tmpfs: convert mem_cgroup shmem to radix-swap
[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 enum {
208         SCAN_BY_LIMIT,
209         SCAN_BY_SYSTEM,
210         NR_SCAN_CONTEXT,
211         SCAN_BY_SHRINK, /* not recorded now */
212 };
213
214 enum {
215         SCAN,
216         SCAN_ANON,
217         SCAN_FILE,
218         ROTATE,
219         ROTATE_ANON,
220         ROTATE_FILE,
221         FREED,
222         FREED_ANON,
223         FREED_FILE,
224         ELAPSED,
225         NR_SCANSTATS,
226 };
227
228 struct scanstat {
229         spinlock_t      lock;
230         unsigned long   stats[NR_SCAN_CONTEXT][NR_SCANSTATS];
231         unsigned long   rootstats[NR_SCAN_CONTEXT][NR_SCANSTATS];
232 };
233
234 const char *scanstat_string[NR_SCANSTATS] = {
235         "scanned_pages",
236         "scanned_anon_pages",
237         "scanned_file_pages",
238         "rotated_pages",
239         "rotated_anon_pages",
240         "rotated_file_pages",
241         "freed_pages",
242         "freed_anon_pages",
243         "freed_file_pages",
244         "elapsed_ns",
245 };
246 #define SCANSTAT_WORD_LIMIT     "_by_limit"
247 #define SCANSTAT_WORD_SYSTEM    "_by_system"
248 #define SCANSTAT_WORD_HIERARCHY "_under_hierarchy"
249
250
251 /*
252  * The memory controller data structure. The memory controller controls both
253  * page cache and RSS per cgroup. We would eventually like to provide
254  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
255  * to help the administrator determine what knobs to tune.
256  *
257  * TODO: Add a water mark for the memory controller. Reclaim will begin when
258  * we hit the water mark. May be even add a low water mark, such that
259  * no reclaim occurs from a cgroup at it's low water mark, this is
260  * a feature that will be implemented much later in the future.
261  */
262 struct mem_cgroup {
263         struct cgroup_subsys_state css;
264         /*
265          * the counter to account for memory usage
266          */
267         struct res_counter res;
268         /*
269          * the counter to account for mem+swap usage.
270          */
271         struct res_counter memsw;
272         /*
273          * Per cgroup active and inactive list, similar to the
274          * per zone LRU lists.
275          */
276         struct mem_cgroup_lru_info info;
277         /*
278          * While reclaiming in a hierarchy, we cache the last child we
279          * reclaimed from.
280          */
281         int last_scanned_child;
282         int last_scanned_node;
283 #if MAX_NUMNODES > 1
284         nodemask_t      scan_nodes;
285         atomic_t        numainfo_events;
286         atomic_t        numainfo_updating;
287 #endif
288         /*
289          * Should the accounting and control be hierarchical, per subtree?
290          */
291         bool use_hierarchy;
292
293         bool            oom_lock;
294         atomic_t        under_oom;
295
296         atomic_t        refcnt;
297
298         int     swappiness;
299         /* OOM-Killer disable */
300         int             oom_kill_disable;
301
302         /* set when res.limit == memsw.limit */
303         bool            memsw_is_minimum;
304
305         /* protect arrays of thresholds */
306         struct mutex thresholds_lock;
307
308         /* thresholds for memory usage. RCU-protected */
309         struct mem_cgroup_thresholds thresholds;
310
311         /* thresholds for mem+swap usage. RCU-protected */
312         struct mem_cgroup_thresholds memsw_thresholds;
313
314         /* For oom notifier event fd */
315         struct list_head oom_notify;
316         /* For recording LRU-scan statistics */
317         struct scanstat scanstat;
318         /*
319          * Should we move charges of a task when a task is moved into this
320          * mem_cgroup ? And what type of charges should we move ?
321          */
322         unsigned long   move_charge_at_immigrate;
323         /*
324          * percpu counter.
325          */
326         struct mem_cgroup_stat_cpu *stat;
327         /*
328          * used when a cpu is offlined or other synchronizations
329          * See mem_cgroup_read_stat().
330          */
331         struct mem_cgroup_stat_cpu nocpu_base;
332         spinlock_t pcp_counter_lock;
333 };
334
335 /* Stuffs for move charges at task migration. */
336 /*
337  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
338  * left-shifted bitmap of these types.
339  */
340 enum move_type {
341         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
342         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
343         NR_MOVE_TYPE,
344 };
345
346 /* "mc" and its members are protected by cgroup_mutex */
347 static struct move_charge_struct {
348         spinlock_t        lock; /* for from, to */
349         struct mem_cgroup *from;
350         struct mem_cgroup *to;
351         unsigned long precharge;
352         unsigned long moved_charge;
353         unsigned long moved_swap;
354         struct task_struct *moving_task;        /* a task moving charges */
355         wait_queue_head_t waitq;                /* a waitq for other context */
356 } mc = {
357         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
358         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
359 };
360
361 static bool move_anon(void)
362 {
363         return test_bit(MOVE_CHARGE_TYPE_ANON,
364                                         &mc.to->move_charge_at_immigrate);
365 }
366
367 static bool move_file(void)
368 {
369         return test_bit(MOVE_CHARGE_TYPE_FILE,
370                                         &mc.to->move_charge_at_immigrate);
371 }
372
373 /*
374  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
375  * limit reclaim to prevent infinite loops, if they ever occur.
376  */
377 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
378 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
379
380 enum charge_type {
381         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
382         MEM_CGROUP_CHARGE_TYPE_MAPPED,
383         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
384         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
385         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
386         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
387         NR_CHARGE_TYPE,
388 };
389
390 /* for encoding cft->private value on file */
391 #define _MEM                    (0)
392 #define _MEMSWAP                (1)
393 #define _OOM_TYPE               (2)
394 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
395 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
396 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
397 /* Used for OOM nofiier */
398 #define OOM_CONTROL             (0)
399
400 /*
401  * Reclaim flags for mem_cgroup_hierarchical_reclaim
402  */
403 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
404 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
405 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
406 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
407 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
408 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
409
410 static void mem_cgroup_get(struct mem_cgroup *mem);
411 static void mem_cgroup_put(struct mem_cgroup *mem);
412 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
413 static void drain_all_stock_async(struct mem_cgroup *mem);
414
415 static struct mem_cgroup_per_zone *
416 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
417 {
418         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
419 }
420
421 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
422 {
423         return &mem->css;
424 }
425
426 static struct mem_cgroup_per_zone *
427 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
428 {
429         int nid = page_to_nid(page);
430         int zid = page_zonenum(page);
431
432         return mem_cgroup_zoneinfo(mem, nid, zid);
433 }
434
435 static struct mem_cgroup_tree_per_zone *
436 soft_limit_tree_node_zone(int nid, int zid)
437 {
438         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
439 }
440
441 static struct mem_cgroup_tree_per_zone *
442 soft_limit_tree_from_page(struct page *page)
443 {
444         int nid = page_to_nid(page);
445         int zid = page_zonenum(page);
446
447         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
448 }
449
450 static void
451 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
452                                 struct mem_cgroup_per_zone *mz,
453                                 struct mem_cgroup_tree_per_zone *mctz,
454                                 unsigned long long new_usage_in_excess)
455 {
456         struct rb_node **p = &mctz->rb_root.rb_node;
457         struct rb_node *parent = NULL;
458         struct mem_cgroup_per_zone *mz_node;
459
460         if (mz->on_tree)
461                 return;
462
463         mz->usage_in_excess = new_usage_in_excess;
464         if (!mz->usage_in_excess)
465                 return;
466         while (*p) {
467                 parent = *p;
468                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
469                                         tree_node);
470                 if (mz->usage_in_excess < mz_node->usage_in_excess)
471                         p = &(*p)->rb_left;
472                 /*
473                  * We can't avoid mem cgroups that are over their soft
474                  * limit by the same amount
475                  */
476                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
477                         p = &(*p)->rb_right;
478         }
479         rb_link_node(&mz->tree_node, parent, p);
480         rb_insert_color(&mz->tree_node, &mctz->rb_root);
481         mz->on_tree = true;
482 }
483
484 static void
485 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
486                                 struct mem_cgroup_per_zone *mz,
487                                 struct mem_cgroup_tree_per_zone *mctz)
488 {
489         if (!mz->on_tree)
490                 return;
491         rb_erase(&mz->tree_node, &mctz->rb_root);
492         mz->on_tree = false;
493 }
494
495 static void
496 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
497                                 struct mem_cgroup_per_zone *mz,
498                                 struct mem_cgroup_tree_per_zone *mctz)
499 {
500         spin_lock(&mctz->lock);
501         __mem_cgroup_remove_exceeded(mem, mz, mctz);
502         spin_unlock(&mctz->lock);
503 }
504
505
506 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
507 {
508         unsigned long long excess;
509         struct mem_cgroup_per_zone *mz;
510         struct mem_cgroup_tree_per_zone *mctz;
511         int nid = page_to_nid(page);
512         int zid = page_zonenum(page);
513         mctz = soft_limit_tree_from_page(page);
514
515         /*
516          * Necessary to update all ancestors when hierarchy is used.
517          * because their event counter is not touched.
518          */
519         for (; mem; mem = parent_mem_cgroup(mem)) {
520                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
521                 excess = res_counter_soft_limit_excess(&mem->res);
522                 /*
523                  * We have to update the tree if mz is on RB-tree or
524                  * mem is over its softlimit.
525                  */
526                 if (excess || mz->on_tree) {
527                         spin_lock(&mctz->lock);
528                         /* if on-tree, remove it */
529                         if (mz->on_tree)
530                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
531                         /*
532                          * Insert again. mz->usage_in_excess will be updated.
533                          * If excess is 0, no tree ops.
534                          */
535                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
536                         spin_unlock(&mctz->lock);
537                 }
538         }
539 }
540
541 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
542 {
543         int node, zone;
544         struct mem_cgroup_per_zone *mz;
545         struct mem_cgroup_tree_per_zone *mctz;
546
547         for_each_node_state(node, N_POSSIBLE) {
548                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
549                         mz = mem_cgroup_zoneinfo(mem, node, zone);
550                         mctz = soft_limit_tree_node_zone(node, zone);
551                         mem_cgroup_remove_exceeded(mem, mz, mctz);
552                 }
553         }
554 }
555
556 static struct mem_cgroup_per_zone *
557 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
558 {
559         struct rb_node *rightmost = NULL;
560         struct mem_cgroup_per_zone *mz;
561
562 retry:
563         mz = NULL;
564         rightmost = rb_last(&mctz->rb_root);
565         if (!rightmost)
566                 goto done;              /* Nothing to reclaim from */
567
568         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
569         /*
570          * Remove the node now but someone else can add it back,
571          * we will to add it back at the end of reclaim to its correct
572          * position in the tree.
573          */
574         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
575         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
576                 !css_tryget(&mz->mem->css))
577                 goto retry;
578 done:
579         return mz;
580 }
581
582 static struct mem_cgroup_per_zone *
583 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
584 {
585         struct mem_cgroup_per_zone *mz;
586
587         spin_lock(&mctz->lock);
588         mz = __mem_cgroup_largest_soft_limit_node(mctz);
589         spin_unlock(&mctz->lock);
590         return mz;
591 }
592
593 /*
594  * Implementation Note: reading percpu statistics for memcg.
595  *
596  * Both of vmstat[] and percpu_counter has threshold and do periodic
597  * synchronization to implement "quick" read. There are trade-off between
598  * reading cost and precision of value. Then, we may have a chance to implement
599  * a periodic synchronizion of counter in memcg's counter.
600  *
601  * But this _read() function is used for user interface now. The user accounts
602  * memory usage by memory cgroup and he _always_ requires exact value because
603  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
604  * have to visit all online cpus and make sum. So, for now, unnecessary
605  * synchronization is not implemented. (just implemented for cpu hotplug)
606  *
607  * If there are kernel internal actions which can make use of some not-exact
608  * value, and reading all cpu value can be performance bottleneck in some
609  * common workload, threashold and synchonization as vmstat[] should be
610  * implemented.
611  */
612 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
613                                  enum mem_cgroup_stat_index idx)
614 {
615         long val = 0;
616         int cpu;
617
618         get_online_cpus();
619         for_each_online_cpu(cpu)
620                 val += per_cpu(mem->stat->count[idx], cpu);
621 #ifdef CONFIG_HOTPLUG_CPU
622         spin_lock(&mem->pcp_counter_lock);
623         val += mem->nocpu_base.count[idx];
624         spin_unlock(&mem->pcp_counter_lock);
625 #endif
626         put_online_cpus();
627         return val;
628 }
629
630 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
631                                          bool charge)
632 {
633         int val = (charge) ? 1 : -1;
634         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
635 }
636
637 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
638 {
639         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
640 }
641
642 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
643 {
644         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
645 }
646
647 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
648                                             enum mem_cgroup_events_index idx)
649 {
650         unsigned long val = 0;
651         int cpu;
652
653         for_each_online_cpu(cpu)
654                 val += per_cpu(mem->stat->events[idx], cpu);
655 #ifdef CONFIG_HOTPLUG_CPU
656         spin_lock(&mem->pcp_counter_lock);
657         val += mem->nocpu_base.events[idx];
658         spin_unlock(&mem->pcp_counter_lock);
659 #endif
660         return val;
661 }
662
663 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
664                                          bool file, int nr_pages)
665 {
666         preempt_disable();
667
668         if (file)
669                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
670         else
671                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
672
673         /* pagein of a big page is an event. So, ignore page size */
674         if (nr_pages > 0)
675                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
676         else {
677                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
678                 nr_pages = -nr_pages; /* for event */
679         }
680
681         __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
682
683         preempt_enable();
684 }
685
686 unsigned long
687 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
688                         unsigned int lru_mask)
689 {
690         struct mem_cgroup_per_zone *mz;
691         enum lru_list l;
692         unsigned long ret = 0;
693
694         mz = mem_cgroup_zoneinfo(mem, nid, zid);
695
696         for_each_lru(l) {
697                 if (BIT(l) & lru_mask)
698                         ret += MEM_CGROUP_ZSTAT(mz, l);
699         }
700         return ret;
701 }
702
703 static unsigned long
704 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
705                         int nid, unsigned int lru_mask)
706 {
707         u64 total = 0;
708         int zid;
709
710         for (zid = 0; zid < MAX_NR_ZONES; zid++)
711                 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
712
713         return total;
714 }
715
716 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
717                         unsigned int lru_mask)
718 {
719         int nid;
720         u64 total = 0;
721
722         for_each_node_state(nid, N_HIGH_MEMORY)
723                 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
724         return total;
725 }
726
727 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
728 {
729         unsigned long val, next;
730
731         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
732         next = this_cpu_read(mem->stat->targets[target]);
733         /* from time_after() in jiffies.h */
734         return ((long)next - (long)val < 0);
735 }
736
737 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
738 {
739         unsigned long val, next;
740
741         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
742
743         switch (target) {
744         case MEM_CGROUP_TARGET_THRESH:
745                 next = val + THRESHOLDS_EVENTS_TARGET;
746                 break;
747         case MEM_CGROUP_TARGET_SOFTLIMIT:
748                 next = val + SOFTLIMIT_EVENTS_TARGET;
749                 break;
750         case MEM_CGROUP_TARGET_NUMAINFO:
751                 next = val + NUMAINFO_EVENTS_TARGET;
752                 break;
753         default:
754                 return;
755         }
756
757         this_cpu_write(mem->stat->targets[target], next);
758 }
759
760 /*
761  * Check events in order.
762  *
763  */
764 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
765 {
766         /* threshold event is triggered in finer grain than soft limit */
767         if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
768                 mem_cgroup_threshold(mem);
769                 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
770                 if (unlikely(__memcg_event_check(mem,
771                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
772                         mem_cgroup_update_tree(mem, page);
773                         __mem_cgroup_target_update(mem,
774                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
775                 }
776 #if MAX_NUMNODES > 1
777                 if (unlikely(__memcg_event_check(mem,
778                         MEM_CGROUP_TARGET_NUMAINFO))) {
779                         atomic_inc(&mem->numainfo_events);
780                         __mem_cgroup_target_update(mem,
781                                 MEM_CGROUP_TARGET_NUMAINFO);
782                 }
783 #endif
784         }
785 }
786
787 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
788 {
789         return container_of(cgroup_subsys_state(cont,
790                                 mem_cgroup_subsys_id), struct mem_cgroup,
791                                 css);
792 }
793
794 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
795 {
796         /*
797          * mm_update_next_owner() may clear mm->owner to NULL
798          * if it races with swapoff, page migration, etc.
799          * So this can be called with p == NULL.
800          */
801         if (unlikely(!p))
802                 return NULL;
803
804         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
805                                 struct mem_cgroup, css);
806 }
807
808 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
809 {
810         struct mem_cgroup *mem = NULL;
811
812         if (!mm)
813                 return NULL;
814         /*
815          * Because we have no locks, mm->owner's may be being moved to other
816          * cgroup. We use css_tryget() here even if this looks
817          * pessimistic (rather than adding locks here).
