + 1 */
4293         i++;
4294
4295         /*
4296          * Iterate forward over array of thresholds starting from
4297          * current_threshold+1 and check if a threshold is crossed.
4298          * If none of thresholds above usage is crossed, we read
4299          * only one element of the array here.
4300          */
4301         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4302                 eventfd_signal(t->entries[i].eventfd, 1);
4303
4304         /* Update current_threshold */
4305         t->current_threshold = i - 1;
4306 unlock:
4307         rcu_read_unlock();
4308 }
4309
4310 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4311 {
4312         while (memcg) {
4313                 __mem_cgroup_threshold(memcg, false);
4314                 if (do_swap_account)
4315                         __mem_cgroup_threshold(memcg, true);
4316
4317                 memcg = parent_mem_cgroup(memcg);
4318         }
4319 }
4320
4321 static int compare_thresholds(const void *a, const void *b)
4322 {
4323         const struct mem_cgroup_threshold *_a = a;
4324         const struct mem_cgroup_threshold *_b = b;
4325
4326         return _a->threshold - _b->threshold;
4327 }
4328
4329 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4330 {
4331         struct mem_cgroup_eventfd_list *ev;
4332
4333         list_for_each_entry(ev, &mem->oom_notify, list)
4334                 eventfd_signal(ev->eventfd, 1);
4335         return 0;
4336 }
4337
4338 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4339 {
4340         struct mem_cgroup *iter;
4341
4342         for_each_mem_cgroup_tree(iter, mem)
4343                 mem_cgroup_oom_notify_cb(iter);
4344 }
4345
4346 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4347         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4348 {
4349         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4350         struct mem_cgroup_thresholds *thresholds;
4351         struct mem_cgroup_threshold_ary *new;
4352         int type = MEMFILE_TYPE(cft->private);
4353         u64 threshold, usage;
4354         int i, size, ret;
4355
4356         ret = res_counter_memparse_write_strategy(args, &threshold);
4357         if (ret)
4358                 return ret;
4359
4360         mutex_lock(&memcg->thresholds_lock);
4361
4362         if (type == _MEM)
4363                 thresholds = &memcg->thresholds;
4364         else if (type == _MEMSWAP)
4365                 thresholds = &memcg->memsw_thresholds;
4366         else
4367                 BUG();
4368
4369         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4370
4371         /* Check if a threshold crossed before adding a new one */
4372         if (thresholds->primary)
4373                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4374
4375         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4376
4377         /* Allocate memory for new array of thresholds */
4378         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4379                         GFP_KERNEL);
4380         if (!new) {
4381                 ret = -ENOMEM;
4382                 goto unlock;
4383         }
4384         new->size = size;
4385
4386         /* Copy thresholds (if any) to new array */
4387         if (thresholds->primary) {
4388                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4389                                 sizeof(struct mem_cgroup_threshold));
4390         }
4391
4392         /* Add new threshold */
4393         new->entries[size - 1].eventfd = eventfd;
4394         new->entries[size - 1].threshold = threshold;
4395
4396         /* Sort thresholds. Registering of new threshold isn't time-critical */
4397         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4398                         compare_thresholds, NULL);
4399
4400         /* Find current threshold */
4401         new->current_threshold = -1;
4402         for (i = 0; i < size; i++) {
4403                 if (new->entries[i].threshold < usage) {
4404                         /*
4405                          * new->current_threshold will not be used until
4406                          * rcu_assign_pointer(), so it's safe to increment
4407                          * it here.
4408                          */
4409                         ++new->current_threshold;
4410                 }
4411         }
4412
4413         /* Free old spare buffer and save old primary buffer as spare */
4414         kfree(thresholds->spare);
4415         thresholds->spare = thresholds->primary;
4416
4417         rcu_assign_pointer(thresholds->primary, new);
4418
4419         /* To be sure that nobody uses thresholds */
4420         synchronize_rcu();
4421
4422 unlock:
4423         mutex_unlock(&memcg->thresholds_lock);
4424
4425         return ret;
4426 }
4427
4428 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4429         struct cftype *cft, struct eventfd_ctx *eventfd)
4430 {
4431         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4432         struct mem_cgroup_thresholds *thresholds;
4433         struct mem_cgroup_threshold_ary *new;
4434         int type = MEMFILE_TYPE(cft->private);
4435         u64 usage;
4436         int i, j, size;
4437
4438         mutex_lock(&memcg->thresholds_lock);
4439         if (type == _MEM)
4440                 thresholds = &memcg->thresholds;
4441         else if (type == _MEMSWAP)
4442                 thresholds = &memcg->memsw_thresholds;
4443         else
4444                 BUG();
4445
4446         /*
4447          * Something went wrong if we trying to unregister a threshold
4448          * if we don't have thresholds
4449          */
4450         BUG_ON(!thresholds);
4451
4452         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4453
4454         /* Check if a threshold crossed before removing */
4455         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4456
4457         /* Calculate new number of threshold */
4458         size = 0;
4459         for (i = 0; i < thresholds->primary->size; i++) {
4460                 if (thresholds->primary->entries[i].eventfd != eventfd)
4461                         size++;
4462         }
4463
4464         new = thresholds->spare;
4465
4466         /* Set thresholds array to NULL if we don't have thresholds */
4467         if (!size) {
4468                 kfree(new);
4469                 new = NULL;
4470                 goto swap_buffers;
4471         }
4472
4473         new->size = size;
4474
4475         /* Copy thresholds and find current threshold */
4476         new->current_threshold = -1;
4477         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4478                 if (thresholds->primary->entries[i].eventfd == eventfd)
4479                         continue;
4480
4481                 new->entries[j] = thresholds->primary->entries[i];
4482                 if (new->entries[j].threshold < usage) {
4483                         /*
4484                          * new->current_threshold will not be used
4485                          * until rcu_assign_pointer(), so it's safe to increment
4486                          * it here.