818          */
819         rcu_read_lock();
820         do {
821                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
822                 if (unlikely(!mem))
823                         break;
824         } while (!css_tryget(&mem->css));
825         rcu_read_unlock();
826         return mem;
827 }
828
829 /* The caller has to guarantee "mem" exists before calling this */
830 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
831 {
832         struct cgroup_subsys_state *css;
833         int found;
834
835         if (!mem) /* ROOT cgroup has the smallest ID */
836                 return root_mem_cgroup; /*css_put/get against root is ignored*/
837         if (!mem->use_hierarchy) {
838                 if (css_tryget(&mem->css))
839                         return mem;
840                 return NULL;
841         }
842         rcu_read_lock();
843         /*
844          * searching a memory cgroup which has the smallest ID under given
845          * ROOT cgroup. (ID >= 1)
846          */
847         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
848         if (css && css_tryget(css))
849                 mem = container_of(css, struct mem_cgroup, css);
850         else
851                 mem = NULL;
852         rcu_read_unlock();
853         return mem;
854 }
855
856 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
857                                         struct mem_cgroup *root,
858                                         bool cond)
859 {
860         int nextid = css_id(&iter->css) + 1;
861         int found;
862         int hierarchy_used;
863         struct cgroup_subsys_state *css;
864
865         hierarchy_used = iter->use_hierarchy;
866
867         css_put(&iter->css);
868         /* If no ROOT, walk all, ignore hierarchy */
869         if (!cond || (root && !hierarchy_used))
870                 return NULL;
871
872         if (!root)
873                 root = root_mem_cgroup;
874
875         do {
876                 iter = NULL;
877                 rcu_read_lock();
878
879                 css = css_get_next(&mem_cgroup_subsys, nextid,
880                                 &root->css, &found);
881                 if (css && css_tryget(css))
882                         iter = container_of(css, struct mem_cgroup, css);
883                 rcu_read_unlock();
884                 /* If css is NULL, no more cgroups will be found */
885                 nextid = found + 1;
886         } while (css && !iter);
887
888         return iter;
889 }
890 /*
891  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
892  * be careful that "break" loop is not allowed. We have reference count.
893  * Instead of that modify "cond" to be false and "continue" to exit the loop.
894  */
895 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
896         for (iter = mem_cgroup_start_loop(root);\
897              iter != NULL;\
898              iter = mem_cgroup_get_next(iter, root, cond))
899
900 #define for_each_mem_cgroup_tree(iter, root) \
901         for_each_mem_cgroup_tree_cond(iter, root, true)
902
903 #define for_each_mem_cgroup_all(iter) \
904         for_each_mem_cgroup_tree_cond(iter, NULL, true)
905
906
907 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
908 {
909         return (mem == root_mem_cgroup);
910 }
911
912 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
913 {
914         struct mem_cgroup *mem;
915
916         if (!mm)
917                 return;
918
919         rcu_read_lock();
920         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
921         if (unlikely(!mem))
922                 goto out;
923
924         switch (idx) {
925         case PGMAJFAULT:
926                 mem_cgroup_pgmajfault(mem, 1);
927                 break;
928         case PGFAULT:
929                 mem_cgroup_pgfault(mem, 1);
930                 break;
931         default:
932                 BUG();
933         }
934 out:
935         rcu_read_unlock();
936 }
937 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
938
939 /*
940  * Following LRU functions are allowed to be used without PCG_LOCK.
941  * Operations are called by routine of global LRU independently from memcg.
942  * What we have to take care of here is validness of pc->mem_cgroup.
943  *
944  * Changes to pc->mem_cgroup happens when
945  * 1. charge
946  * 2. moving account
947  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
948  * It is added to LRU before charge.
949  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
950  * When moving account, the page is not on LRU. It's isolated.
951  */
952
953 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
954 {
955         struct page_cgroup *pc;
956         struct mem_cgroup_per_zone *mz;
957
958         if (mem_cgroup_disabled())
959                 return;
960         pc = lookup_page_cgroup(page);
961         /* can happen while we handle swapcache. */
962         if (!TestClearPageCgroupAcctLRU(pc))
963                 return;
964         VM_BUG_ON(!pc->mem_cgroup);
965         /*
966          * We don't check PCG_USED bit. It's cleared when the "page" is finally
967          * removed from global LRU.
968          */
969         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
970         /* huge page split is done under lru_lock. so, we have no races. */
971         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
972         if (mem_cgroup_is_root(pc->mem_cgroup))
973                 return;
974         VM_BUG_ON(list_empty(&pc->lru));
975         list_del_init(&pc->lru);
976 }
977
978 void mem_cgroup_del_lru(struct page *page)
979 {
980         mem_cgroup_del_lru_list(page, page_lru(page));
981 }
982
983 /*
984  * Writeback is about to end against a page which has been marked for immediate
985  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
986  * inactive list.
987  */
988 void mem_cgroup_rotate_reclaimable_page(struct page *page)
989 {
990         struct mem_cgroup_per_zone *mz;
991         struct page_cgroup *pc;
992         enum lru_list lru = page_lru(page);
993
994         if (mem_cgroup_disabled())
995                 return;
996
997         pc = lookup_page_cgroup(page);
998         /* unused or root page is not rotated. */
999         if (!PageCgroupUsed(pc))
1000                 return;
1001         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1002         smp_rmb();
1003         if (mem_cgroup_is_root(pc->mem_cgroup))
1004                 return;
1005         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1006         list_move_tail(&pc->lru, &mz->lists[lru]);
1007 }
1008
1009 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1010 {
1011         struct mem_cgroup_per_zone *mz;
1012         struct page_cgroup *pc;
1013
1014         if (mem_cgroup_disabled())
1015                 return;
1016
1017         pc = lookup_page_cgroup(page);
1018         /* unused or root page is not rotated. */
1019         if (!PageCgroupUsed(pc))
1020                 return;
1021         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1022         smp_rmb();
1023         if (mem_cgroup_is_root(pc->mem_cgroup))
1024                 return;
1025         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1026         list_move(&pc->lru, &mz->lists[lru]);
1027 }
1028
1029 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1030 {
1031         struct page_cgroup *pc;
1032         struct mem_cgroup_per_zone *mz;
1033
1034         if (mem_cgroup_disabled())
1035                 return;
1036         pc = lookup_page_cgroup(page);
1037         VM_BUG_ON(PageCgroupAcctLRU(pc));
1038         if (!PageCgroupUsed(pc))
1039                 return;
1040         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1041         smp_rmb();
1042         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1043         /* huge page split is done under lru_lock. so, we have no races. */
1044         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1045         SetPageCgroupAcctLRU(pc);
1046         if (mem_cgroup_is_root(pc->mem_cgroup))
1047                 return;
1048         list_add(&pc->lru, &mz->lists[lru]);
1049 }
1050
1051 /*
1052  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1053  * while it's linked to lru because the page may be reused after it's fully
1054  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1055  * It's done under lock_page and expected that zone->lru_lock isnever held.
1056  */
1057 static void mem_cgroup_lru_del_before_commit(struct page *page)
1058 {
1059         unsigned long flags;
1060         struct zone *zone = page_zone(page);
1061         struct page_cgroup *pc = lookup_page_cgroup(page);
1062
1063         /*
1064          * Doing this check without taking ->lru_lock seems wrong but this
1065          * is safe. Because if page_cgroup's USED bit is unset, the page
1066          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1067          * set, the commit after this will fail, anyway.
1068          * This all charge/uncharge is done under some mutual execustion.
1069          * So, we don't need to taking care of changes in USED bit.
1070          */
1071         if (likely(!PageLRU(page)))
1072                 return;
1073
1074         spin_lock_irqsave(&zone->lru_lock, flags);
1075         /*
1076          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1077          * is guarded by lock_page() because the page is SwapCache.
1078          */
1079         if (!PageCgroupUsed(pc))
1080                 mem_cgroup_del_lru_list(page, page_lru(page));
1081         spin_unlock_irqrestore(&zone->lru_lock, flags);
1082 }
1083
1084 static void mem_cgroup_lru_add_after_commit(struct page *page)
1085 {
1086         unsigned long flags;
1087         struct zone *zone = page_zone(page);
1088         struct page_cgroup *pc = lookup_page_cgroup(page);
1089
1090         /* taking care of that the page is added to LRU while we commit it */
1091         if (likely(!PageLRU(page)))
1092                 return;
1093         spin_lock_irqsave(&zone->lru_lock, flags);
1094         /* link when the page is linked to LRU but page_cgroup isn't */
1095         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1096                 mem_cgroup_add_lru_list(page, page_lru(page));
1097         spin_unlock_irqrestore(&zone->lru_lock, flags);
1098 }
1099
1100
1101 void mem_cgroup_move_lists(struct page *page,
1102                            enum lru_list from, enum lru_list to)
1103 {
1104         if (mem_cgroup_disabled())
1105                 return;
1106         mem_cgroup_del_lru_list(page, from);
1107         mem_cgroup_add_lru_list(page, to);
1108 }
1109
1110 /*
1111  * Checks whether given mem is same or in the root_mem's
1112  * hierarchy subtree
1113  */
1114 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_mem,
1115                 struct mem_cgroup *mem)
1116 {
1117         if (root_mem != mem) {
1118                 return (root_mem->use_hierarchy &&
1119                         css_is_ancestor(&mem->css, &root_mem->css));
1120         }
1121
1122         return true;
1123 }
1124
1125 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1126 {
1127         int ret;
1128         struct mem_cgroup *curr = NULL;
1129         struct task_struct *p;
1130
1131         p = find_lock_task_mm(task);
1132         if (!p)
1133                 return 0;
1134         curr = try_get_mem_cgroup_from_mm(p->mm);
1135         task_unlock(p);
1136         if (!curr)
1137                 return 0;
1138         /*
1139          * We should check use_hierarchy of "mem" not "curr". Because checking
1140          * use_hierarchy of "curr" here make this function true if hierarchy is
1141          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1142          * hierarchy(even if use_hierarchy is disabled in "mem").
1143          */
1144         ret = mem_cgroup_same_or_subtree(mem, curr);
1145         css_put(&curr->css);
1146         return ret;
1147 }
1148
1149 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1150 {
1151         unsigned long active;
1152         unsigned long inactive;
1153         unsigned long gb;
1154         unsigned long inactive_ratio;
1155
1156         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1157         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1158
1159         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1160         if (gb)
1161                 inactive_ratio = int_sqrt(10 * gb);
1162         else
1163                 inactive_ratio = 1;
1164
1165         if (present_pages) {
1166                 present_pages[0] = inactive;
1167                 present_pages[1] = active;
1168         }
1169
1170         return inactive_ratio;
1171 }
1172
1173 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1174 {
1175         unsigned long active;
1176         unsigned long inactive;
1177         unsigned long present_pages[2];
1178         unsigned long inactive_ratio;
1179
1180         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1181
1182         inactive = present_pages[0];
1183         active = present_pages[1];
1184
1185         if (inactive * inactive_ratio < active)
1186                 return 1;
1187
1188         return 0;
1189 }
1190
1191 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1192 {
1193         unsigned long active;
1194         unsigned long inactive;
1195
1196         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1197         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1198
1199         return (active > inactive);
1200 }
1201
1202 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1203                                                       struct zone *zone)
1204 {
1205         int nid = zone_to_nid(zone);
1206         int zid = zone_idx(zone);
1207         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1208
1209         return &mz->reclaim_stat;
1210 }
1211
1212 struct zone_reclaim_stat *
1213 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1214 {
1215         struct page_cgroup *pc;
1216         struct mem_cgroup_per_zone *mz;
1217
1218         if (mem_cgroup_disabled())
1219                 return NULL;
1220
1221         pc = lookup_page_cgroup(page);
1222         if (!PageCgroupUsed(pc))
1223                 return NULL;
1224         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1225         smp_rmb();
1226         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1227         return &mz->reclaim_stat;
1228 }
1229
1230 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1231                                         struct list_head *dst,
1232                                         unsigned long *scanned, int order,
1233                                         int mode, struct zone *z,
1234                                         struct mem_cgroup *mem_cont,
1235                                         int active, int file)
1236 {
1237         unsigned long nr_taken = 0;
1238         struct page *page;
1239         unsigned long scan;
1240         LIST_HEAD(pc_list);
1241         struct list_head *src;
1242         struct page_cgroup *pc, *tmp;
1243         int nid = zone_to_nid(z);
1244         int zid = zone_idx(z);
1245         struct mem_cgroup_per_zone *mz;
1246         int lru = LRU_FILE * file + active;
1247         int ret;
1248
1249         BUG_ON(!mem_cont);
1250         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1251         src = &mz->lists[lru];
1252
1253         scan = 0;
1254         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1255                 if (scan >= nr_to_scan)
1256                         break;
1257
1258                 if (unlikely(!PageCgroupUsed(pc)))
1259                         continue;
1260
1261                 page = lookup_cgroup_page(pc);
1262
1263                 if (unlikely(!PageLRU(page)))
1264                         continue;
1265
1266                 scan++;
1267                 ret = __isolate_lru_page(page, mode, file);
1268                 switch (ret) {
1269                 case 0:
1270                         list_move(&page->lru, dst);
1271                         mem_cgroup_del_lru(page);
1272                         nr_taken += hpage_nr_pages(page);
1273                         break;
1274                 case -EBUSY:
1275                         /* we don't affect global LRU but rotate in our LRU */
1276                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1277                         break;
1278                 default:
1279                         break;
1280                 }
1281         }
1282
1283         *scanned = scan;
1284
1285         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1286                                       0, 0, 0, mode);
1287
1288         return nr_taken;
1289 }
1290
1291 #define mem_cgroup_from_res_counter(counter, member)    \
1292         container_of(counter, struct mem_cgroup, member)
1293
1294 /**
1295  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1296  * @mem: the memory cgroup
1297  *
1298  * Returns the maximum amount of memory @mem can be charged with, in
1299  * pages.
1300  */
1301 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1302 {
1303         unsigned long long margin;
1304
1305         margin = res_counter_margin(&mem->res);
1306         if (do_swap_account)
1307                 margin = min(margin, res_counter_margin(&mem->memsw));
1308         return margin >> PAGE_SHIFT;
1309 }
1310
1311 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1312 {
1313         struct cgroup *cgrp = memcg->css.cgroup;
1314
1315         /* root ? */
1316         if (cgrp->parent == NULL)
1317                 return vm_swappiness;
1318
1319         return memcg->swappiness;
1320 }
1321
1322 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1323 {
1324         int cpu;
1325
1326         get_online_cpus();
1327         spin_lock(&mem->pcp_counter_lock);
1328         for_each_online_cpu(cpu)
1329                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1330         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1331         spin_unlock(&mem->pcp_counter_lock);
1332         put_online_cpus();
1333
1334         synchronize_rcu();
1335 }
1336
1337 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1338 {
1339         int cpu;
1340
1341         if (!mem)
1342                 return;
1343         get_online_cpus();
1344         spin_lock(&mem->pcp_counter_lock);
1345         for_each_online_cpu(cpu)
1346                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1347         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1348         spin_unlock(&mem->pcp_counter_lock);
1349         put_online_cpus();
1350 }
1351 /*
1352  * 2 routines for checking "mem" is under move_account() or not.
1353  *
1354  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1355  *                        for avoiding race in accounting. If true,
1356  *                        pc->mem_cgroup may be overwritten.
1357  *
1358  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1359  *                        under hierarchy of moving cgroups. This is for
1360  *                        waiting at hith-memory prressure caused by "move".
1361  */
1362
1363 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1364 {
1365         VM_BUG_ON(!rcu_read_lock_held());
1366         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1367 }
1368
1369 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1370 {
1371         struct mem_cgroup *from;
1372         struct mem_cgroup *to;
1373         bool ret = false;
1374         /*
1375          * Unlike task_move routines, we access mc.to, mc.from not under
1376          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1377          */
1378         spin_lock(&mc.lock);
1379         from = mc.from;
1380         to = mc.to;
1381         if (!from)
1382                 goto unlock;
1383
1384         ret = mem_cgroup_same_or_subtree(mem, from)
1385                 || mem_cgroup_same_or_subtree(mem, to);
1386 unlock:
1387         spin_unlock(&mc.lock);
1388         return ret;
1389 }
1390
1391 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1392 {
1393         if (mc.moving_task && current != mc.moving_task) {
1394                 if (mem_cgroup_under_move(mem)) {
1395                         DEFINE_WAIT(wait);
1396                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1397                         /* moving charge context might have finished. */
1398                         if (mc.moving_task)
1399                                 schedule();
1400                         finish_wait(&mc.waitq, &wait);
1401                         return true;
1402                 }
1403         }
1404         return false;
1405 }
1406
1407 /**
1408  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1409  * @memcg: The memory cgroup that went over limit
1410  * @p: Task that is going to be killed
1411  *
1412  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1413  * enabled
1414  */
1415 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1416 {
1417         struct cgroup *task_cgrp;
1418         struct cgroup *mem_cgrp;
1419         /*
1420          * Need a buffer in BSS, can't rely on allocations. The code relies
1421          * on the assumption that OOM is serialized for memory controller.
1422          * If this assumption is broken, revisit this code.
1423          */
1424         static char memcg_name[PATH_MAX];
1425         int ret;
1426
1427         if (!memcg || !p)
1428                 return;
1429
1430
1431         rcu_read_lock();
1432
1433         mem_cgrp = memcg->css.cgroup;
1434         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1435
1436         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1437         if (ret < 0) {
1438                 /*
1439                  * Unfortunately, we are unable to convert to a useful name
1440                  * But we'll still print out the usage information
1441                  */
1442                 rcu_read_unlock();
1443                 goto done;
1444         }
1445         rcu_read_unlock();
1446
1447         printk(KERN_INFO "Task in %s killed", memcg_name);
1448
1449         rcu_read_lock();
1450         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1451         if (ret < 0) {
1452                 rcu_read_unlock();
1453                 goto done;
1454         }
1455         rcu_read_unlock();
1456
1457         /*
1458          * Continues from above, so we don't need an KERN_ level
1459          */
1460         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1461 done:
1462
1463         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1464                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1465                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1466                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1467         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1468                 "failcnt %llu\n",
1469                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1470                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1471                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1472 }
1473
1474 /*
1475  * This function returns the number of memcg under hierarchy tree. Returns
1476  * 1(self count) if no children.
1477  */
1478 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1479 {
1480         int num = 0;
1481         struct mem_cgroup *iter;
1482
1483         for_each_mem_cgroup_tree(iter, mem)
1484                 num++;
1485         return num;
1486 }
1487
1488 /*
1489  * Return the memory (and swap, if configured) limit for a memcg.
1490  */
1491 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1492 {
1493         u64 limit;
1494         u64 memsw;
1495
1496         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1497         limit += total_swap_pages << PAGE_SHIFT;
1498
1499         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1500         /*
1501          * If memsw is finite and limits the amount of swap space available
1502          * to this memcg, return that limit.
1503          */
1504         return min(limit, memsw);
1505 }
1506
1507 /*
1508  * Visit the first child (need not be the first child as per the ordering
1509  * of the cgroup list, since we track last_scanned_child) of @mem and use
1510  * that to reclaim free pages from.
1511  */
1512 static struct mem_cgroup *
1513 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1514 {
1515         struct mem_cgroup *ret = NULL;
1516         struct cgroup_subsys_state *css;
1517         int nextid, found;
1518
1519         if (!root_mem->use_hierarchy) {
1520                 css_get(&root_mem->css);
1521                 ret = root_mem;
1522         }
1523
1524         while (!ret) {
1525                 rcu_read_lock();
1526                 nextid = root_mem->last_scanned_child + 1;
1527                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1528                                    &found);
1529                 if (css && css_tryget(css))
1530                         ret = container_of(css, struct mem_cgroup, css);
1531
1532                 rcu_read_unlock();
1533                 /* Updates scanning parameter */
1534                 if (!css) {
1535                         /* this means start scan from ID:1 */
1536                         root_mem->last_scanned_child = 0;
1537                 } else
1538                         root_mem->last_scanned_child = found;
1539         }
1540
1541         return ret;
1542 }
1543
1544 /**
1545  * test_mem_cgroup_node_reclaimable
1546  * @mem: the target memcg
1547  * @nid: the node ID to be checked.
1548  * @noswap : specify true here if the user wants flle only information.
1549  *
1550  * This function returns whether the specified memcg contains any
1551  * reclaimable pages on a node. Returns true if there are any reclaimable
1552  * pages in the node.
1553  */
1554 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1555                 int nid, bool noswap)
1556 {
1557         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1558                 return true;
1559         if (noswap || !total_swap_pages)
1560                 return false;
1561         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1562                 return true;
1563         return false;
1564
1565 }
1566 #if MAX_NUMNODES > 1
1567
1568 /*
1569  * Always updating the nodemask is not very good - even if we have an empty
1570  * list or the wrong list here, we can start from some node and traverse all
1571  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1572  *
1573  */
1574 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1575 {
1576         int nid;
1577         /*
1578          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1579          * pagein/pageout changes since the last update.
1580          */
1581         if (!atomic_read(&mem->numainfo_events))
1582                 return;
1583         if (atomic_inc_return(&mem->numainfo_updating) > 1)
1584                 return;
1585
1586         /* make a nodemask where this memcg uses memory from */
1587         mem->scan_nodes = node_states[N_HIGH_MEMORY];
1588
1589         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1590
1591                 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1592                         node_clear(nid, mem->scan_nodes);
1593         }
1594
1595         atomic_set(&mem->numainfo_events, 0);
1596         atomic_set(&mem->numainfo_updating, 0);
1597 }
1598
1599 /*
1600  * Selecting a node where we start reclaim from. Because what we need is just
1601  * reducing usage counter, start from anywhere is O,K. Considering
1602  * memory reclaim from current node, there are pros. and cons.
1603  *
1604  * Freeing memory from current node means freeing memory from a node which
1605  * we'll use or we've used. So, it may make LRU bad. And if several threads
1606  * hit limits, it will see a contention on a node. But freeing from remote
1607  * node means more costs for memory reclaim because of memory latency.
1608  *
1609  * Now, we use round-robin. Better algorithm is welcomed.
1610  */
1611 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1612 {
1613         int node;
1614
1615         mem_cgroup_may_update_nodemask(mem);
1616         node = mem->last_scanned_node;
1617
1618         node = next_node(node, mem->scan_nodes);
1619         if (node == MAX_NUMNODES)
1620                 node = first_node(mem->scan_nodes);
1621         /*
1622          * We call this when we hit limit, not when pages are added to LRU.
1623          * No LRU may hold pages because all pages are UNEVICTABLE or
1624          * memcg is too small and all pages are not on LRU. In that case,
1625          * we use curret node.
1626          */
1627         if (unlikely(node == MAX_NUMNODES))
1628                 node = numa_node_id();
1629
1630         mem->last_scanned_node = node;
1631         return node;
1632 }
1633
1634 /*
1635  * Check all nodes whether it contains reclaimable pages or not.
1636  * For quick scan, we make use of scan_nodes. This will allow us to skip
1637  * unused nodes. But scan_nodes is lazily updated and may not cotain
1638  * enough new information. We need to do double check.
1639  */
1640 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1641 {
1642         int nid;
1643
1644         /*
1645          * quick check...making use of scan_node.
1646          * We can skip unused nodes.
1647          */
1648         if (!nodes_empty(mem->scan_nodes)) {
1649                 for (nid = first_node(mem->scan_nodes);
1650                      nid < MAX_NUMNODES;
1651                      nid = next_node(nid, mem->scan_nodes)) {
1652
1653                         if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1654                                 return true;
1655                 }
1656         }
1657         /*
1658          * Check rest of nodes.
1659          */
1660         for_each_node_state(nid, N_HIGH_MEMORY) {
1661                 if (node_isset(nid, mem->scan_nodes))
1662                         continue;
1663                 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1664                         return true;
1665         }
1666         return false;
1667 }
1668
1669 #else
1670 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1671 {
1672         return 0;
1673 }
1674
1675 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1676 {
1677         return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1678 }
1679 #endif
1680
1681 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1682                            struct memcg_scanrecord *rec)
1683 {
1684
1685         stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1686         stats[SCAN_ANON] += rec->nr_scanned[0];
1687         stats[SCAN_FILE] += rec->nr_scanned[1];
1688
1689         stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1690         stats[ROTATE_ANON] += rec->nr_rotated[0];
1691         stats[ROTATE_FILE] += rec->nr_rotated[1];
1692
1693         stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1694         stats[FREED_ANON] += rec->nr_freed[0];
1695         stats[FREED_FILE] += rec->nr_freed[1];
1696
1697         stats[ELAPSED] += rec->elapsed;
1698 }
1699
1700 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1701 {
1702         struct mem_cgroup *mem;
1703         int context = rec->context;
1704
1705         if (context >= NR_SCAN_CONTEXT)
1706                 return;
1707
1708         mem = rec->mem;
1709         spin_lock(&mem->scanstat.lock);
1710         __mem_cgroup_record_scanstat(mem->scanstat.stats[context], rec);
1711         spin_unlock(&mem->scanstat.lock);
1712
1713         mem = rec->root;
1714         spin_lock(&mem->scanstat.lock);
1715         __mem_cgroup_record_scanstat(mem->scanstat.rootstats[context], rec);
1716         spin_unlock(&mem->scanstat.lock);
1717 }
1718
1719 /*
1720  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1721  * we reclaimed from, so that we don't end up penalizing one child extensively
1722  * based on its position in the children list.
1723  *
1724  * root_mem is the original ancestor that we've been reclaim from.
1725  *
1726  * We give up and return to the caller when we visit root_mem twice.
1727  * (other groups can be removed while we're walking....)
1728  *
1729  * If shrink==true, for avoiding to free too much, this returns immedieately.
1730  */
1731 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1732                                                 struct zone *zone,
1733                                                 gfp_t gfp_mask,
1734                                                 unsigned long reclaim_options,
1735                                                 unsigned long *total_scanned)
1736 {
1737         struct mem_cgroup *victim;
1738         int ret, total = 0;
1739         int loop = 0;
1740         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1741         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1742         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1743         struct memcg_scanrecord rec;
1744         unsigned long excess;
1745         unsigned long scanned;
1746
1747         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1748
1749         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1750         if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1751                 noswap = true;
1752
1753         if (shrink)
1754                 rec.context = SCAN_BY_SHRINK;
1755         else if (check_soft)
1756                 rec.context = SCAN_BY_SYSTEM;
1757         else
1758                 rec.context = SCAN_BY_LIMIT;
1759
1760         rec.root = root_mem;
1761
1762         while (1) {
1763                 victim = mem_cgroup_select_victim(root_mem);
1764                 if (victim == root_mem) {
1765                         loop++;
1766                         /*
1767                          * We are not draining per cpu cached charges during
1768                          * soft limit reclaim  because global reclaim doesn't
1769                          * care about charges. It tries to free some memory and
1770                          * charges will not give any.
1771                          */
1772                         if (!check_soft && loop >= 1)
1773                                 drain_all_stock_async(root_mem);
1774                         if (loop >= 2) {
1775                                 /*
1776                                  * If we have not been able to reclaim
1777                                  * anything, it might because there are
1778                                  * no reclaimable pages under this hierarchy
1779                                  */
1780                                 if (!check_soft || !total) {
1781                                         css_put(&victim->css);
1782                                         break;
1783                                 }
1784                                 /*
1785                                  * We want to do more targeted reclaim.
1786                                  * excess >> 2 is not to excessive so as to
1787                                  * reclaim too much, nor too less that we keep
1788                                  * coming back to reclaim from this cgroup
1789                                  */
1790                                 if (total >= (excess >> 2) ||
1791                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1792                                         css_put(&victim->css);
1793                                         break;
1794                                 }
1795                         }
1796                 }
1797                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1798                         /* this cgroup's local usage == 0 */
1799                         css_put(&victim->css);
1800                         continue;
1801                 }
1802                 rec.mem = victim;
1803                 rec.nr_scanned[0] = 0;
1804                 rec.nr_scanned[1] = 0;
1805                 rec.nr_rotated[0] = 0;
1806                 rec.nr_rotated[1] = 0;
1807                 rec.nr_freed[0] = 0;
1808                 rec.nr_freed[1] = 0;
1809                 rec.elapsed = 0;
1810                 /* we use swappiness of local cgroup */
1811                 if (check_soft) {
1812                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1813                                 noswap, zone, &rec, &scanned);
1814                         *total_scanned += scanned;
1815                 } else
1816                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1817                                                 noswap, &rec);
1818                 mem_cgroup_record_scanstat(&rec);
1819                 css_put(&victim->css);
1820                 /*
1821                  * At shrinking usage, we can't check we should stop here or
1822                  * reclaim more. It's depends on callers. last_scanned_child
1823                  * will work enough for keeping fairness under tree.
1824                  */
1825                 if (shrink)
1826                         return ret;
1827                 total += ret;
1828                 if (check_soft) {
1829                         if (!res_counter_soft_limit_excess(&root_mem->res))
1830                                 return total;
1831                 } else if (mem_cgroup_margin(root_mem))
1832                         return total;
1833         }
1834         return total;
1835 }
1836
1837 /*
1838  * Check OOM-Killer is already running under our hierarchy.
1839  * If someone is running, return false.
1840  * Has to be called with memcg_oom_lock
1841  */
1842 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1843 {
1844         int lock_count = -1;
1845         struct mem_cgroup *iter, *failed = NULL;
1846         bool cond = true;
1847
1848         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1849                 bool locked = iter->oom_lock;
1850
1851                 iter->oom_lock = true;
1852                 if (lock_count == -1)
1853                         lock_count = iter->oom_lock;
1854                 else if (lock_count != locked) {
1855                         /*
1856                          * this subtree of our hierarchy is already locked
1857                          * so we cannot give a lock.
1858                          */
1859                         lock_count = 0;
1860                         failed = iter;
1861                         cond = false;
1862                 }
1863         }
1864
1865         if (!failed)
1866                 goto done;
1867
1868         /*
1869          * OK, we failed to lock the whole subtree so we have to clean up
1870          * what we set up to the failing subtree
1871          */
1872         cond = true;
1873         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1874                 if (iter == failed) {
1875                         cond = false;
1876                         continue;
1877                 }
1878                 iter->oom_lock = false;
1879         }
1880 done:
1881         return lock_count;
1882 }
1883
1884 /*
1885  * Has to be called with memcg_oom_lock
1886  */
1887 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1888 {
1889         struct mem_cgroup *iter;
1890
1891         for_each_mem_cgroup_tree(iter, mem)
1892                 iter->oom_lock = false;
1893         return 0;
1894 }
1895
1896 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1897 {
1898         struct mem_cgroup *iter;
1899
1900         for_each_mem_cgroup_tree(iter, mem)
1901                 atomic_inc(&iter->under_oom);
1902 }
1903
1904 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1905 {
1906         struct mem_cgroup *iter;
1907
1908         /*
1909          * When a new child is created while the hierarchy is under oom,
1910          * mem_cgroup_oom_lock() may not be called. We have to use
1911          * atomic_add_unless() here.
1912          */
1913         for_each_mem_cgroup_tree(iter, mem)
1914                 atomic_add_unless(&iter->under_oom, -1, 0);
1915 }
1916
1917 static DEFINE_SPINLOCK(memcg_oom_lock);
1918 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1919
1920 struct oom_wait_info {
1921         struct mem_cgroup *mem;
1922         wait_queue_t    wait;
1923 };
1924
1925 static int memcg_oom_wake_function(wait_queue_t *wait,
1926         unsigned mode, int sync, void *arg)
1927 {
1928         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg,
1929                           *oom_wait_mem;
1930         struct oom_wait_info *oom_wait_info;
1931
1932         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1933         oom_wait_mem = oom_wait_info->mem;
1934
1935         /*
1936          * Both of oom_wait_info->mem and wake_mem are stable under us.
1937          * Then we can use css_is_ancestor without taking care of RCU.
1938          */
1939         if (!mem_cgroup_same_or_subtree(oom_wait_mem, wake_mem)
1940                         && !mem_cgroup_same_or_subtree(wake_mem, oom_wait_mem))
1941                 return 0;
1942         return autoremove_wake_function(wait, mode, sync, arg);
1943 }
1944
1945 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1946 {
1947         /* for filtering, pass "mem" as argument. */
1948         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1949 }
1950
1951 static void memcg_oom_recover(struct mem_cgroup *mem)
1952 {
1953         if (mem && atomic_read(&mem->under_oom))
1954                 memcg_wakeup_oom(mem);
1955 }
1956
1957 /*
1958  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1959  */
1960 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1961 {
1962         struct oom_wait_info owait;
1963         bool locked, need_to_kill;
1964
1965         owait.mem = mem;
1966         owait.wait.flags = 0;
1967         owait.wait.func = memcg_oom_wake_function;
1968         owait.wait.private = current;
1969         INIT_LIST_HEAD(&owait.wait.task_list);
1970         need_to_kill = true;
1971         mem_cgroup_mark_under_oom(mem);
1972
1973         /* At first, try to OOM lock hierarchy under mem.*/
1974         spin_lock(&memcg_oom_lock);
1975         locked = mem_cgroup_oom_lock(mem);
1976         /*
1977          * Even if signal_pending(), we can't quit charge() loop without
1978          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1979          * under OOM is always welcomed, use TASK_KILLABLE here.
1980          */
1981         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1982         if (!locked || mem->oom_kill_disable)
1983                 need_to_kill = false;
1984         if (locked)
1985                 mem_cgroup_oom_notify(mem);
1986         spin_unlock(&memcg_oom_lock);
1987
1988         if (need_to_kill) {
1989                 finish_wait(&memcg_oom_waitq, &owait.wait);
1990                 mem_cgroup_out_of_memory(mem, mask);
1991         } else {
1992                 schedule();
1993                 finish_wait(&memcg_oom_waitq, &owait.wait);
1994         }
1995         spin_lock(&memcg_oom_lock);
1996         if (locked)
1997                 mem_cgroup_oom_unlock(mem);
1998         memcg_wakeup_oom(mem);
1999         spin_unlock(&memcg_oom_lock);
2000
2001         mem_cgroup_unmark_under_oom(mem);
2002
2003         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2004                 return false;
2005         /* Give chance to dying process */
2006         schedule_timeout(1);
2007         return true;
2008 }
2009
2010 /*
2011  * Currently used to update mapped file statistics, but the routine can be
2012  * generalized to update other statistics as well.
2013  *
2014  * Notes: Race condition
2015  *
2016  * We usually use page_cgroup_lock() for accessing page_cgroup member but
2017  * it tends to be costly. But considering some conditions, we doesn't need
2018  * to do so _always_.
2019  *
2020  * Considering "charge", lock_page_cgroup() is not required because all
2021  * file-stat operations happen after a page is attached to radix-tree. There
2022  * are no race with "charge".
2023  *
2024  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2025  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2026  * if there are race with "uncharge". Statistics itself is properly handled
2027  * by flags.
2028  *
2029  * Considering "move", this is an only case we see a race. To make the race
2030  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2031  * possibility of race condition. If there is, we take a lock.