4487                          */
4488                         ++new->current_threshold;
4489                 }
4490                 j++;
4491         }
4492
4493 swap_buffers:
4494         /* Swap primary and spare array */
4495         thresholds->spare = thresholds->primary;
4496         rcu_assign_pointer(thresholds->primary, new);
4497
4498         /* To be sure that nobody uses thresholds */
4499         synchronize_rcu();
4500
4501         mutex_unlock(&memcg->thresholds_lock);
4502 }
4503
4504 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4505         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4506 {
4507         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4508         struct mem_cgroup_eventfd_list *event;
4509         int type = MEMFILE_TYPE(cft->private);
4510
4511         BUG_ON(type != _OOM_TYPE);
4512         event = kmalloc(sizeof(*event), GFP_KERNEL);
4513         if (!event)
4514                 return -ENOMEM;
4515
4516         spin_lock(&memcg_oom_lock);
4517
4518         event->eventfd = eventfd;
4519         list_add(&event->list, &memcg->oom_notify);
4520
4521         /* already in OOM ? */
4522         if (atomic_read(&memcg->under_oom))
4523                 eventfd_signal(eventfd, 1);
4524         spin_unlock(&memcg_oom_lock);
4525
4526         return 0;
4527 }
4528
4529 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4530         struct cftype *cft, struct eventfd_ctx *eventfd)
4531 {
4532         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4533         struct mem_cgroup_eventfd_list *ev, *tmp;
4534         int type = MEMFILE_TYPE(cft->private);
4535
4536         BUG_ON(type != _OOM_TYPE);
4537
4538         spin_lock(&memcg_oom_lock);
4539
4540         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4541                 if (ev->eventfd == eventfd) {
4542                         list_del(&ev->list);
4543                         kfree(ev);
4544                 }
4545         }
4546
4547         spin_unlock(&memcg_oom_lock);
4548 }
4549
4550 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4551         struct cftype *cft,  struct cgroup_map_cb *cb)
4552 {
4553         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4554
4555         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4556
4557         if (atomic_read(&mem->under_oom))
4558                 cb->fill(cb, "under_oom", 1);
4559         else
4560                 cb->fill(cb, "under_oom", 0);
4561         return 0;
4562 }
4563
4564 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4565         struct cftype *cft, u64 val)
4566 {
4567         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4568         struct mem_cgroup *parent;
4569
4570         /* cannot set to root cgroup and only 0 and 1 are allowed */
4571         if (!cgrp->parent || !((val == 0) || (val == 1)))
4572                 return -EINVAL;
4573
4574         parent = mem_cgroup_from_cont(cgrp->parent);
4575
4576         cgroup_lock();
4577         /* oom-kill-disable is a flag for subhierarchy. */
4578         if ((parent->use_hierarchy) ||
4579             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4580                 cgroup_unlock();
4581                 return -EINVAL;
4582         }
4583         mem->oom_kill_disable = val;
4584         if (!val)
4585                 memcg_oom_recover(mem);
4586         cgroup_unlock();
4587         return 0;
4588 }
4589
4590 #ifdef CONFIG_NUMA
4591 static const struct file_operations mem_control_numa_stat_file_operations = {
4592         .read = seq_read,
4593         .llseek = seq_lseek,
4594         .release = single_release,
4595 };
4596
4597 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4598 {
4599         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4600
4601         file->f_op = &mem_control_numa_stat_file_operations;
4602         return single_open(file, mem_control_numa_stat_show, cont);
4603 }
4604 #endif /* CONFIG_NUMA */
4605
4606 static struct cftype mem_cgroup_files[] = {
4607         {
4608                 .name = "usage_in_bytes",
4609                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4610                 .read_u64 = mem_cgroup_read,
4611                 .register_event = mem_cgroup_usage_register_event,
4612                 .unregister_event = mem_cgroup_usage_unregister_event,
4613         },
4614         {
4615                 .name = "max_usage_in_bytes",
4616                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4617                 .trigger = mem_cgroup_reset,