2032  */
2033
2034 void mem_cgroup_update_page_stat(struct page *page,
2035                                  enum mem_cgroup_page_stat_item idx, int val)
2036 {
2037         struct mem_cgroup *mem;
2038         struct page_cgroup *pc = lookup_page_cgroup(page);
2039         bool need_unlock = false;
2040         unsigned long uninitialized_var(flags);
2041
2042         if (unlikely(!pc))
2043                 return;
2044
2045         rcu_read_lock();
2046         mem = pc->mem_cgroup;
2047         if (unlikely(!mem || !PageCgroupUsed(pc)))
2048                 goto out;
2049         /* pc->mem_cgroup is unstable ? */
2050         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
2051                 /* take a lock against to access pc->mem_cgroup */
2052                 move_lock_page_cgroup(pc, &flags);
2053                 need_unlock = true;
2054                 mem = pc->mem_cgroup;
2055                 if (!mem || !PageCgroupUsed(pc))
2056                         goto out;
2057         }
2058
2059         switch (idx) {
2060         case MEMCG_NR_FILE_MAPPED:
2061                 if (val > 0)
2062                         SetPageCgroupFileMapped(pc);
2063                 else if (!page_mapped(page))
2064                         ClearPageCgroupFileMapped(pc);
2065                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2066                 break;
2067         default:
2068                 BUG();
2069         }
2070
2071         this_cpu_add(mem->stat->count[idx], val);
2072
2073 out:
2074         if (unlikely(need_unlock))
2075                 move_unlock_page_cgroup(pc, &flags);
2076         rcu_read_unlock();
2077         return;
2078 }
2079 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2080
2081 /*
2082  * size of first charge trial. "32" comes from vmscan.c's magic value.
2083  * TODO: maybe necessary to use big numbers in big irons.
2084  */
2085 #define CHARGE_BATCH    32U
2086 struct memcg_stock_pcp {
2087         struct mem_cgroup *cached; /* this never be root cgroup */
2088         unsigned int nr_pages;
2089         struct work_struct work;
2090         unsigned long flags;
2091 #define FLUSHING_CACHED_CHARGE  (0)
2092 };
2093 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2094
2095 /*
2096  * Try to consume stocked charge on this cpu. If success, one page is consumed
2097  * from local stock and true is returned. If the stock is 0 or charges from a
2098  * cgroup which is not current target, returns false. This stock will be
2099  * refilled.
2100  */
2101 static bool consume_stock(struct mem_cgroup *mem)
2102 {
2103         struct memcg_stock_pcp *stock;
2104         bool ret = true;
2105
2106         stock = &get_cpu_var(memcg_stock);
2107         if (mem == stock->cached && stock->nr_pages)
2108                 stock->nr_pages--;
2109         else /* need to call res_counter_charge */
2110                 ret = false;
2111         put_cpu_var(memcg_stock);
2112         return ret;
2113 }
2114
2115 /*
2116  * Returns stocks cached in percpu to res_counter and reset cached information.
2117  */
2118 static void drain_stock(struct memcg_stock_pcp *stock)
2119 {
2120         struct mem_cgroup *old = stock->cached;
2121
2122         if (stock->nr_pages) {
2123                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2124
2125                 res_counter_uncharge(&old->res, bytes);
2126                 if (do_swap_account)
2127                         res_counter_uncharge(&old->memsw, bytes);
2128                 stock->nr_pages = 0;
2129         }
2130         stock->cached = NULL;
2131 }
2132
2133 /*
2134  * This must be called under preempt disabled or must be called by
2135  * a thread which is pinned to local cpu.
2136  */
2137 static void drain_local_stock(struct work_struct *dummy)
2138 {
2139         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2140         drain_stock(stock);
2141         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2142 }
2143
2144 /*
2145  * Cache charges(val) which is from res_counter, to local per_cpu area.
2146  * This will be consumed by consume_stock() function, later.
2147  */
2148 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2149 {
2150         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2151
2152         if (stock->cached != mem) { /* reset if necessary */
2153                 drain_stock(stock);
2154                 stock->cached = mem;
2155         }
2156         stock->nr_pages += nr_pages;
2157         put_cpu_var(memcg_stock);
2158 }
2159
2160 /*
2161  * Drains all per-CPU charge caches for given root_mem resp. subtree
2162  * of the hierarchy under it. sync flag says whether we should block
2163  * until the work is done.
2164  */
2165 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2166 {
2167         int cpu, curcpu;
2168
2169         /* Notify other cpus that system-wide "drain" is running */
2170         get_online_cpus();
2171         /*
2172          * Get a hint for avoiding draining charges on the current cpu,
2173          * which must be exhausted by our charging.  It is not required that
2174          * this be a precise check, so we use raw_smp_processor_id() instead of
2175          * getcpu()/putcpu().
2176          */
2177         curcpu = raw_smp_processor_id();
2178         for_each_online_cpu(cpu) {
2179                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2180                 struct mem_cgroup *mem;
2181
2182                 mem = stock->cached;
2183                 if (!mem || !stock->nr_pages)
2184                         continue;
2185                 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2186                         continue;
2187                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2188                         if (cpu == curcpu)
2189                                 drain_local_stock(&stock->work);
2190                         else
2191                                 schedule_work_on(cpu, &stock->work);
2192                 }
2193         }
2194
2195         if (!sync)
2196                 goto out;
2197
2198         for_each_online_cpu(cpu) {
2199                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2200                 if (mem_cgroup_same_or_subtree(root_mem, stock->cached) &&
2201                                 test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2202                         flush_work(&stock->work);
2203         }
2204 out:
2205         put_online_cpus();
2206 }
2207
2208 /*
2209  * Tries to drain stocked charges in other cpus. This function is asynchronous
2210  * and just put a work per cpu for draining localy on each cpu. Caller can
2211  * expects some charges will be back to res_counter later but cannot wait for
2212  * it.
2213  */
2214 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2215 {
2216         drain_all_stock(root_mem, false);
2217 }
2218
2219 /* This is a synchronous drain interface. */
2220 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2221 {
2222         /* called when force_empty is called */
2223         drain_all_stock(root_mem, true);
2224 }
2225
2226 /*
2227  * This function drains percpu counter value from DEAD cpu and
2228  * move it to local cpu. Note that this function can be preempted.
2229  */
2230 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2231 {
2232         int i;
2233
2234         spin_lock(&mem->pcp_counter_lock);
2235         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2236                 long x = per_cpu(mem->stat->count[i], cpu);
2237
2238                 per_cpu(mem->stat->count[i], cpu) = 0;
2239                 mem->nocpu_base.count[i] += x;
2240         }
2241         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2242                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2243
2244                 per_cpu(mem->stat->events[i], cpu) = 0;
2245                 mem->nocpu_base.events[i] += x;
2246         }
2247         /* need to clear ON_MOVE value, works as a kind of lock. */
2248         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2249         spin_unlock(&mem->pcp_counter_lock);
2250 }
2251
2252 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2253 {
2254         int idx = MEM_CGROUP_ON_MOVE;
2255
2256         spin_lock(&mem->pcp_counter_lock);
2257         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2258         spin_unlock(&mem->pcp_counter_lock);
2259 }
2260
2261 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2262                                         unsigned long action,
2263                                         void *hcpu)
2264 {
2265         int cpu = (unsigned long)hcpu;
2266         struct memcg_stock_pcp *stock;
2267         struct mem_cgroup *iter;
2268
2269         if ((action == CPU_ONLINE)) {
2270                 for_each_mem_cgroup_all(iter)
2271                         synchronize_mem_cgroup_on_move(iter, cpu);
2272                 return NOTIFY_OK;
2273         }
2274
2275         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2276                 return NOTIFY_OK;
2277
2278         for_each_mem_cgroup_all(iter)
2279                 mem_cgroup_drain_pcp_counter(iter, cpu);
2280
2281         stock = &per_cpu(memcg_stock, cpu);
2282         drain_stock(stock);
2283         return NOTIFY_OK;
2284 }
2285
2286
2287 /* See __mem_cgroup_try_charge() for details */
2288 enum {
2289         CHARGE_OK,              /* success */
2290         CHARGE_RETRY,           /* need to retry but retry is not bad */
2291         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2292         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2293         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2294 };
2295
2296 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2297                                 unsigned int nr_pages, bool oom_check)
2298 {
2299         unsigned long csize = nr_pages * PAGE_SIZE;
2300         struct mem_cgroup *mem_over_limit;
2301         struct res_counter *fail_res;
2302         unsigned long flags = 0;
2303         int ret;
2304
2305         ret = res_counter_charge(&mem->res, csize, &fail_res);
2306
2307         if (likely(!ret)) {
2308                 if (!do_swap_account)
2309                         return CHARGE_OK;
2310                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2311                 if (likely(!ret))
2312                         return CHARGE_OK;
2313
2314                 res_counter_uncharge(&mem->res, csize);
2315                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2316                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2317         } else
2318                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2319         /*
2320          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2321          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2322          *
2323          * Never reclaim on behalf of optional batching, retry with a
2324          * single page instead.
2325          */
2326         if (nr_pages == CHARGE_BATCH)
2327                 return CHARGE_RETRY;
2328
2329         if (!(gfp_mask & __GFP_WAIT))
2330                 return CHARGE_WOULDBLOCK;
2331
2332         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2333                                               gfp_mask, flags, NULL);
2334         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2335                 return CHARGE_RETRY;
2336         /*
2337          * Even though the limit is exceeded at this point, reclaim
2338          * may have been able to free some pages.  Retry the charge
2339          * before killing the task.
2340          *
2341          * Only for regular pages, though: huge pages are rather
2342          * unlikely to succeed so close to the limit, and we fall back
2343          * to regular pages anyway in case of failure.
2344          */
2345         if (nr_pages == 1 && ret)
2346                 return CHARGE_RETRY;
2347
2348         /*
2349          * At task move, charge accounts can be doubly counted. So, it's
2350          * better to wait until the end of task_move if something is going on.
2351          */
2352         if (mem_cgroup_wait_acct_move(mem_over_limit))
2353                 return CHARGE_RETRY;
2354
2355         /* If we don't need to call oom-killer at el, return immediately */
2356         if (!oom_check)
2357                 return CHARGE_NOMEM;
2358         /* check OOM */
2359         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2360                 return CHARGE_OOM_DIE;
2361
2362         return CHARGE_RETRY;
2363 }
2364
2365 /*
2366  * Unlike exported interface, "oom" parameter is added. if oom==true,
2367  * oom-killer can be invoked.
2368  */
2369 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2370                                    gfp_t gfp_mask,
2371                                    unsigned int nr_pages,
2372                                    struct mem_cgroup **memcg,
2373                                    bool oom)
2374 {
2375         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2376         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2377         struct mem_cgroup *mem = NULL;
2378         int ret;
2379
2380         /*
2381          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2382          * in system level. So, allow to go ahead dying process in addition to
2383          * MEMDIE process.
2384          */
2385         if (unlikely(test_thread_flag(TIF_MEMDIE)
2386                      || fatal_signal_pending(current)))
2387                 goto bypass;
2388
2389         /*
2390          * We always charge the cgroup the mm_struct belongs to.
2391          * The mm_struct's mem_cgroup changes on task migration if the
2392          * thread group leader migrates. It's possible that mm is not
2393          * set, if so charge the init_mm (happens for pagecache usage).
2394          */
2395         if (!*memcg && !mm)
2396                 goto bypass;
2397 again:
2398         if (*memcg) { /* css should be a valid one */
2399                 mem = *memcg;
2400                 VM_BUG_ON(css_is_removed(&mem->css));
2401                 if (mem_cgroup_is_root(mem))
2402                         goto done;
2403                 if (nr_pages == 1 && consume_stock(mem))
2404                         goto done;
2405                 css_get(&mem->css);
2406         } else {
2407                 struct task_struct *p;
2408
2409                 rcu_read_lock();
2410                 p = rcu_dereference(mm->owner);
2411                 /*
2412                  * Because we don't have task_lock(), "p" can exit.
2413                  * In that case, "mem" can point to root or p can be NULL with
2414                  * race with swapoff. Then, we have small risk of mis-accouning.
2415                  * But such kind of mis-account by race always happens because
2416                  * we don't have cgroup_mutex(). It's overkill and we allo that
2417                  * small race, here.
2418                  * (*) swapoff at el will charge against mm-struct not against
2419                  * task-struct. So, mm->owner can be NULL.
2420                  */
2421                 mem = mem_cgroup_from_task(p);
2422                 if (!mem || mem_cgroup_is_root(mem)) {
2423                         rcu_read_unlock();
2424                         goto done;
2425                 }
2426                 if (nr_pages == 1 && consume_stock(mem)) {
2427                         /*
2428                          * It seems dagerous to access memcg without css_get().
2429                          * But considering how consume_stok works, it's not
2430                          * necessary. If consume_stock success, some charges
2431                          * from this memcg are cached on this cpu. So, we
2432                          * don't need to call css_get()/css_tryget() before
2433                          * calling consume_stock().
2434                          */
2435                         rcu_read_unlock();
2436                         goto done;
2437                 }
2438                 /* after here, we may be blocked. we need to get refcnt */
2439                 if (!css_tryget(&mem->css)) {
2440                         rcu_read_unlock();
2441                         goto again;
2442                 }
2443                 rcu_read_unlock();
2444         }
2445
2446         do {
2447                 bool oom_check;
2448
2449                 /* If killed, bypass charge */
2450                 if (fatal_signal_pending(current)) {
2451                         css_put(&mem->css);
2452                         goto bypass;
2453                 }
2454
2455                 oom_check = false;
2456                 if (oom && !nr_oom_retries) {
2457                         oom_check = true;
2458                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2459                 }
2460
2461                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2462                 switch (ret) {
2463                 case CHARGE_OK:
2464                         break;
2465                 case CHARGE_RETRY: /* not in OOM situation but retry */
2466                         batch = nr_pages;
2467                         css_put(&mem->css);
2468                         mem = NULL;
2469                         goto again;
2470                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2471                         css_put(&mem->css);
2472                         goto nomem;
2473                 case CHARGE_NOMEM: /* OOM routine works */
2474                         if (!oom) {
2475                                 css_put(&mem->css);
2476                                 goto nomem;
2477                         }
2478                         /* If oom, we never return -ENOMEM */
2479                         nr_oom_retries--;
2480                         break;
2481                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2482                         css_put(&mem->css);
2483                         goto bypass;
2484                 }
2485         } while (ret != CHARGE_OK);
2486
2487         if (batch > nr_pages)
2488                 refill_stock(mem, batch - nr_pages);
2489         css_put(&mem->css);
2490 done:
2491         *memcg = mem;
2492         return 0;
2493 nomem:
2494         *memcg = NULL;
2495         return -ENOMEM;
2496 bypass:
2497         *memcg = NULL;
2498         return 0;
2499 }
2500
2501 /*
2502  * Somemtimes we have to undo a charge we got by try_charge().
2503  * This function is for that and do uncharge, put css's refcnt.
2504  * gotten by try_charge().
2505  */
2506 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2507                                        unsigned int nr_pages)
2508 {
2509         if (!mem_cgroup_is_root(mem)) {
2510                 unsigned long bytes = nr_pages * PAGE_SIZE;
2511
2512                 res_counter_uncharge(&mem->res, bytes);
2513                 if (do_swap_account)
2514                         res_counter_uncharge(&mem->memsw, bytes);
2515         }
2516 }
2517
2518 /*
2519  * A helper function to get mem_cgroup from ID. must be called under
2520  * rcu_read_lock(). The caller must check css_is_removed() or some if
2521  * it's concern. (dropping refcnt from swap can be called against removed
2522  * memcg.)
2523  */
2524 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2525 {
2526         struct cgroup_subsys_state *css;
2527
2528         /* ID 0 is unused ID */
2529         if (!id)
2530                 return NULL;
2531         css = css_lookup(&mem_cgroup_subsys, id);
2532         if (!css)
2533                 return NULL;
2534         return container_of(css, struct mem_cgroup, css);
2535 }
2536
2537 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2538 {
2539         struct mem_cgroup *mem = NULL;
2540         struct page_cgroup *pc;
2541         unsigned short id;
2542         swp_entry_t ent;
2543
2544         VM_BUG_ON(!PageLocked(page));
2545
2546         pc = lookup_page_cgroup(page);
2547         lock_page_cgroup(pc);
2548         if (PageCgroupUsed(pc)) {
2549                 mem = pc->mem_cgroup;
2550                 if (mem && !css_tryget(&mem->css))
2551                         mem = NULL;
2552         } else if (PageSwapCache(page)) {
2553                 ent.val = page_private(page);
2554                 id = lookup_swap_cgroup(ent);
2555                 rcu_read_lock();
2556                 mem = mem_cgroup_lookup(id);
2557                 if (mem && !css_tryget(&mem->css))
2558                         mem = NULL;
2559                 rcu_read_unlock();
2560         }
2561         unlock_page_cgroup(pc);
2562         return mem;
2563 }
2564
2565 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2566                                        struct page *page,
2567                                        unsigned int nr_pages,
2568                                        struct page_cgroup *pc,
2569                                        enum charge_type ctype)
2570 {
2571         lock_page_cgroup(pc);
2572         if (unlikely(PageCgroupUsed(pc))) {
2573                 unlock_page_cgroup(pc);
2574                 __mem_cgroup_cancel_charge(mem, nr_pages);
2575                 return;
2576         }
2577         /*
2578          * we don't need page_cgroup_lock about tail pages, becase they are not
2579          * accessed by any other context at this point.
2580          */
2581         pc->mem_cgroup = mem;
2582         /*
2583          * We access a page_cgroup asynchronously without lock_page_cgroup().
2584          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2585          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2586          * before USED bit, we need memory barrier here.
2587          * See mem_cgroup_add_lru_list(), etc.
2588          */
2589         smp_wmb();
2590         switch (ctype) {
2591         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2592         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2593                 SetPageCgroupCache(pc);
2594                 SetPageCgroupUsed(pc);
2595                 break;
2596         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2597                 ClearPageCgroupCache(pc);
2598                 SetPageCgroupUsed(pc);
2599                 break;
2600         default:
2601                 break;
2602         }
2603
2604         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2605         unlock_page_cgroup(pc);
2606         /*
2607          * "charge_statistics" updated event counter. Then, check it.