4618                 .read_u64 = mem_cgroup_read,
4619         },
4620         {
4621                 .name = "limit_in_bytes",
4622                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4623                 .write_string = mem_cgroup_write,
4624                 .read_u64 = mem_cgroup_read,
4625         },
4626         {
4627                 .name = "soft_limit_in_bytes",
4628                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4629                 .write_string = mem_cgroup_write,
4630                 .read_u64 = mem_cgroup_read,
4631         },
4632         {
4633                 .name = "failcnt",
4634                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4635                 .trigger = mem_cgroup_reset,
4636                 .read_u64 = mem_cgroup_read,
4637         },
4638         {
4639                 .name = "stat",
4640                 .read_map = mem_control_stat_show,
4641         },
4642         {
4643                 .name = "force_empty",
4644                 .trigger = mem_cgroup_force_empty_write,
4645         },
4646         {
4647                 .name = "use_hierarchy",
4648                 .write_u64 = mem_cgroup_hierarchy_write,
4649                 .read_u64 = mem_cgroup_hierarchy_read,
4650         },
4651         {
4652                 .name = "swappiness",
4653                 .read_u64 = mem_cgroup_swappiness_read,
4654                 .write_u64 = mem_cgroup_swappiness_write,
4655         },
4656         {
4657                 .name = "move_charge_at_immigrate",
4658                 .read_u64 = mem_cgroup_move_charge_read,
4659                 .write_u64 = mem_cgroup_move_charge_write,
4660         },
4661         {
4662                 .name = "oom_control",
4663                 .read_map = mem_cgroup_oom_control_read,
4664                 .write_u64 = mem_cgroup_oom_control_write,
4665                 .register_event = mem_cgroup_oom_register_event,
4666                 .unregister_event = mem_cgroup_oom_unregister_event,
4667                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4668         },
4669 #ifdef CONFIG_NUMA
4670         {
4671                 .name = "numa_stat",
4672                 .open = mem_control_numa_stat_open,
4673                 .mode = S_IRUGO,
4674         },
4675 #endif
4676 };
4677
4678 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4679 static struct cftype memsw_cgroup_files[] = {
4680         {
4681                 .name = "memsw.usage_in_bytes",
4682                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4683                 .read_u64 = mem_cgroup_read,
4684                 .register_event = mem_cgroup_usage_register_event,
4685                 .unregister_event = mem_cgroup_usage_unregister_event,
4686         },
4687         {
4688                 .name = "memsw.max_usage_in_bytes",
4689                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4690                 .trigger = mem_cgroup_reset,
4691                 .read_u64 = mem_cgroup_read,
4692         },
4693         {
4694                 .name = "memsw.limit_in_bytes",
4695                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4696                 .write_string = mem_cgroup_write,
4697                 .read_u64 = mem_cgroup_read,
4698         },
4699         {
4700                 .name = "memsw.failcnt",
4701                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4702                 .trigger = mem_cgroup_reset,
4703                 .read_u64 = mem_cgroup_read,
4704         },
4705 };
4706
4707 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4708 {
4709         if (!do_swap_account)
4710                 return 0;
4711         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4712                                 ARRAY_SIZE(memsw_cgroup_files));
4713 };
4714 #else
4715 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4716 {
4717         return 0;
4718 }
4719 #endif
4720
4721 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4722 {
4723         struct mem_cgroup_per_node *pn;
4724         struct mem_cgroup_per_zone *mz;
4725         enum lru_list l;
4726         int zone, tmp = node;
4727         /*
4728          * This routine is called against possible nodes.
4729          * But it's BUG to call kmalloc() against offline node.
4730          *
4731          * TODO: this routine can waste much memory for nodes which will
4732          *       never be onlined. It's better to use memory hotplug callback
4733          *       function.
4734          */
4735         if (!node_state(node, N_NORMAL_MEMORY))
4736                 tmp = -1;
4737         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4738         if (!pn)
4739                 return 1;
4740
4741         mem->info.nodeinfo[node] = pn;
4742         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4743                 mz = &pn->zoneinfo[zone];
4744                 for_each_lru(l)
4745                         INIT_LIST_HEAD(&mz->lists[l]);
4746                 mz->usage_in_excess = 0;
4747                 mz->on_tree = false;
4748                 mz->mem = mem;
4749         }
4750         return 0;
4751 }
4752
4753 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4754 {
4755         kfree(mem->info.nodeinfo[node]);
4756 }
4757
4758 static struct mem_cgroup *mem_cgroup_alloc(void)
4759 {
4760         struct mem_cgroup *mem;
4761         int size = sizeof(struct mem_cgroup);
4762
4763         /* Can be very big if MAX_NUMNODES is very big */
4764         if (size < PAGE_SIZE)
4765                 mem = kzalloc(size, GFP_KERNEL);
4766         else
4767                 mem = vzalloc(size);
4768
4769         if (!mem)
4770                 return NULL;
4771
4772         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4773         if (!mem->stat)
4774                 goto out_free;
4775         spin_lock_init(&mem->pcp_counter_lock);
4776         return mem;
4777
4778 out_free:
4779         if (size < PAGE_SIZE)
4780                 kfree(mem);
4781         else
4782                 vfree(mem);
4783         return NULL;
4784 }
4785
4786 /*
4787  * At destroying mem_cgroup, references from swap_cgroup can remain.
4788  * (scanning all at force_empty is too costly...)
4789  *
4790  * Instead of clearing all references at force_empty, we remember
4791  * the number of reference from swap_cgroup and free mem_cgroup when
4792  * it goes down to 0.
4793  *
4794  * Removal of cgroup itself succeeds regardless of refs from swap.
4795  */
4796
4797 static void __mem_cgroup_free(struct mem_cgroup *mem)
4798 {
4799         int node;
4800
4801         mem_cgroup_remove_from_trees(mem);
4802         free_css_id(&mem_cgroup_subsys, &mem->css);
4803
4804         for_each_node_state(node, N_POSSIBLE)
4805                 free_mem_cgroup_per_zone_info(mem, node);
4806
4807         free_percpu(mem->stat);
4808         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4809                 kfree(mem);
4810         else
4811                 vfree(mem);
4812 }
4813
4814 static void mem_cgroup_get(struct mem_cgroup *mem)
4815 {
4816         atomic_inc(&mem->refcnt);
4817 }
4818
4819 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4820 {
4821         if (atomic_sub_and_test(count, &mem->refcnt)) {
4822                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4823                 __mem_cgroup_free(mem);
4824                 if (parent)
4825                         mem_cgroup_put(parent);
4826         }
4827 }
4828
4829 static void mem_cgroup_put(struct mem_cgroup *mem)
4830 {
4831         __mem_cgroup_put(mem, 1);
4832 }
4833
4834 /*
4835  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4836  */
4837 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4838 {
4839         if (!mem->res.parent)
4840                 return NULL;
4841         return mem_cgroup_from_res_counter(mem->res.parent, res);
4842 }
4843
4844 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4845 static void __init enable_swap_cgroup(void)
4846 {
4847         if (!mem_cgroup_disabled() && really_do_swap_account)
4848                 do_swap_account = 1;
4849 }
4850 #else
4851 static void __init enable_swap_cgroup(void)
4852 {
4853 }
4854 #endif
4855
4856 static int mem_cgroup_soft_limit_tree_init(void)
4857 {
4858         struct mem_cgroup_tree_per_node *rtpn;
4859         struct mem_cgroup_tree_per_zone *rtpz;
4860         int tmp, node, zone;
4861
4862         for_each_node_state(node, N_POSSIBLE) {
4863                 tmp = node;
4864                 if (!node_state(node, N_NORMAL_MEMORY))
4865                         tmp = -1;
4866                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4867                 if (!rtpn)
4868                         return 1;
4869
4870                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4871
4872                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4873                         rtpz = &rtpn->rb_tree_per_zone[zone];
4874                         rtpz->rb_root = RB_ROOT;
4875                         spin_lock_init(&rtpz->lock);
4876                 }
4877         }
4878         return 0;
4879 }
4880
4881 static struct cgroup_subsys_state * __ref
4882 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4883 {
4884         struct mem_cgroup *mem, *parent;
4885         long error = -ENOMEM;
4886         int node;
4887
4888         mem = mem_cgroup_alloc();
4889         if (!mem)
4890                 return ERR_PTR(error);
4891
4892         for_each_node_state(node, N_POSSIBLE)
4893                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4894                         goto free_out;
4895
4896         /* root ? */
4897         if (cont->parent == NULL) {
4898                 int cpu;
4899                 enable_swap_cgroup();
4900                 parent = NULL;
4901                 if (mem_cgroup_soft_limit_tree_init())
4902                         goto free_out;
4903                 root_mem_cgroup = mem;
4904                 for_each_possible_cpu(cpu) {
4905                         struct memcg_stock_pcp *stock =
4906                                                 &per_cpu(memcg_stock, cpu);
4907                         INIT_WORK(&stock->work, drain_local_stock);
4908                 }
4909                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4910         } else {
4911                 parent = mem_cgroup_from_cont(cont->parent);
4912                 mem->use_hierarchy = parent->use_hierarchy;
4913                 mem->oom_kill_disable = parent->oom_kill_disable;
4914         }
4915
4916         if (parent && parent->use_hierarchy) {
4917                 res_counter_init(&mem->res, &parent->res);
4918                 res_counter_init(&mem->memsw, &parent->memsw);
4919                 /*
4920                  * We increment refcnt of the parent to ensure that we can
4921                  * safely access it on res_counter_charge/uncharge.