2608          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2609          * if they exceeds softlimit.
2610          */
2611         memcg_check_events(mem, page);
2612 }
2613
2614 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2615
2616 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2617                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2618 /*
2619  * Because tail pages are not marked as "used", set it. We're under
2620  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2621  */
2622 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2623 {
2624         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2625         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2626         unsigned long flags;
2627
2628         if (mem_cgroup_disabled())
2629                 return;
2630         /*
2631          * We have no races with charge/uncharge but will have races with
2632          * page state accounting.
2633          */
2634         move_lock_page_cgroup(head_pc, &flags);
2635
2636         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2637         smp_wmb(); /* see __commit_charge() */
2638         if (PageCgroupAcctLRU(head_pc)) {
2639                 enum lru_list lru;
2640                 struct mem_cgroup_per_zone *mz;
2641
2642                 /*
2643                  * LRU flags cannot be copied because we need to add tail
2644                  *.page to LRU by generic call and our hook will be called.
2645                  * We hold lru_lock, then, reduce counter directly.
2646                  */
2647                 lru = page_lru(head);
2648                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2649                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2650         }
2651         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2652         move_unlock_page_cgroup(head_pc, &flags);
2653 }
2654 #endif
2655
2656 /**
2657  * mem_cgroup_move_account - move account of the page
2658  * @page: the page
2659  * @nr_pages: number of regular pages (>1 for huge pages)
2660  * @pc: page_cgroup of the page.
2661  * @from: mem_cgroup which the page is moved from.
2662  * @to: mem_cgroup which the page is moved to. @from != @to.
2663  * @uncharge: whether we should call uncharge and css_put against @from.
2664  *
2665  * The caller must confirm following.
2666  * - page is not on LRU (isolate_page() is useful.)
2667  * - compound_lock is held when nr_pages > 1
2668  *
2669  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2670  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2671  * true, this function does "uncharge" from old cgroup, but it doesn't if
2672  * @uncharge is false, so a caller should do "uncharge".
2673  */
2674 static int mem_cgroup_move_account(struct page *page,
2675                                    unsigned int nr_pages,
2676                                    struct page_cgroup *pc,
2677                                    struct mem_cgroup *from,
2678                                    struct mem_cgroup *to,
2679                                    bool uncharge)
2680 {
2681         unsigned long flags;
2682         int ret;
2683
2684         VM_BUG_ON(from == to);
2685         VM_BUG_ON(PageLRU(page));
2686         /*
2687          * The page is isolated from LRU. So, collapse function
2688          * will not handle this page. But page splitting can happen.
2689          * Do this check under compound_page_lock(). The caller should
2690          * hold it.
2691          */
2692         ret = -EBUSY;
2693         if (nr_pages > 1 && !PageTransHuge(page))
2694                 goto out;
2695
2696         lock_page_cgroup(pc);
2697
2698         ret = -EINVAL;
2699         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2700                 goto unlock;
2701
2702         move_lock_page_cgroup(pc, &flags);
2703
2704         if (PageCgroupFileMapped(pc)) {
2705                 /* Update mapped_file data for mem_cgroup */
2706                 preempt_disable();
2707                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2708                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2709                 preempt_enable();
2710         }
2711         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2712         if (uncharge)
2713                 /* This is not "cancel", but cancel_charge does all we need. */
2714                 __mem_cgroup_cancel_charge(from, nr_pages);
2715
2716         /* caller should have done css_get */
2717         pc->mem_cgroup = to;
2718         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2719         /*
2720          * We charges against "to" which may not have any tasks. Then, "to"
2721          * can be under rmdir(). But in current implementation, caller of
2722          * this function is just force_empty() and move charge, so it's
2723          * guaranteed that "to" is never removed. So, we don't check rmdir
2724          * status here.
2725          */
2726         move_unlock_page_cgroup(pc, &flags);
2727         ret = 0;
2728 unlock:
2729         unlock_page_cgroup(pc);
2730         /*
2731          * check events
2732          */
2733         memcg_check_events(to, page);
2734         memcg_check_events(from, page);
2735 out:
2736         return ret;
2737 }
2738
2739 /*
2740  * move charges to its parent.
2741  */
2742
2743 static int mem_cgroup_move_parent(struct page *page,
2744                                   struct page_cgroup *pc,
2745                                   struct mem_cgroup *child,
2746                                   gfp_t gfp_mask)
2747 {
2748         struct cgroup *cg = child->css.cgroup;
2749         struct cgroup *pcg = cg->parent;
2750         struct mem_cgroup *parent;
2751         unsigned int nr_pages;
2752         unsigned long uninitialized_var(flags);
2753         int ret;
2754
2755         /* Is ROOT ? */
2756         if (!pcg)
2757                 return -EINVAL;
2758
2759         ret = -EBUSY;
2760         if (!get_page_unless_zero(page))
2761                 goto out;
2762         if (isolate_lru_page(page))
2763                 goto put;
2764
2765         nr_pages = hpage_nr_pages(page);
2766
2767         parent = mem_cgroup_from_cont(pcg);
2768         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2769         if (ret || !parent)
2770                 goto put_back;
2771
2772         if (nr_pages > 1)
2773                 flags = compound_lock_irqsave(page);
2774
2775         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2776         if (ret)
2777                 __mem_cgroup_cancel_charge(parent, nr_pages);
2778
2779         if (nr_pages > 1)
2780                 compound_unlock_irqrestore(page, flags);
2781 put_back:
2782         putback_lru_page(page);
2783 put:
2784         put_page(page);
2785 out:
2786         return ret;
2787 }
2788
2789 /*
2790  * Charge the memory controller for page usage.
2791  * Return
2792  * 0 if the charge was successful
2793  * < 0 if the cgroup is over its limit
2794  */
2795 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2796                                 gfp_t gfp_mask, enum charge_type ctype)
2797 {
2798         struct mem_cgroup *mem = NULL;
2799         unsigned int nr_pages = 1;
2800         struct page_cgroup *pc;
2801         bool oom = true;
2802         int ret;
2803
2804         if (PageTransHuge(page)) {
2805                 nr_pages <<= compound_order(page);
2806                 VM_BUG_ON(!PageTransHuge(page));
2807                 /*
2808                  * Never OOM-kill a process for a huge page.  The
2809                  * fault handler will fall back to regular pages.
2810                  */
2811                 oom = false;
2812         }
2813
2814         pc = lookup_page_cgroup(page);
2815         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2816
2817         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2818         if (ret || !mem)
2819                 return ret;
2820
2821         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2822         return 0;
2823 }
2824
2825 int mem_cgroup_newpage_charge(struct page *page,
2826                               struct mm_struct *mm, gfp_t gfp_mask)
2827 {
2828         if (mem_cgroup_disabled())
2829                 return 0;
2830         /*
2831          * If already mapped, we don't have to account.
2832          * If page cache, page->mapping has address_space.
2833          * But page->mapping may have out-of-use anon_vma pointer,
2834          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2835          * is NULL.
2836          */
2837         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2838                 return 0;
2839         if (unlikely(!mm))
2840                 mm = &init_mm;
2841         return mem_cgroup_charge_common(page, mm, gfp_mask,
2842                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2843 }
2844
2845 static void
2846 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2847                                         enum charge_type ctype);
2848
2849 static void
2850 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2851                                         enum charge_type ctype)
2852 {
2853         struct page_cgroup *pc = lookup_page_cgroup(page);
2854         /*
2855          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2856          * is already on LRU. It means the page may on some other page_cgroup's
2857          * LRU. Take care of it.
2858          */
2859         mem_cgroup_lru_del_before_commit(page);
2860         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2861         mem_cgroup_lru_add_after_commit(page);
2862         return;
2863 }
2864
2865 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2866                                 gfp_t gfp_mask)
2867 {
2868         struct mem_cgroup *mem = NULL;
2869         int ret;
2870
2871         if (mem_cgroup_disabled())
2872                 return 0;
2873         if (PageCompound(page))
2874                 return 0;
2875
2876         if (unlikely(!mm))
2877                 mm = &init_mm;
2878
2879         if (page_is_file_cache(page)) {
2880                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2881                 if (ret || !mem)
2882                         return ret;
2883
2884                 /*
2885                  * FUSE reuses pages without going through the final
2886                  * put that would remove them from the LRU list, make
2887                  * sure that they get relinked properly.
2888                  */
2889                 __mem_cgroup_commit_charge_lrucare(page, mem,
2890                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2891                 return ret;
2892         }
2893         /* shmem */
2894         if (PageSwapCache(page)) {
2895                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2896                 if (!ret)
2897                         __mem_cgroup_commit_charge_swapin(page, mem,
2898                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2899         } else
2900                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2901                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2902
2903         return ret;
2904 }
2905
2906 /*
2907  * While swap-in, try_charge -> commit or cancel, the page is locked.
2908  * And when try_charge() successfully returns, one refcnt to memcg without
2909  * struct page_cgroup is acquired. This refcnt will be consumed by
2910  * "commit()" or removed by "cancel()"
2911  */
2912 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2913                                  struct page *page,
2914                                  gfp_t mask, struct mem_cgroup **ptr)
2915 {
2916         struct mem_cgroup *mem;
2917         int ret;
2918
2919         *ptr = NULL;
2920
2921         if (mem_cgroup_disabled())
2922                 return 0;
2923
2924         if (!do_swap_account)
2925                 goto charge_cur_mm;
2926         /*
2927          * A racing thread's fault, or swapoff, may have already updated
2928          * the pte, and even removed page from swap cache: in those cases
2929          * do_swap_page()'s pte_same() test will fail; but there's also a
2930          * KSM case which does need to charge the page.
2931          */
2932         if (!PageSwapCache(page))
2933                 goto charge_cur_mm;
2934         mem = try_get_mem_cgroup_from_page(page);
2935         if (!mem)
2936                 goto charge_cur_mm;
2937         *ptr = mem;
2938         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2939         css_put(&mem->css);
2940         return ret;
2941 charge_cur_mm:
2942         if (unlikely(!mm))
2943                 mm = &init_mm;
2944         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2945 }
2946
2947 static void
2948 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2949                                         enum charge_type ctype)
2950 {
2951         if (mem_cgroup_disabled())
2952                 return;
2953         if (!ptr)
2954                 return;
2955         cgroup_exclude_rmdir(&ptr->css);
2956
2957         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2958         /*
2959          * Now swap is on-memory. This means this page may be
2960          * counted both as mem and swap....double count.
2961          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2962          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2963          * may call delete_from_swap_cache() before reach here.
2964          */
2965         if (do_swap_account && PageSwapCache(page)) {
2966                 swp_entry_t ent = {.val = page_private(page)};
2967                 unsigned short id;
2968                 struct mem_cgroup *memcg;
2969
2970                 id = swap_cgroup_record(ent, 0);
2971                 rcu_read_lock();
2972                 memcg = mem_cgroup_lookup(id);
2973                 if (memcg) {
2974                         /*
2975                          * This recorded memcg can be obsolete one. So, avoid
2976                          * calling css_tryget
2977                          */
2978                         if (!mem_cgroup_is_root(memcg))
2979                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2980                         mem_cgroup_swap_statistics(memcg, false);
2981                         mem_cgroup_put(memcg);
2982                 }
2983                 rcu_read_unlock();
2984         }
2985         /*
2986          * At swapin, we may charge account against cgroup which has no tasks.
2987          * So, rmdir()->pre_destroy() can be called while we do this charge.
2988          * In that case, we need to call pre_destroy() again. check it here.
2989          */
2990         cgroup_release_and_wakeup_rmdir(&ptr->css);
2991 }
2992
2993 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2994 {
2995         __mem_cgroup_commit_charge_swapin(page, ptr,
2996                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2997 }
2998
2999 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3000 {
3001         if (mem_cgroup_disabled())
3002                 return;
3003         if (!mem)
3004                 return;
3005         __mem_cgroup_cancel_charge(mem, 1);
3006 }
3007
3008 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3009                                    unsigned int nr_pages,
3010                                    const enum charge_type ctype)
3011 {
3012         struct memcg_batch_info *batch = NULL;
3013         bool uncharge_memsw = true;
3014
3015         /* If swapout, usage of swap doesn't decrease */
3016         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3017                 uncharge_memsw = false;
3018
3019         batch = &current->memcg_batch;
3020         /*
3021          * In usual, we do css_get() when we remember memcg pointer.
3022          * But in this case, we keep res->usage until end of a series of
3023          * uncharges. Then, it's ok to ignore memcg's refcnt.
3024          */
3025         if (!batch->memcg)
3026                 batch->memcg = mem;
3027         /*
3028          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3029          * In those cases, all pages freed continuously can be expected to be in
3030          * the same cgroup and we have chance to coalesce uncharges.
3031          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3032          * because we want to do uncharge as soon as possible.
3033          */
3034
3035         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3036                 goto direct_uncharge;
3037
3038         if (nr_pages > 1)
3039                 goto direct_uncharge;
3040
3041         /*
3042          * In typical case, batch->memcg == mem. This means we can
3043          * merge a series of uncharges to an uncharge of res_counter.
3044          * If not, we uncharge res_counter ony by one.
3045          */
3046         if (batch->memcg != mem)
3047                 goto direct_uncharge;
3048         /* remember freed charge and uncharge it later */
3049         batch->nr_pages++;
3050         if (uncharge_memsw)
3051                 batch->memsw_nr_pages++;
3052         return;
3053 direct_uncharge:
3054         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3055         if (uncharge_memsw)
3056                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3057         if (unlikely(batch->memcg != mem))
3058                 memcg_oom_recover(mem);
3059         return;
3060 }
3061
3062 /*
3063  * uncharge if !page_mapped(page)
3064  */
3065 static struct mem_cgroup *
3066 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3067 {
3068         struct mem_cgroup *mem = NULL;
3069         unsigned int nr_pages = 1;
3070         struct page_cgroup *pc;
3071
3072         if (mem_cgroup_disabled())
3073                 return NULL;
3074
3075         if (PageSwapCache(page))
3076                 return NULL;
3077
3078         if (PageTransHuge(page)) {
3079                 nr_pages <<= compound_order(page);
3080                 VM_BUG_ON(!PageTransHuge(page));
3081         }
3082         /*
3083          * Check if our page_cgroup is valid
3084          */
3085         pc = lookup_page_cgroup(page);
3086         if (unlikely(!pc || !PageCgroupUsed(pc)))
3087                 return NULL;
3088
3089         lock_page_cgroup(pc);
3090
3091         mem = pc->mem_cgroup;
3092
3093         if (!PageCgroupUsed(pc))
3094                 goto unlock_out;
3095
3096         switch (ctype) {
3097         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3098         case MEM_CGROUP_CHARGE_TYPE_DROP:
3099                 /* See mem_cgroup_prepare_migration() */
3100                 if (page_mapped(page) || PageCgroupMigration(pc))
3101                         goto unlock_out;
3102                 break;
3103         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3104                 if (!PageAnon(page)) {  /* Shared memory */
3105                         if (page->mapping && !page_is_file_cache(page))
3106                                 goto unlock_out;
3107                 } else if (page_mapped(page)) /* Anon */
3108                                 goto unlock_out;
3109                 break;
3110         default:
3111                 break;
3112         }
3113
3114         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3115
3116         ClearPageCgroupUsed(pc);
3117         /*
3118          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3119          * freed from LRU. This is safe because uncharged page is expected not
3120          * to be reused (freed soon). Exception is SwapCache, it's handled by
3121          * special functions.
3122          */
3123
3124         unlock_page_cgroup(pc);
3125         /*
3126          * even after unlock, we have mem->res.usage here and this memcg
3127          * will never be freed.
3128          */
3129         memcg_check_events(mem, page);
3130         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3131                 mem_cgroup_swap_statistics(mem, true);
3132                 mem_cgroup_get(mem);
3133         }
3134         if (!mem_cgroup_is_root(mem))
3135                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3136
3137         return mem;
3138
3139 unlock_out:
3140         unlock_page_cgroup(pc);
3141         return NULL;
3142 }
3143
3144 void mem_cgroup_uncharge_page(struct page *page)
3145 {
3146         /* early check. */
3147         if (page_mapped(page))
3148                 return;
3149         if (page->mapping && !PageAnon(page))
3150                 return;
3151         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3152 }
3153
3154 void mem_cgroup_uncharge_cache_page(struct page *page)
3155 {
3156         VM_BUG_ON(page_mapped(page));
3157         VM_BUG_ON(page->mapping);
3158         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3159 }
3160
3161 /*
3162  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3163  * In that cases, pages are freed continuously and we can expect pages
3164  * are in the same memcg. All these calls itself limits the number of
3165  * pages freed at once, then uncharge_start/end() is called properly.
3166  * This may be called prural(2) times in a context,
3167  */
3168
3169 void mem_cgroup_uncharge_start(void)
3170 {
3171         current->memcg_batch.do_batch++;
3172         /* We can do nest. */
3173         if (current->memcg_batch.do_batch == 1) {
3174                 current->memcg_batch.memcg = NULL;
3175                 current->memcg_batch.nr_pages = 0;
3176                 current->memcg_batch.memsw_nr_pages = 0;
3177         }
3178 }
3179
3180 void mem_cgroup_uncharge_end(void)
3181 {
3182         struct memcg_batch_info *batch = &current->memcg_batch;
3183
3184         if (!batch->do_batch)
3185                 return;
3186
3187         batch->do_batch--;
3188         if (batch->do_batch) /* If stacked, do nothing. */
3189                 return;
3190
3191         if (!batch->memcg)
3192                 return;
3193         /*
3194          * This "batch->memcg" is valid without any css_get/put etc...