4922                  * This refcnt will be decremented when freeing this
4923                  * mem_cgroup(see mem_cgroup_put).
4924                  */
4925                 mem_cgroup_get(parent);
4926         } else {
4927                 res_counter_init(&mem->res, NULL);
4928                 res_counter_init(&mem->memsw, NULL);
4929         }
4930         mem->last_scanned_child = 0;
4931         mem->last_scanned_node = MAX_NUMNODES;
4932         INIT_LIST_HEAD(&mem->oom_notify);
4933
4934         if (parent)
4935                 mem->swappiness = mem_cgroup_swappiness(parent);
4936         atomic_set(&mem->refcnt, 1);
4937         mem->move_charge_at_immigrate = 0;
4938         mutex_init(&mem->thresholds_lock);
4939         return &mem->css;
4940 free_out:
4941         __mem_cgroup_free(mem);
4942         return ERR_PTR(error);
4943 }
4944
4945 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4946                                         struct cgroup *cont)
4947 {
4948         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4949
4950         return mem_cgroup_force_empty(mem, false);
4951 }
4952
4953 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4954                                 struct cgroup *cont)
4955 {
4956         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4957
4958         mem_cgroup_put(mem);
4959 }
4960
4961 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4962                                 struct cgroup *cont)
4963 {
4964         int ret;
4965
4966         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4967                                 ARRAY_SIZE(mem_cgroup_files));
4968
4969         if (!ret)
4970                 ret = register_memsw_files(cont, ss);
4971         return ret;
4972 }
4973
4974 #ifdef CONFIG_MMU
4975 /* Handlers for move charge at task migration. */
4976 #define PRECHARGE_COUNT_AT_ONCE 256
4977 static int mem_cgroup_do_precharge(unsigned long count)
4978 {
4979         int ret = 0;
4980         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4981         struct mem_cgroup *mem = mc.to;
4982
4983         if (mem_cgroup_is_root(mem)) {
4984                 mc.precharge += count;
4985                 /* we don't need css_get for root */
4986                 return ret;
4987         }
4988         /* try to charge at once */
4989         if (count > 1) {
4990                 struct res_counter *dummy;
4991                 /*
4992                  * "mem" cannot be under rmdir() because we've already checked
4993                  * by cgroup_lock_live_cgroup() that it is not removed and we
4994                  * are still under the same cgroup_mutex. So we can postpone
4995                  * css_get().
4996                  */
4997                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4998                         goto one_by_one;
4999                 if (do_swap_account && res_counter_charge(&mem->memsw,
5000                                                 PAGE_SIZE * count, &dummy)) {
5001                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5002                         goto one_by_one;
5003                 }
5004                 mc.precharge += count;
5005                 return ret;
5006         }
5007 one_by_one:
5008         /* fall back to one by one charge */
5009         while (count--) {
5010                 if (signal_pending(current)) {
5011                         ret = -EINTR;
5012                         break;
5013                 }
5014                 if (!batch_count--) {
5015                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5016                         cond_resched();
5017                 }
5018                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5019                 if (ret || !mem)
5020                         /* mem_cgroup_clear_mc() will do uncharge later */
5021                         return -ENOMEM;
5022                 mc.precharge++;
5023         }
5024         return ret;
5025 }
5026
5027 /**
5028  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5029  * @vma: the vma the pte to be checked belongs
5030  * @addr: the address corresponding to the pte to be checked
5031  * @ptent: the pte to be checked
5032  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5033  *
5034  * Returns
5035  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5036  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5037  *     move charge. if @target is not NULL, the page is stored in target->page
5038  *     with extra refcnt got(Callers should handle it).
5039  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5040  *     target for charge migration. if @target is not NULL, the entry is stored
5041  *     in target->ent.
5042  *
5043  * Called with pte lock held.
5044  */
5045 union mc_target {
5046         struct page     *page;
5047         swp_entry_t     ent;
5048 };
5049
5050 enum mc_target_type {
5051         MC_TARGET_NONE, /* not used */
5052         MC_TARGET_PAGE,
5053         MC_TARGET_SWAP,
5054 };
5055
5056 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5057                                                 unsigned long addr, pte_t ptent)
5058 {
5059         struct page *page = vm_normal_page(vma, addr, ptent);
5060
5061         if (!page || !page_mapped(page))
5062                 return NULL;
5063         if (PageAnon(page)) {
5064                 /* we don't move shared anon */
5065                 if (!move_anon() || page_mapcount(page) > 2)
5066                         return NULL;
5067         } else if (!move_file())
5068                 /* we ignore mapcount for file pages */
5069                 return NULL;
5070         if (!get_page_unless_zero(page))
5071                 return NULL;
5072
5073         return page;
5074 }
5075
5076 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5077                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5078 {
5079         int usage_count;
5080         struct page *page = NULL;
5081         swp_entry_t ent = pte_to_swp_entry(ptent);
5082
5083         if (!move_anon() || non_swap_entry(ent))
5084                 return NULL;
5085         usage_count = mem_cgroup_count_swap_user(ent, &page);
5086         if (usage_count > 1) { /* we don't move shared anon */
5087                 if (page)
5088                         put_page(page);
5089                 return NULL;
5090         }
5091         if (do_swap_account)
5092                 entry->val = ent.val;
5093
5094         return page;
5095 }
5096
5097 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5098                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5099 {
5100         struct page *page = NULL;
5101         struct inode *inode;
5102         struct address_space *mapping;
5103         pgoff_t pgoff;
5104
5105         if (!vma->vm_file) /* anonymous vma */
5106                 return NULL;
5107         if (!move_file())
5108                 return NULL;
5109
5110         inode = vma->vm_file->f_path.dentry->d_inode;
5111         mapping = vma->vm_file->f_mapping;
5112         if (pte_none(ptent))
5113                 pgoff = linear_page_index(vma, addr);
5114         else /* pte_file(ptent) is true */
5115                 pgoff = pte_to_pgoff(ptent);
5116
5117         /* page is moved even if it's not RSS of this task(page-faulted). */
5118         page = find_get_page(mapping, pgoff);
5119
5120 #ifdef CONFIG_SWAP
5121         /* shmem/tmpfs may report page out on swap: account for that too. */
5122         if (radix_tree_exceptional_entry(page)) {
5123                 swp_entry_t swap = radix_to_swp_entry(page);
5124                 if (do_swap_account)
5125                         *entry = swap;
5126                 page = find_get_page(&swapper_space, swap.val);
5127         }
5128 #endif
5129         return page;
5130 }
5131
5132 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5133                 unsigned long addr, pte_t ptent, union mc_target *target)
5134 {
5135         struct page *page = NULL;
5136         struct page_cgroup *pc;
5137         int ret = 0;
5138         swp_entry_t ent = { .val = 0 };
5139
5140         if (pte_present(ptent))
5141                 page = mc_handle_present_pte(vma, addr, ptent);
5142         else if (is_swap_pte(ptent))
5143                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5144         else if (pte_none(ptent) || pte_file(ptent))
5145                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5146
5147         if (!page && !ent.val)
5148                 return 0;
5149         if (page) {
5150                 pc = lookup_page_cgroup(page);
5151                 /*
5152                  * Do only loose check w/o page_cgroup lock.
5153                  * mem_cgroup_move_account() checks the pc is valid or not under
5154                  * the lock.
5155                  */
5156                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5157                         ret = MC_TARGET_PAGE;
5158                         if (target)
5159                                 target->page = page;
5160                 }
5161                 if (!ret || !target)
5162                         put_page(page);
5163         }
5164         /* There is a swap entry and a page doesn't exist or isn't charged */
5165         if (ent.val && !ret &&
5166                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5167                 ret = MC_TARGET_SWAP;
5168                 if (target)
5169                         target->ent = ent;
5170         }
5171         return ret;
5172 }
5173
5174 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5175                                         unsigned long addr, unsigned long end,
5176                                         struct mm_walk *walk)
5177 {
5178         struct vm_area_struct *vma = walk->private;
5179         pte_t *pte;
5180         spinlock_t *ptl;
5181
5182         split_huge_page_pmd(walk->mm, pmd);
5183
5184         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5185         for (; addr != end; pte++, addr += PAGE_SIZE)
5186                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5187                         mc.precharge++; /* increment precharge temporarily */
5188         pte_unmap_unlock(pte - 1, ptl);
5189         cond_resched();
5190
5191         return 0;
5192 }
5193
5194 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5195 {
5196         unsigned long precharge;
5197         struct vm_area_struct *vma;
5198
5199         down_read(&mm->mmap_sem);
5200         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5201                 struct mm_walk mem_cgroup_count_precharge_walk = {
5202                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5203                         .mm = mm,
5204                         .private = vma,
5205                 };
5206                 if (is_vm_hugetlb_page(vma))
5207                         continue;
5208                 walk_page_range(vma->vm_start, vma->vm_end,
5209                                         &mem_cgroup_count_precharge_walk);
5210         }
5211         up_read(&mm->mmap_sem);
5212
5213         precharge = mc.precharge;
5214         mc.precharge = 0;
5215
5216         return precharge;
5217 }
5218
5219 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5220 {
5221         unsigned long precharge = mem_cgroup_count_precharge(mm);
5222
5223         VM_BUG_ON(mc.moving_task);
5224         mc.moving_task = current;
5225         return mem_cgroup_do_precharge(precharge);
5226 }
5227
5228 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5229 static void __mem_cgroup_clear_mc(void)
5230 {
5231         struct mem_cgroup *from = mc.from;
5232         struct mem_cgroup *to = mc.to;
5233
5234         /* we must uncharge all the leftover precharges from mc.to */
5235         if (mc.precharge) {
5236                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5237                 mc.precharge = 0;
5238         }
5239         /*
5240          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5241          * we must uncharge here.