3195          * bacause we hide charges behind us.
3196          */
3197         if (batch->nr_pages)
3198                 res_counter_uncharge(&batch->memcg->res,
3199                                      batch->nr_pages * PAGE_SIZE);
3200         if (batch->memsw_nr_pages)
3201                 res_counter_uncharge(&batch->memcg->memsw,
3202                                      batch->memsw_nr_pages * PAGE_SIZE);
3203         memcg_oom_recover(batch->memcg);
3204         /* forget this pointer (for sanity check) */
3205         batch->memcg = NULL;
3206 }
3207
3208 #ifdef CONFIG_SWAP
3209 /*
3210  * called after __delete_from_swap_cache() and drop "page" account.
3211  * memcg information is recorded to swap_cgroup of "ent"
3212  */
3213 void
3214 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3215 {
3216         struct mem_cgroup *memcg;
3217         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3218
3219         if (!swapout) /* this was a swap cache but the swap is unused ! */
3220                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3221
3222         memcg = __mem_cgroup_uncharge_common(page, ctype);
3223
3224         /*
3225          * record memcg information,  if swapout && memcg != NULL,
3226          * mem_cgroup_get() was called in uncharge().
3227          */
3228         if (do_swap_account && swapout && memcg)
3229                 swap_cgroup_record(ent, css_id(&memcg->css));
3230 }
3231 #endif
3232
3233 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3234 /*
3235  * called from swap_entry_free(). remove record in swap_cgroup and
3236  * uncharge "memsw" account.
3237  */
3238 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3239 {
3240         struct mem_cgroup *memcg;
3241         unsigned short id;
3242
3243         if (!do_swap_account)
3244                 return;
3245
3246         id = swap_cgroup_record(ent, 0);
3247         rcu_read_lock();
3248         memcg = mem_cgroup_lookup(id);
3249         if (memcg) {
3250                 /*
3251                  * We uncharge this because swap is freed.
3252                  * This memcg can be obsolete one. We avoid calling css_tryget
3253                  */
3254                 if (!mem_cgroup_is_root(memcg))
3255                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3256                 mem_cgroup_swap_statistics(memcg, false);
3257                 mem_cgroup_put(memcg);
3258         }
3259         rcu_read_unlock();
3260 }
3261
3262 /**
3263  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3264  * @entry: swap entry to be moved
3265  * @from:  mem_cgroup which the entry is moved from
3266  * @to:  mem_cgroup which the entry is moved to
3267  * @need_fixup: whether we should fixup res_counters and refcounts.
3268  *
3269  * It succeeds only when the swap_cgroup's record for this entry is the same
3270  * as the mem_cgroup's id of @from.
3271  *
3272  * Returns 0 on success, -EINVAL on failure.
3273  *
3274  * The caller must have charged to @to, IOW, called res_counter_charge() about
3275  * both res and memsw, and called css_get().
3276  */
3277 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3278                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3279 {
3280         unsigned short old_id, new_id;
3281
3282         old_id = css_id(&from->css);
3283         new_id = css_id(&to->css);
3284
3285         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3286                 mem_cgroup_swap_statistics(from, false);
3287                 mem_cgroup_swap_statistics(to, true);
3288                 /*
3289                  * This function is only called from task migration context now.
3290                  * It postpones res_counter and refcount handling till the end
3291                  * of task migration(mem_cgroup_clear_mc()) for performance
3292                  * improvement. But we cannot postpone mem_cgroup_get(to)
3293                  * because if the process that has been moved to @to does
3294                  * swap-in, the refcount of @to might be decreased to 0.
3295                  */
3296                 mem_cgroup_get(to);
3297                 if (need_fixup) {
3298                         if (!mem_cgroup_is_root(from))
3299                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3300                         mem_cgroup_put(from);
3301                         /*
3302                          * we charged both to->res and to->memsw, so we should
3303                          * uncharge to->res.
3304                          */
3305                         if (!mem_cgroup_is_root(to))
3306                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3307                 }
3308                 return 0;
3309         }
3310         return -EINVAL;
3311 }
3312 #else
3313 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3314                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3315 {
3316         return -EINVAL;
3317 }
3318 #endif
3319
3320 /*
3321  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3322  * page belongs to.
3323  */
3324 int mem_cgroup_prepare_migration(struct page *page,
3325         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3326 {
3327         struct mem_cgroup *mem = NULL;
3328         struct page_cgroup *pc;
3329         enum charge_type ctype;
3330         int ret = 0;
3331
3332         *ptr = NULL;
3333
3334         VM_BUG_ON(PageTransHuge(page));
3335         if (mem_cgroup_disabled())
3336                 return 0;
3337
3338         pc = lookup_page_cgroup(page);
3339         lock_page_cgroup(pc);
3340         if (PageCgroupUsed(pc)) {
3341                 mem = pc->mem_cgroup;
3342                 css_get(&mem->css);
3343                 /*
3344                  * At migrating an anonymous page, its mapcount goes down
3345                  * to 0 and uncharge() will be called. But, even if it's fully
3346                  * unmapped, migration may fail and this page has to be
3347                  * charged again. We set MIGRATION flag here and delay uncharge
3348                  * until end_migration() is called
3349                  *
3350                  * Corner Case Thinking
3351                  * A)
3352                  * When the old page was mapped as Anon and it's unmap-and-freed
3353                  * while migration was ongoing.
3354                  * If unmap finds the old page, uncharge() of it will be delayed
3355                  * until end_migration(). If unmap finds a new page, it's
3356                  * uncharged when it make mapcount to be 1->0. If unmap code
3357                  * finds swap_migration_entry, the new page will not be mapped
3358                  * and end_migration() will find it(mapcount==0).
3359                  *
3360                  * B)
3361                  * When the old page was mapped but migraion fails, the kernel
3362                  * remaps it. A charge for it is kept by MIGRATION flag even
3363                  * if mapcount goes down to 0. We can do remap successfully
3364                  * without charging it again.
3365                  *
3366                  * C)
3367                  * The "old" page is under lock_page() until the end of
3368                  * migration, so, the old page itself will not be swapped-out.
3369                  * If the new page is swapped out before end_migraton, our
3370                  * hook to usual swap-out path will catch the event.
3371                  */
3372                 if (PageAnon(page))
3373                         SetPageCgroupMigration(pc);
3374         }
3375         unlock_page_cgroup(pc);
3376         /*
3377          * If the page is not charged at this point,
3378          * we return here.
3379          */
3380         if (!mem)
3381                 return 0;
3382
3383         *ptr = mem;
3384         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3385         css_put(&mem->css);/* drop extra refcnt */
3386         if (ret || *ptr == NULL) {
3387                 if (PageAnon(page)) {
3388                         lock_page_cgroup(pc);
3389                         ClearPageCgroupMigration(pc);
3390                         unlock_page_cgroup(pc);
3391                         /*
3392                          * The old page may be fully unmapped while we kept it.
3393                          */
3394                         mem_cgroup_uncharge_page(page);
3395                 }
3396                 return -ENOMEM;
3397         }
3398         /*
3399          * We charge new page before it's used/mapped. So, even if unlock_page()
3400          * is called before end_migration, we can catch all events on this new
3401          * page. In the case new page is migrated but not remapped, new page's
3402          * mapcount will be finally 0 and we call uncharge in end_migration().
3403          */
3404         pc = lookup_page_cgroup(newpage);
3405         if (PageAnon(page))
3406                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3407         else if (page_is_file_cache(page))
3408                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3409         else
3410                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3411         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3412         return ret;
3413 }
3414
3415 /* remove redundant charge if migration failed*/
3416 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3417         struct page *oldpage, struct page *newpage, bool migration_ok)
3418 {
3419         struct page *used, *unused;
3420         struct page_cgroup *pc;
3421
3422         if (!mem)
3423                 return;
3424         /* blocks rmdir() */
3425         cgroup_exclude_rmdir(&mem->css);
3426         if (!migration_ok) {
3427                 used = oldpage;
3428                 unused = newpage;
3429         } else {
3430                 used = newpage;
3431                 unused = oldpage;
3432         }
3433         /*
3434          * We disallowed uncharge of pages under migration because mapcount
3435          * of the page goes down to zero, temporarly.
3436          * Clear the flag and check the page should be charged.
3437          */
3438         pc = lookup_page_cgroup(oldpage);
3439         lock_page_cgroup(pc);
3440         ClearPageCgroupMigration(pc);
3441         unlock_page_cgroup(pc);
3442
3443         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3444
3445         /*
3446          * If a page is a file cache, radix-tree replacement is very atomic
3447          * and we can skip this check. When it was an Anon page, its mapcount
3448          * goes down to 0. But because we added MIGRATION flage, it's not
3449          * uncharged yet. There are several case but page->mapcount check
3450          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3451          * check. (see prepare_charge() also)
3452          */
3453         if (PageAnon(used))
3454                 mem_cgroup_uncharge_page(used);
3455         /*
3456          * At migration, we may charge account against cgroup which has no
3457          * tasks.
3458          * So, rmdir()->pre_destroy() can be called while we do this charge.
3459          * In that case, we need to call pre_destroy() again. check it here.
3460          */
3461         cgroup_release_and_wakeup_rmdir(&mem->css);
3462 }
3463
3464 #ifdef CONFIG_DEBUG_VM
3465 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3466 {
3467         struct page_cgroup *pc;
3468
3469         pc = lookup_page_cgroup(page);
3470         if (likely(pc) && PageCgroupUsed(pc))
3471                 return pc;
3472         return NULL;
3473 }
3474
3475 bool mem_cgroup_bad_page_check(struct page *page)
3476 {
3477         if (mem_cgroup_disabled())
3478                 return false;
3479
3480         return lookup_page_cgroup_used(page) != NULL;
3481 }
3482
3483 void mem_cgroup_print_bad_page(struct page *page)
3484 {
3485         struct page_cgroup *pc;
3486
3487         pc = lookup_page_cgroup_used(page);
3488         if (pc) {
3489                 int ret = -1;
3490                 char *path;
3491
3492                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3493                        pc, pc->flags, pc->mem_cgroup);
3494
3495                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3496                 if (path) {
3497                         rcu_read_lock();
3498                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3499                                                         path, PATH_MAX);
3500                         rcu_read_unlock();
3501                 }
3502
3503                 printk(KERN_CONT "(%s)\n",
3504                                 (ret < 0) ? "cannot get the path" : path);
3505                 kfree(path);
3506         }
3507 }
3508 #endif
3509
3510 static DEFINE_MUTEX(set_limit_mutex);
3511
3512 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3513                                 unsigned long long val)
3514 {
3515         int retry_count;
3516         u64 memswlimit, memlimit;
3517         int ret = 0;
3518         int children = mem_cgroup_count_children(memcg);
3519         u64 curusage, oldusage;
3520         int enlarge;
3521
3522         /*
3523          * For keeping hierarchical_reclaim simple, how long we should retry
3524          * is depends on callers. We set our retry-count to be function
3525          * of # of children which we should visit in this loop.
3526          */
3527         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3528
3529         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3530
3531         enlarge = 0;
3532         while (retry_count) {
3533                 if (signal_pending(current)) {
3534                         ret = -EINTR;
3535                         break;
3536                 }
3537                 /*
3538                  * Rather than hide all in some function, I do this in
3539                  * open coded manner. You see what this really does.
3540                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3541                  */
3542                 mutex_lock(&set_limit_mutex);
3543                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3544                 if (memswlimit < val) {
3545                         ret = -EINVAL;
3546                         mutex_unlock(&set_limit_mutex);
3547                         break;
3548                 }
3549
3550                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3551                 if (memlimit < val)
3552                         enlarge = 1;
3553
3554                 ret = res_counter_set_limit(&memcg->res, val);
3555                 if (!ret) {
3556                         if (memswlimit == val)
3557                                 memcg->memsw_is_minimum = true;
3558                         else
3559                                 memcg->memsw_is_minimum = false;
3560                 }
3561                 mutex_unlock(&set_limit_mutex);
3562
3563                 if (!ret)
3564                         break;
3565
3566                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3567                                                 MEM_CGROUP_RECLAIM_SHRINK,
3568                                                 NULL);
3569                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3570                 /* Usage is reduced ? */
3571                 if (curusage >= oldusage)
3572                         retry_count--;
3573                 else
3574                         oldusage = curusage;
3575         }
3576         if (!ret && enlarge)
3577                 memcg_oom_recover(memcg);
3578
3579         return ret;
3580 }
3581
3582 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3583                                         unsigned long long val)
3584 {
3585         int retry_count;
3586         u64 memlimit, memswlimit, oldusage, curusage;
3587         int children = mem_cgroup_count_children(memcg);
3588         int ret = -EBUSY;
3589         int enlarge = 0;
3590
3591         /* see mem_cgroup_resize_res_limit */
3592         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3593         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3594         while (retry_count) {
3595                 if (signal_pending(current)) {
3596                         ret = -EINTR;
3597                         break;
3598                 }
3599                 /*
3600                  * Rather than hide all in some function, I do this in
3601                  * open coded manner. You see what this really does.
3602                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3603                  */
3604                 mutex_lock(&set_limit_mutex);
3605                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3606                 if (memlimit > val) {
3607                         ret = -EINVAL;
3608                         mutex_unlock(&set_limit_mutex);
3609                         break;
3610                 }
3611                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3612                 if (memswlimit < val)
3613                         enlarge = 1;
3614                 ret = res_counter_set_limit(&memcg->memsw, val);
3615                 if (!ret) {
3616                         if (memlimit == val)
3617                                 memcg->memsw_is_minimum = true;
3618                         else
3619                                 memcg->memsw_is_minimum = false;
3620                 }
3621                 mutex_unlock(&set_limit_mutex);
3622
3623                 if (!ret)
3624                         break;
3625
3626                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3627                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3628                                                 MEM_CGROUP_RECLAIM_SHRINK,
3629                                                 NULL);
3630                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3631                 /* Usage is reduced ? */
3632                 if (curusage >= oldusage)
3633                         retry_count--;
3634                 else
3635                         oldusage = curusage;
3636         }
3637         if (!ret && enlarge)
3638                 memcg_oom_recover(memcg);
3639         return ret;
3640 }
3641
3642 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3643                                             gfp_t gfp_mask,
3644                                             unsigned long *total_scanned)
3645 {
3646         unsigned long nr_reclaimed = 0;
3647         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3648         unsigned long reclaimed;
3649         int loop = 0;
3650         struct mem_cgroup_tree_per_zone *mctz;
3651         unsigned long long excess;
3652         unsigned long nr_scanned;
3653
3654         if (order > 0)
3655                 return 0;
3656
3657         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3658         /*
3659          * This loop can run a while, specially if mem_cgroup's continuously
3660          * keep exceeding their soft limit and putting the system under
3661          * pressure
3662          */
3663         do {
3664                 if (next_mz)
3665                         mz = next_mz;
3666                 else
3667                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3668                 if (!mz)
3669                         break;
3670
3671                 nr_scanned = 0;
3672                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3673                                                 gfp_mask,
3674                                                 MEM_CGROUP_RECLAIM_SOFT,
3675                                                 &nr_scanned);
3676                 nr_reclaimed += reclaimed;
3677                 *total_scanned += nr_scanned;
3678                 spin_lock(&mctz->lock);
3679
3680                 /*
3681                  * If we failed to reclaim anything from this memory cgroup
3682                  * it is time to move on to the next cgroup
3683                  */
3684                 next_mz = NULL;
3685                 if (!reclaimed) {
3686                         do {
3687                                 /*
3688                                  * Loop until we find yet another one.
3689                                  *
3690                                  * By the time we get the soft_limit lock
3691                                  * again, someone might have aded the
3692                                  * group back on the RB tree. Iterate to
3693                                  * make sure we get a different mem.
3694                                  * mem_cgroup_largest_soft_limit_node returns
3695                                  * NULL if no other cgroup is present on
3696                                  * the tree
3697                                  */
3698                                 next_mz =
3699                                 __mem_cgroup_largest_soft_limit_node(mctz);
3700                                 if (next_mz == mz)
3701                                         css_put(&next_mz->mem->css);
3702                                 else /* next_mz == NULL or other memcg */
3703                                         break;
3704                         } while (1);
3705                 }
3706                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3707                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3708                 /*
3709                  * One school of thought says that we should not add
3710                  * back the node to the tree if reclaim returns 0.
3711                  * But our reclaim could return 0, simply because due
3712                  * to priority we are exposing a smaller subset of
3713                  * memory to reclaim from. Consider this as a longer
3714                  * term TODO.
3715                  */
3716                 /* If excess == 0, no tree ops */
3717                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3718                 spin_unlock(&mctz->lock);
3719                 css_put(&mz->mem->css);
3720                 loop++;
3721                 /*
3722                  * Could not reclaim anything and there are no more
3723                  * mem cgroups to try or we seem to be looping without
3724                  * reclaiming anything.
3725                  */
3726                 if (!nr_reclaimed &&
3727                         (next_mz == NULL ||
3728                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3729                         break;
3730         } while (!nr_reclaimed);
3731         if (next_mz)
3732                 css_put(&next_mz->mem->css);
3733         return nr_reclaimed;
3734 }
3735
3736 /*
3737  * This routine traverse page_cgroup in given list and drop them all.