5242          */
5243         if (mc.moved_charge) {
5244                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5245                 mc.moved_charge = 0;
5246         }
5247         /* we must fixup refcnts and charges */
5248         if (mc.moved_swap) {
5249                 /* uncharge swap account from the old cgroup */
5250                 if (!mem_cgroup_is_root(mc.from))
5251                         res_counter_uncharge(&mc.from->memsw,
5252                                                 PAGE_SIZE * mc.moved_swap);
5253                 __mem_cgroup_put(mc.from, mc.moved_swap);
5254
5255                 if (!mem_cgroup_is_root(mc.to)) {
5256                         /*
5257                          * we charged both to->res and to->memsw, so we should
5258                          * uncharge to->res.
5259                          */
5260                         res_counter_uncharge(&mc.to->res,
5261                                                 PAGE_SIZE * mc.moved_swap);
5262                 }
5263                 /* we've already done mem_cgroup_get(mc.to) */
5264                 mc.moved_swap = 0;
5265         }
5266         memcg_oom_recover(from);
5267         memcg_oom_recover(to);
5268         wake_up_all(&mc.waitq);
5269 }
5270
5271 static void mem_cgroup_clear_mc(void)
5272 {
5273         struct mem_cgroup *from = mc.from;
5274
5275         /*
5276          * we must clear moving_task before waking up waiters at the end of
5277          * task migration.
5278          */
5279         mc.moving_task = NULL;
5280         __mem_cgroup_clear_mc();
5281         spin_lock(&mc.lock);
5282         mc.from = NULL;
5283         mc.to = NULL;
5284         spin_unlock(&mc.lock);
5285         mem_cgroup_end_move(from);
5286 }
5287
5288 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5289                                 struct cgroup *cgroup,
5290                                 struct task_struct *p)
5291 {
5292         int ret = 0;
5293         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5294
5295         if (mem->move_charge_at_immigrate) {
5296                 struct mm_struct *mm;
5297                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5298
5299                 VM_BUG_ON(from == mem);
5300
5301                 mm = get_task_mm(p);
5302                 if (!mm)
5303                         return 0;
5304                 /* We move charges only when we move a owner of the mm */
5305                 if (mm->owner == p) {
5306                         VM_BUG_ON(mc.from);
5307                         VM_BUG_ON(mc.to);
5308                         VM_BUG_ON(mc.precharge);
5309                         VM_BUG_ON(mc.moved_charge);
5310                         VM_BUG_ON(mc.moved_swap);
5311                         mem_cgroup_start_move(from);
5312                         spin_lock(&mc.lock);
5313                         mc.from = from;
5314                         mc.to = mem;
5315                         spin_unlock(&mc.lock);
5316                         /* We set mc.moving_task later */
5317
5318                         ret = mem_cgroup_precharge_mc(mm);
5319                         if (ret)
5320                                 mem_cgroup_clear_mc();
5321                 }
5322                 mmput(mm);
5323         }
5324         return ret;
5325 }
5326
5327 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5328                                 struct cgroup *cgroup,
5329                                 struct task_struct *p)
5330 {
5331         mem_cgroup_clear_mc();
5332 }
5333
5334 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5335                                 unsigned long addr, unsigned long end,
5336                                 struct mm_walk *walk)
5337 {
5338         int ret = 0;
5339         struct vm_area_struct *vma = walk->private;
5340         pte_t *pte;
5341         spinlock_t *ptl;
5342
5343         split_huge_page_pmd(walk->mm, pmd);
5344 retry:
5345         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5346         for (; addr != end; addr += PAGE_SIZE) {
5347                 pte_t ptent = *(pte++);
5348                 union mc_target target;
5349                 int type;
5350                 struct page *page;
5351                 struct page_cgroup *pc;
5352                 swp_entry_t ent;
5353
5354                 if (!mc.precharge)
5355                         break;
5356
5357                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5358                 switch (type) {
5359                 case MC_TARGET_PAGE:
5360                         page = target.page;
5361                         if (isolate_lru_page(page))
5362                                 goto put;
5363                         pc = lookup_page_cgroup(page);
5364                         if (!mem_cgroup_move_account(page, 1, pc,
5365                                                      mc.from, mc.to, false)) {
5366                                 mc.precharge--;
5367                                 /* we uncharge from mc.from later. */
5368                                 mc.moved_charge++;
5369                         }
5370                         putback_lru_page(page);
5371 put:                    /* is_target_pte_for_mc() gets the page */
5372                         put_page(page);
5373                         break;
5374                 case MC_TARGET_SWAP:
5375                         ent = target.ent;
5376                         if (!mem_cgroup_move_swap_account(ent,
5377                                                 mc.from, mc.to, false)) {
5378                                 mc.precharge--;
5379                                 /* we fixup refcnts and charges later. */
5380                                 mc.moved_swap++;
5381                         }
5382                         break;
5383                 default:
5384                         break;
5385                 }
5386         }
5387         pte_unmap_unlock(pte - 1, ptl);
5388         cond_resched();
5389
5390         if (addr != end) {
5391                 /*
5392                  * We have consumed all precharges we got in can_attach().