3738  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3739  */
3740 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3741                                 int node, int zid, enum lru_list lru)
3742 {
3743         struct zone *zone;
3744         struct mem_cgroup_per_zone *mz;
3745         struct page_cgroup *pc, *busy;
3746         unsigned long flags, loop;
3747         struct list_head *list;
3748         int ret = 0;
3749
3750         zone = &NODE_DATA(node)->node_zones[zid];
3751         mz = mem_cgroup_zoneinfo(mem, node, zid);
3752         list = &mz->lists[lru];
3753
3754         loop = MEM_CGROUP_ZSTAT(mz, lru);
3755         /* give some margin against EBUSY etc...*/
3756         loop += 256;
3757         busy = NULL;
3758         while (loop--) {
3759                 struct page *page;
3760
3761                 ret = 0;
3762                 spin_lock_irqsave(&zone->lru_lock, flags);
3763                 if (list_empty(list)) {
3764                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3765                         break;
3766                 }
3767                 pc = list_entry(list->prev, struct page_cgroup, lru);
3768                 if (busy == pc) {
3769                         list_move(&pc->lru, list);
3770                         busy = NULL;
3771                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3772                         continue;
3773                 }
3774                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3775
3776                 page = lookup_cgroup_page(pc);
3777
3778                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3779                 if (ret == -ENOMEM)
3780                         break;
3781
3782                 if (ret == -EBUSY || ret == -EINVAL) {
3783                         /* found lock contention or "pc" is obsolete. */
3784                         busy = pc;
3785                         cond_resched();
3786                 } else
3787                         busy = NULL;
3788         }
3789
3790         if (!ret && !list_empty(list))
3791                 return -EBUSY;
3792         return ret;
3793 }
3794
3795 /*
3796  * make mem_cgroup's charge to be 0 if there is no task.
3797  * This enables deleting this mem_cgroup.
3798  */
3799 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3800 {
3801         int ret;
3802         int node, zid, shrink;
3803         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3804         struct cgroup *cgrp = mem->css.cgroup;
3805
3806         css_get(&mem->css);
3807
3808         shrink = 0;
3809         /* should free all ? */
3810         if (free_all)
3811                 goto try_to_free;
3812 move_account:
3813         do {
3814                 ret = -EBUSY;
3815                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3816                         goto out;
3817                 ret = -EINTR;
3818                 if (signal_pending(current))
3819                         goto out;
3820                 /* This is for making all *used* pages to be on LRU. */
3821                 lru_add_drain_all();
3822                 drain_all_stock_sync(mem);
3823                 ret = 0;
3824                 mem_cgroup_start_move(mem);
3825                 for_each_node_state(node, N_HIGH_MEMORY) {
3826                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3827                                 enum lru_list l;
3828                                 for_each_lru(l) {
3829                                         ret = mem_cgroup_force_empty_list(mem,
3830                                                         node, zid, l);
3831                                         if (ret)
3832                                                 break;
3833                                 }
3834                         }
3835                         if (ret)
3836                                 break;
3837                 }
3838                 mem_cgroup_end_move(mem);
3839                 memcg_oom_recover(mem);
3840                 /* it seems parent cgroup doesn't have enough mem */
3841                 if (ret == -ENOMEM)
3842                         goto try_to_free;
3843                 cond_resched();
3844         /* "ret" should also be checked to ensure all lists are empty. */
3845         } while (mem->res.usage > 0 || ret);
3846 out:
3847         css_put(&mem->css);
3848         return ret;
3849
3850 try_to_free:
3851         /* returns EBUSY if there is a task or if we come here twice. */
3852         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3853                 ret = -EBUSY;
3854                 goto out;
3855         }
3856         /* we call try-to-free pages for make this cgroup empty */
3857         lru_add_drain_all();
3858         /* try to free all pages in this cgroup */
3859         shrink = 1;
3860         while (nr_retries && mem->res.usage > 0) {
3861                 struct memcg_scanrecord rec;
3862                 int progress;
3863
3864                 if (signal_pending(current)) {
3865                         ret = -EINTR;
3866                         goto out;
3867                 }
3868                 rec.context = SCAN_BY_SHRINK;
3869                 rec.mem = mem;
3870                 rec.root = mem;
3871                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3872                                                 false, &rec);
3873                 if (!progress) {
3874                         nr_retries--;
3875                         /* maybe some writeback is necessary */
3876                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3877                 }
3878
3879         }
3880         lru_add_drain();
3881         /* try move_account...there may be some *locked* pages. */
3882         goto move_account;
3883 }
3884
3885 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3886 {
3887         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3888 }
3889
3890
3891 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3892 {
3893         return mem_cgroup_from_cont(cont)->use_hierarchy;
3894 }
3895
3896 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3897                                         u64 val)
3898 {
3899         int retval = 0;
3900         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3901         struct cgroup *parent = cont->parent;
3902         struct mem_cgroup *parent_mem = NULL;
3903
3904         if (parent)
3905                 parent_mem = mem_cgroup_from_cont(parent);
3906
3907         cgroup_lock();
3908         /*
3909          * If parent's use_hierarchy is set, we can't make any modifications
3910          * in the child subtrees. If it is unset, then the change can
3911          * occur, provided the current cgroup has no children.
3912          *
3913          * For the root cgroup, parent_mem is NULL, we allow value to be
3914          * set if there are no children.
3915          */
3916         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3917                                 (val == 1 || val == 0)) {
3918                 if (list_empty(&cont->children))
3919                         mem->use_hierarchy = val;
3920                 else
3921                         retval = -EBUSY;
3922         } else
3923                 retval = -EINVAL;
3924         cgroup_unlock();
3925
3926         return retval;
3927 }
3928
3929
3930 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3931                                                enum mem_cgroup_stat_index idx)
3932 {
3933         struct mem_cgroup *iter;
3934         long val = 0;
3935
3936         /* Per-cpu values can be negative, use a signed accumulator */
3937         for_each_mem_cgroup_tree(iter, mem)
3938                 val += mem_cgroup_read_stat(iter, idx);
3939
3940         if (val < 0) /* race ? */
3941                 val = 0;
3942         return val;
3943 }
3944
3945 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3946 {
3947         u64 val;
3948
3949         if (!mem_cgroup_is_root(mem)) {
3950                 if (!swap)
3951                         return res_counter_read_u64(&mem->res, RES_USAGE);
3952                 else
3953                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3954         }
3955
3956         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3957         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3958
3959         if (swap)
3960                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3961
3962         return val << PAGE_SHIFT;
3963 }
3964
3965 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3966 {
3967         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3968         u64 val;
3969         int type, name;
3970
3971         type = MEMFILE_TYPE(cft->private);
3972         name = MEMFILE_ATTR(cft->private);
3973         switch (type) {
3974         case _MEM:
3975                 if (name == RES_USAGE)
3976                         val = mem_cgroup_usage(mem, false);
3977                 else
3978                         val = res_counter_read_u64(&mem->res, name);
3979                 break;
3980         case _MEMSWAP:
3981                 if (name == RES_USAGE)
3982                         val = mem_cgroup_usage(mem, true);
3983                 else
3984                         val = res_counter_read_u64(&mem->memsw, name);
3985                 break;
3986         default:
3987                 BUG();
3988                 break;
3989         }
3990         return val;
3991 }
3992 /*
3993  * The user of this function is...
3994  * RES_LIMIT.
3995  */
3996 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3997                             const char *buffer)
3998 {
3999         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4000         int type, name;
4001         unsigned long long val;
4002         int ret;
4003
4004         type = MEMFILE_TYPE(cft->private);
4005         name = MEMFILE_ATTR(cft->private);
4006         switch (name) {
4007         case RES_LIMIT:
4008                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4009                         ret = -EINVAL;
4010                         break;
4011                 }
4012                 /* This function does all necessary parse...reuse it */
4013                 ret = res_counter_memparse_write_strategy(buffer, &val);
4014                 if (ret)
4015                         break;
4016                 if (type == _MEM)
4017                         ret = mem_cgroup_resize_limit(memcg, val);
4018                 else
4019                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
4020                 break;
4021         case RES_SOFT_LIMIT:
4022                 ret = res_counter_memparse_write_strategy(buffer, &val);
4023                 if (ret)
4024                         break;
4025                 /*
4026                  * For memsw, soft limits are hard to implement in terms
4027                  * of semantics, for now, we support soft limits for
4028                  * control without swap
4029                  */
4030                 if (type == _MEM)
4031                         ret = res_counter_set_soft_limit(&memcg->res, val);
4032                 else
4033                         ret = -EINVAL;
4034                 break;
4035         default:
4036                 ret = -EINVAL; /* should be BUG() ? */
4037                 break;
4038         }
4039         return ret;
4040 }
4041
4042 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4043                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4044 {
4045         struct cgroup *cgroup;
4046         unsigned long long min_limit, min_memsw_limit, tmp;
4047
4048         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4049         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4050         cgroup = memcg->css.cgroup;
4051         if (!memcg->use_hierarchy)
4052                 goto out;
4053
4054         while (cgroup->parent) {
4055                 cgroup = cgroup->parent;
4056                 memcg = mem_cgroup_from_cont(cgroup);
4057                 if (!memcg->use_hierarchy)
4058                         break;
4059                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4060                 min_limit = min(min_limit, tmp);
4061                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4062                 min_memsw_limit = min(min_memsw_limit, tmp);
4063         }
4064 out:
4065         *mem_limit = min_limit;
4066         *memsw_limit = min_memsw_limit;
4067         return;
4068 }
4069
4070 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4071 {
4072         struct mem_cgroup *mem;
4073         int type, name;
4074
4075         mem = mem_cgroup_from_cont(cont);
4076         type = MEMFILE_TYPE(event);
4077         name = MEMFILE_ATTR(event);
4078         switch (name) {
4079         case RES_MAX_USAGE:
4080                 if (type == _MEM)
4081                         res_counter_reset_max(&mem->res);
4082                 else
4083                         res_counter_reset_max(&mem->memsw);
4084                 break;
4085         case RES_FAILCNT:
4086                 if (type == _MEM)
4087                         res_counter_reset_failcnt(&mem->res);
4088                 else
4089                         res_counter_reset_failcnt(&mem->memsw);
4090                 break;
4091         }
4092
4093         return 0;
4094 }
4095
4096 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4097                                         struct cftype *cft)
4098 {
4099         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4100 }
4101
4102 #ifdef CONFIG_MMU
4103 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4104                                         struct cftype *cft, u64 val)
4105 {
4106         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4107
4108         if (val >= (1 << NR_MOVE_TYPE))
4109                 return -EINVAL;
4110         /*
4111          * We check this value several times in both in can_attach() and
4112          * attach(), so we need cgroup lock to prevent this value from being
4113          * inconsistent.
4114          */
4115         cgroup_lock();
4116         mem->move_charge_at_immigrate = val;
4117         cgroup_unlock();
4118
4119         return 0;
4120 }
4121 #else
4122 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4123                                         struct cftype *cft, u64 val)
4124 {
4125         return -ENOSYS;
4126 }
4127 #endif
4128
4129
4130 /* For read statistics */
4131 enum {
4132         MCS_CACHE,
4133         MCS_RSS,
4134         MCS_FILE_MAPPED,
4135         MCS_PGPGIN,
4136         MCS_PGPGOUT,
4137         MCS_SWAP,
4138         MCS_PGFAULT,
4139         MCS_PGMAJFAULT,
4140         MCS_INACTIVE_ANON,
4141         MCS_ACTIVE_ANON,
4142         MCS_INACTIVE_FILE,
4143         MCS_ACTIVE_FILE,
4144         MCS_UNEVICTABLE,
4145         NR_MCS_STAT,
4146 };
4147
4148 struct mcs_total_stat {
4149         s64 stat[NR_MCS_STAT];
4150 };
4151
4152 struct {
4153         char *local_name;
4154         char *total_name;
4155 } memcg_stat_strings[NR_MCS_STAT] = {
4156         {"cache", "total_cache"},
4157         {"rss", "total_rss"},
4158         {"mapped_file", "total_mapped_file"},
4159         {"pgpgin", "total_pgpgin"},
4160         {"pgpgout", "total_pgpgout"},
4161         {"swap", "total_swap"},
4162         {"pgfault", "total_pgfault"},
4163         {"pgmajfault", "total_pgmajfault"},
4164         {"inactive_anon", "total_inactive_anon"},
4165         {"active_anon", "total_active_anon"},
4166         {"inactive_file", "total_inactive_file"},
4167         {"active_file", "total_active_file"},
4168         {"unevictable", "total_unevictable"}
4169 };
4170
4171
4172 static void
4173 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4174 {
4175         s64 val;
4176
4177         /* per cpu stat */
4178         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4179         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4180         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4181         s->stat[MCS_RSS] += val * PAGE_SIZE;
4182         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4183         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4184         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4185         s->stat[MCS_PGPGIN] += val;
4186         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4187         s->stat[MCS_PGPGOUT] += val;
4188         if (do_swap_account) {
4189                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4190                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4191         }
4192         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4193         s->stat[MCS_PGFAULT] += val;
4194         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4195         s->stat[MCS_PGMAJFAULT] += val;
4196
4197         /* per zone stat */
4198         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4199         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4200         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4201         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4202         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4203         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4204         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4205         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4206         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4207         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4208 }
4209
4210 static void
4211 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4212 {
4213         struct mem_cgroup *iter;
4214
4215         for_each_mem_cgroup_tree(iter, mem)
4216                 mem_cgroup_get_local_stat(iter, s);
4217 }
4218
4219 #ifdef CONFIG_NUMA
4220 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4221 {
4222         int nid;
4223         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4224         unsigned long node_nr;
4225         struct cgroup *cont = m->private;
4226         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4227
4228         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4229         seq_printf(m, "total=%lu", total_nr);
4230         for_each_node_state(nid, N_HIGH_MEMORY) {
4231                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4232                 seq_printf(m, " N%d=%lu", nid, node_nr);
4233         }
4234         seq_putc(m, '\n');
4235
4236         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4237         seq_printf(m, "file=%lu", file_nr);
4238         for_each_node_state(nid, N_HIGH_MEMORY) {
4239                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4240                                 LRU_ALL_FILE);
4241                 seq_printf(m, " N%d=%lu", nid, node_nr);
4242         }
4243         seq_putc(m, '\n');
4244
4245         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4246         seq_printf(m, "anon=%lu", anon_nr);
4247         for_each_node_state(nid, N_HIGH_MEMORY) {
4248                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4249                                 LRU_ALL_ANON);
4250                 seq_printf(m, " N%d=%lu", nid, node_nr);
4251         }
4252         seq_putc(m, '\n');
4253
4254         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4255         seq_printf(m, "unevictable=%lu", unevictable_nr);
4256         for_each_node_state(nid, N_HIGH_MEMORY) {
4257                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4258                                 BIT(LRU_UNEVICTABLE));
4259                 seq_printf(m, " N%d=%lu", nid, node_nr);
4260         }
4261         seq_putc(m, '\n');
4262         return 0;
4263 }
4264 #endif /* CONFIG_NUMA */
4265
4266 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4267                                  struct cgroup_map_cb *cb)
4268 {
4269         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4270         struct mcs_total_stat mystat;
4271         int i;
4272
4273         memset(&mystat, 0, sizeof(mystat));
4274         mem_cgroup_get_local_stat(mem_cont, &mystat);
4275
4276
4277         for (i = 0; i < NR_MCS_STAT; i++) {
4278                 if (i == MCS_SWAP && !do_swap_account)
4279                         continue;
4280                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4281         }
4282
4283         /* Hierarchical information */
4284         {
4285                 unsigned long long limit, memsw_limit;
4286                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4287                 cb->fill(cb, "hierarchical_memory_limit", limit);
4288                 if (do_swap_account)
4289                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4290         }
4291
4292         memset(&mystat, 0, sizeof(mystat));
4293         mem_cgroup_get_total_stat(mem_cont, &mystat);
4294         for (i = 0; i < NR_MCS_STAT; i++) {
4295                 if (i == MCS_SWAP && !do_swap_account)
4296                         continue;
4297                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4298         }
4299
4300 #ifdef CONFIG_DEBUG_VM
4301         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4302
4303         {
4304                 int nid, zid;
4305                 struct mem_cgroup_per_zone *mz;
4306                 unsigned long recent_rotated[2] = {0, 0};
4307                 unsigned long recent_scanned[2] = {0, 0};
4308
4309                 for_each_online_node(nid)
4310                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4311                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4312
4313                                 recent_rotated[0] +=
4314                                         mz->reclaim_stat.recent_rotated[0];
4315                                 recent_rotated[1] +=
4316                                         mz->reclaim_stat.recent_rotated[1];
4317                                 recent_scanned[0] +=
4318                                         mz->reclaim_stat.recent_scanned[0];
4319                                 recent_scanned[1] +=
4320                                         mz->reclaim_stat.recent_scanned[1];
4321                         }
4322                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4323                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4324                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4325                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4326         }
4327 #endif
4328
4329         return 0;
4330 }
4331
4332 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4333 {
4334         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4335
4336         return mem_cgroup_swappiness(memcg);
4337 }
4338
4339 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4340                                        u64 val)
4341 {
4342         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4343         struct mem_cgroup *parent;
4344
4345         if (val > 100)
4346                 return -EINVAL;
4347
4348         if (cgrp->parent == NULL)
4349                 return -EINVAL;
4350
4351         parent = mem_cgroup_from_cont(cgrp->parent);
4352
4353         cgroup_lock();
4354
4355         /* If under hierarchy, only empty-root can set this value */
4356         if ((parent->use_hierarchy) ||
4357             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4358                 cgroup_unlock();
4359                 return -EINVAL;
4360         }
4361
4362         memcg->swappiness = val;
4363
4364         cgroup_unlock();
4365
4366         return 0;
4367 }
4368
4369 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4370 {
4371         struct mem_cgroup_threshold_ary *t;
4372         u64 usage;
4373         int i;
4374
4375         rcu_read_lock();
4376         if (!swap)
4377                 t = rcu_dereference(memcg->thresholds.primary);
4378         else
4379                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4380
4381         if (!t)
4382                 goto unlock;
4383
4384         usage = mem_cgroup_usage(memcg, swap);
4385
4386         /*
4387          * current_threshold points to threshold just below usage.