5393                  * We try charge one by one, but don't do any additional
5394                  * charges to mc.to if we have failed in charge once in attach()
5395                  * phase.
5396                  */
5397                 ret = mem_cgroup_do_precharge(1);
5398                 if (!ret)
5399                         goto retry;
5400         }
5401
5402         return ret;
5403 }
5404
5405 static void mem_cgroup_move_charge(struct mm_struct *mm)
5406 {
5407         struct vm_area_struct *vma;
5408
5409         lru_add_drain_all();
5410 retry:
5411         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5412                 /*
5413                  * Someone who are holding the mmap_sem might be waiting in
5414                  * waitq. So we cancel all extra charges, wake up all waiters,
5415                  * and retry. Because we cancel precharges, we might not be able
5416                  * to move enough charges, but moving charge is a best-effort
5417                  * feature anyway, so it wouldn't be a big problem.
5418                  */
5419                 __mem_cgroup_clear_mc();
5420                 cond_resched();
5421                 goto retry;
5422         }
5423         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5424                 int ret;
5425                 struct mm_walk mem_cgroup_move_charge_walk = {
5426                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5427                         .mm = mm,
5428                         .private = vma,
5429                 };
5430                 if (is_vm_hugetlb_page(vma))
5431                         continue;
5432                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5433                                                 &mem_cgroup_move_charge_walk);
5434                 if (ret)
5435                         /*
5436                          * means we have consumed all precharges and failed in
5437                          * doing additional charge. Just abandon here.
5438                          */
5439                         break;
5440         }
5441         up_read(&mm->mmap_sem);
5442 }
5443
5444 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5445                                 struct cgroup *cont,
5446                                 struct cgroup *old_cont,
5447                                 struct task_struct *p)
5448 {
5449         struct mm_struct *mm = get_task_mm(p);
5450
5451         if (mm) {
5452                 if (mc.to)
5453                         mem_cgroup_move_charge(mm);
5454                 put_swap_token(mm);
5455                 mmput(mm);
5456         }
5457         if (mc.to)
5458                 mem_cgroup_clear_mc();
5459 }
5460 #else   /* !CONFIG_MMU */
5461 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5462                                 struct cgroup *cgroup,
5463                                 struct task_struct *p)
5464 {
5465         return 0;
5466 }
5467 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5468                                 struct cgroup *cgroup,
5469                                 struct task_struct *p)
5470 {
5471 }
5472 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5473                                 struct cgroup *cont,
5474                                 struct cgroup *old_cont,
5475                                 struct task_struct *p)
5476 {
5477 }
5478 #endif
5479
5480 struct cgroup_subsys mem_cgroup_subsys = {
5481         .name = "memory",
5482         .subsys_id = mem_cgroup_subsys_id,
5483         .create = mem_cgroup_create,
5484         .pre_destroy = mem_cgroup_pre_destroy,
5485         .destroy = mem_cgroup_destroy,
5486         .populate = mem_cgroup_populate,
5487         .can_attach = mem_cgroup_can_attach,
5488         .cancel_attach = mem_cgroup_cancel_attach,
5489         .attach = mem_cgroup_move_task,
5490         .early_init = 0,
5491         .use_id = 1,
5492 };
5493
5494 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5495 static int __init enable_swap_account(char *s)
5496 {
5497         /* consider enabled if no parameter or 1 is given */
5498         if (!strcmp(s, "1"))
5499                 really_do_swap_account = 1;
5500         else if (!strcmp(s, "0"))
5501                 really_do_swap_account = 0;
5502         return 1;
5503 }
5504 __setup("swapaccount=", enable_swap_account);
5505
5506 #endif