4388          * If it's not true, a threshold was crossed after last
4389          * call of __mem_cgroup_threshold().
4390          */
4391         i = t->current_threshold;
4392
4393         /*
4394          * Iterate backward over array of thresholds starting from
4395          * current_threshold and check if a threshold is crossed.
4396          * If none of thresholds below usage is crossed, we read
4397          * only one element of the array here.
4398          */
4399         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4400                 eventfd_signal(t->entries[i].eventfd, 1);
4401
4402         /* i = current_threshold + 1 */
4403         i++;
4404
4405         /*
4406          * Iterate forward over array of thresholds starting from
4407          * current_threshold+1 and check if a threshold is crossed.
4408          * If none of thresholds above usage is crossed, we read
4409          * only one element of the array here.
4410          */
4411         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4412                 eventfd_signal(t->entries[i].eventfd, 1);
4413
4414         /* Update current_threshold */
4415         t->current_threshold = i - 1;
4416 unlock:
4417         rcu_read_unlock();
4418 }
4419
4420 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4421 {
4422         while (memcg) {
4423                 __mem_cgroup_threshold(memcg, false);
4424                 if (do_swap_account)
4425                         __mem_cgroup_threshold(memcg, true);
4426
4427                 memcg = parent_mem_cgroup(memcg);
4428         }
4429 }
4430
4431 static int compare_thresholds(const void *a, const void *b)
4432 {
4433         const struct mem_cgroup_threshold *_a = a;
4434         const struct mem_cgroup_threshold *_b = b;
4435
4436         return _a->threshold - _b->threshold;
4437 }
4438
4439 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4440 {
4441         struct mem_cgroup_eventfd_list *ev;
4442
4443         list_for_each_entry(ev, &mem->oom_notify, list)
4444                 eventfd_signal(ev->eventfd, 1);
4445         return 0;
4446 }
4447
4448 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4449 {
4450         struct mem_cgroup *iter;
4451
4452         for_each_mem_cgroup_tree(iter, mem)
4453                 mem_cgroup_oom_notify_cb(iter);
4454 }
4455
4456 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4457         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4458 {
4459         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4460         struct mem_cgroup_thresholds *thresholds;
4461         struct mem_cgroup_threshold_ary *new;
4462         int type = MEMFILE_TYPE(cft->private);
4463         u64 threshold, usage;
4464         int i, size, ret;
4465
4466         ret = res_counter_memparse_write_strategy(args, &threshold);
4467         if (ret)
4468                 return ret;
4469
4470         mutex_lock(&memcg->thresholds_lock);
4471
4472         if (type == _MEM)
4473                 thresholds = &memcg->thresholds;
4474         else if (type == _MEMSWAP)
4475                 thresholds = &memcg->memsw_thresholds;
4476         else
4477                 BUG();
4478
4479         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4480
4481         /* Check if a threshold crossed before adding a new one */
4482         if (thresholds->primary)
4483                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4484
4485         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4486
4487         /* Allocate memory for new array of thresholds */
4488         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4489                         GFP_KERNEL);
4490         if (!new) {
4491                 ret = -ENOMEM;
4492                 goto unlock;
4493         }
4494         new->size = size;
4495
4496         /* Copy thresholds (if any) to new array */
4497         if (thresholds->primary) {
4498                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4499                                 sizeof(struct mem_cgroup_threshold));
4500         }
4501
4502         /* Add new threshold */
4503         new->entries[size - 1].eventfd = eventfd;
4504         new->entries[size - 1].threshold = threshold;
4505
4506         /* Sort thresholds. Registering of new threshold isn't time-critical */
4507         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4508                         compare_thresholds, NULL);
4509
4510         /* Find current threshold */
4511         new->current_threshold = -1;
4512         for (i = 0; i < size; i++) {
4513                 if (new->entries[i].threshold < usage) {
4514                         /*
4515                          * new->current_threshold will not be used until
4516                          * rcu_assign_pointer(), so it's safe to increment
4517                          * it here.
4518                          */
4519                         ++new->current_threshold;
4520                 }
4521         }
4522
4523         /* Free old spare buffer and save old primary buffer as spare */
4524         kfree(thresholds->spare);
4525         thresholds->spare = thresholds->primary;
4526
4527         rcu_assign_pointer(thresholds->primary, new);
4528
4529         /* To be sure that nobody uses thresholds */
4530         synchronize_rcu();
4531
4532 unlock:
4533         mutex_unlock(&memcg->thresholds_lock);
4534
4535         return ret;
4536 }
4537
4538 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4539         struct cftype *cft, struct eventfd_ctx *eventfd)
4540 {
4541         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4542         struct mem_cgroup_thresholds *thresholds;
4543         struct mem_cgroup_threshold_ary *new;
4544         int type = MEMFILE_TYPE(cft->private);
4545         u64 usage;
4546         int i, j, size;
4547
4548         mutex_lock(&memcg->thresholds_lock);
4549         if (type == _MEM)
4550                 thresholds = &memcg->thresholds;
4551         else if (type == _MEMSWAP)
4552                 thresholds = &memcg->memsw_thresholds;
4553         else
4554                 BUG();
4555
4556         /*
4557          * Something went wrong if we trying to unregister a threshold
4558          * if we don't have thresholds
4559          */
4560         BUG_ON(!thresholds);
4561
4562         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4563
4564         /* Check if a threshold crossed before removing */
4565         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4566
4567         /* Calculate new number of threshold */
4568         size = 0;
4569         for (i = 0; i < thresholds->primary->size; i++) {
4570                 if (thresholds->primary->entries[i].eventfd != eventfd)
4571                         size++;
4572         }
4573
4574         new = thresholds->spare;
4575
4576         /* Set thresholds array to NULL if we don't have thresholds */
4577         if (!size) {
4578                 kfree(new);
4579                 new = NULL;
4580                 goto swap_buffers;
4581         }
4582
4583         new->size = size;
4584
4585         /* Copy thresholds and find current threshold */
4586         new->current_threshold = -1;
4587         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4588                 if (thresholds->primary->entries[i].eventfd == eventfd)
4589                         continue;
4590
4591                 new->entries[j] = thresholds->primary->entries[i];
4592                 if (new->entries[j].threshold < usage) {
4593                         /*
4594                          * new->current_threshold will not be used
4595                          * until rcu_assign_pointer(), so it's safe to increment
4596                          * it here.
4597                          */
4598                         ++new->current_threshold;
4599                 }
4600                 j++;
4601         }
4602
4603 swap_buffers:
4604         /* Swap primary and spare array */
4605         thresholds->spare = thresholds->primary;
4606         rcu_assign_pointer(thresholds->primary, new);
4607
4608         /* To be sure that nobody uses thresholds */
4609         synchronize_rcu();
4610
4611         mutex_unlock(&memcg->thresholds_lock);
4612 }
4613
4614 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4615         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4616 {
4617         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4618         struct mem_cgroup_eventfd_list *event;
4619         int type = MEMFILE_TYPE(cft->private);
4620
4621         BUG_ON(type != _OOM_TYPE);
4622         event = kmalloc(sizeof(*event), GFP_KERNEL);
4623         if (!event)
4624                 return -ENOMEM;
4625
4626         spin_lock(&memcg_oom_lock);
4627
4628         event->eventfd = eventfd;
4629         list_add(&event->list, &memcg->oom_notify);
4630
4631         /* already in OOM ? */
4632         if (atomic_read(&memcg->under_oom))
4633                 eventfd_signal(eventfd, 1);
4634         spin_unlock(&memcg_oom_lock);
4635
4636         return 0;
4637 }
4638
4639 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4640         struct cftype *cft, struct eventfd_ctx *eventfd)
4641 {
4642         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4643         struct mem_cgroup_eventfd_list *ev, *tmp;
4644         int type = MEMFILE_TYPE(cft->private);
4645
4646         BUG_ON(type != _OOM_TYPE);
4647
4648         spin_lock(&memcg_oom_lock);
4649
4650         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4651                 if (ev->eventfd == eventfd) {
4652                         list_del(&ev->list);
4653                         kfree(ev);
4654                 }
4655         }
4656
4657         spin_unlock(&memcg_oom_lock);
4658 }
4659
4660 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4661         struct cftype *cft,  struct cgroup_map_cb *cb)
4662 {
4663         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4664
4665         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4666
4667         if (atomic_read(&mem->under_oom))
4668                 cb->fill(cb, "under_oom", 1);
4669         else
4670                 cb->fill(cb, "under_oom", 0);
4671         return 0;
4672 }
4673
4674 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4675         struct cftype *cft, u64 val)
4676 {
4677         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4678         struct mem_cgroup *parent;
4679
4680         /* cannot set to root cgroup and only 0 and 1 are allowed */
4681         if (!cgrp->parent || !((val == 0) || (val == 1)))
4682                 return -EINVAL;
4683
4684         parent = mem_cgroup_from_cont(cgrp->parent);
4685
4686         cgroup_lock();
4687         /* oom-kill-disable is a flag for subhierarchy. */
4688         if ((parent->use_hierarchy) ||
4689             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4690                 cgroup_unlock();
4691                 return -EINVAL;
4692         }
4693         mem->oom_kill_disable = val;
4694         if (!val)
4695                 memcg_oom_recover(mem);
4696         cgroup_unlock();
4697         return 0;
4698 }
4699
4700 #ifdef CONFIG_NUMA
4701 static const struct file_operations mem_control_numa_stat_file_operations = {
4702         .read = seq_read,
4703         .llseek = seq_lseek,
4704         .release = single_release,
4705 };
4706
4707 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4708 {
4709         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4710
4711         file->f_op = &mem_control_numa_stat_file_operations;
4712         return single_open(file, mem_control_numa_stat_show, cont);
4713 }
4714 #endif /* CONFIG_NUMA */
4715
4716 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4717                                 struct cftype *cft,
4718                                 struct cgroup_map_cb *cb)
4719 {
4720         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4721         char string[64];
4722         int i;
4723
4724         for (i = 0; i < NR_SCANSTATS; i++) {
4725                 strcpy(string, scanstat_string[i]);
4726                 strcat(string, SCANSTAT_WORD_LIMIT);
4727                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4728         }
4729
4730         for (i = 0; i < NR_SCANSTATS; i++) {
4731                 strcpy(string, scanstat_string[i]);
4732                 strcat(string, SCANSTAT_WORD_SYSTEM);
4733                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4734         }
4735
4736         for (i = 0; i < NR_SCANSTATS; i++) {
4737                 strcpy(string, scanstat_string[i]);
4738                 strcat(string, SCANSTAT_WORD_LIMIT);
4739                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4740                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4741         }
4742         for (i = 0; i < NR_SCANSTATS; i++) {
4743                 strcpy(string, scanstat_string[i]);
4744                 strcat(string, SCANSTAT_WORD_SYSTEM);
4745                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4746                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4747         }
4748         return 0;
4749 }
4750
4751 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4752                                 unsigned int event)
4753 {
4754         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4755
4756         spin_lock(&mem->scanstat.lock);
4757         memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4758         memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4759         spin_unlock(&mem->scanstat.lock);
4760         return 0;
4761 }
4762
4763
4764 static struct cftype mem_cgroup_files[] = {
4765         {
4766                 .name = "usage_in_bytes",
4767                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4768                 .read_u64 = mem_cgroup_read,
4769                 .register_event = mem_cgroup_usage_register_event,
4770                 .unregister_event = mem_cgroup_usage_unregister_event,
4771         },
4772         {
4773                 .name = "max_usage_in_bytes",
4774                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4775                 .trigger = mem_cgroup_reset,
4776                 .read_u64 = mem_cgroup_read,
4777         },
4778         {
4779                 .name = "limit_in_bytes",
4780                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4781                 .write_string = mem_cgroup_write,
4782                 .read_u64 = mem_cgroup_read,
4783         },
4784         {
4785                 .name = "soft_limit_in_bytes",
4786                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4787                 .write_string = mem_cgroup_write,
4788                 .read_u64 = mem_cgroup_read,
4789         },
4790         {
4791                 .name = "failcnt",
4792                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4793                 .trigger = mem_cgroup_reset,
4794                 .read_u64 = mem_cgroup_read,
4795         },
4796         {
4797                 .name = "stat",
4798                 .read_map = mem_control_stat_show,
4799         },
4800         {
4801                 .name = "force_empty",
4802                 .trigger = mem_cgroup_force_empty_write,
4803         },
4804         {
4805                 .name = "use_hierarchy",
4806                 .write_u64 = mem_cgroup_hierarchy_write,
4807                 .read_u64 = mem_cgroup_hierarchy_read,
4808         },
4809         {
4810                 .name = "swappiness",
4811                 .read_u64 = mem_cgroup_swappiness_read,
4812                 .write_u64 = mem_cgroup_swappiness_write,
4813         },
4814         {
4815                 .name = "move_charge_at_immigrate",
4816                 .read_u64 = mem_cgroup_move_charge_read,
4817                 .write_u64 = mem_cgroup_move_charge_write,
4818         },
4819         {
4820                 .name = "oom_control",
4821                 .read_map = mem_cgroup_oom_control_read,
4822                 .write_u64 = mem_cgroup_oom_control_write,
4823                 .register_event = mem_cgroup_oom_register_event,
4824                 .unregister_event = mem_cgroup_oom_unregister_event,
4825                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4826         },
4827 #ifdef CONFIG_NUMA
4828         {
4829                 .name = "numa_stat",
4830                 .open = mem_control_numa_stat_open,
4831                 .mode = S_IRUGO,
4832         },
4833 #endif
4834         {
4835                 .name = "vmscan_stat",
4836                 .read_map = mem_cgroup_vmscan_stat_read,
4837                 .trigger = mem_cgroup_reset_vmscan_stat,
4838         },
4839 };
4840
4841 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4842 static struct cftype memsw_cgroup_files[] = {
4843         {
4844                 .name = "memsw.usage_in_bytes",
4845                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4846                 .read_u64 = mem_cgroup_read,
4847                 .register_event = mem_cgroup_usage_register_event,
4848                 .unregister_event = mem_cgroup_usage_unregister_event,
4849         },
4850         {
4851                 .name = "memsw.max_usage_in_bytes",
4852                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4853                 .trigger = mem_cgroup_reset,
4854                 .read_u64 = mem_cgroup_read,
4855         },
4856         {
4857                 .name = "memsw.limit_in_bytes",
4858                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4859                 .write_string = mem_cgroup_write,
4860                 .read_u64 = mem_cgroup_read,
4861         },
4862         {
4863                 .name = "memsw.failcnt",
4864                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4865                 .trigger = mem_cgroup_reset,
4866                 .read_u64 = mem_cgroup_read,
4867         },
4868 };
4869
4870 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4871 {
4872         if (!do_swap_account)
4873                 return 0;
4874         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4875                                 ARRAY_SIZE(memsw_cgroup_files));
4876 };
4877 #else
4878 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4879 {
4880         return 0;
4881 }
4882 #endif
4883
4884 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4885 {
4886         struct mem_cgroup_per_node *pn;
4887         struct mem_cgroup_per_zone *mz;
4888         enum lru_list l;
4889         int zone, tmp = node;
4890         /*
4891          * This routine is called against possible nodes.
4892          * But it's BUG to call kmalloc() against offline node.
4893          *
4894          * TODO: this routine can waste much memory for nodes which will
4895          *       never be onlined. It's better to use memory hotplug callback
4896          *       function.
4897          */
4898         if (!node_state(node, N_NORMAL_MEMORY))
4899                 tmp = -1;
4900         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4901         if (!pn)
4902                 return 1;
4903
4904         mem->info.nodeinfo[node] = pn;
4905         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4906                 mz = &pn->zoneinfo[zone];
4907                 for_each_lru(l)
4908                         INIT_LIST_HEAD(&mz->lists[l]);
4909                 mz->usage_in_excess = 0;
4910                 mz->on_tree = false;
4911                 mz->mem = mem;
4912         }
4913         return 0;
4914 }
4915
4916 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4917 {
4918         kfree(mem->info.nodeinfo[node]);
4919 }
4920
4921 static struct mem_cgroup *mem_cgroup_alloc(void)
4922 {
4923         struct mem_cgroup *mem;
4924         int size = sizeof(struct mem_cgroup);
4925
4926         /* Can be very big if MAX_NUMNODES is very big */
4927         if (size < PAGE_SIZE)
4928                 mem = kzalloc(size, GFP_KERNEL);
4929         else
4930                 mem = vzalloc(size);
4931
4932         if (!mem)
4933                 return NULL;
4934
4935         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4936         if (!mem->stat)
4937                 goto out_free;
4938         spin_lock_init(&mem->pcp_counter_lock);
4939         return mem;
4940
4941 out_free:
4942         if (size < PAGE_SIZE)
4943                 kfree(mem);
4944         else
4945                 vfree(mem);
4946         return NULL;
4947 }
4948
4949 /*
4950  * At destroying mem_cgroup, references from swap_cgroup can remain.
4951  * (scanning all at force_empty is too costly...)
4952  *
4953  * Instead of clearing all references at force_empty, we remember
4954  * the number of reference from swap_cgroup and free mem_cgroup when
4955  * it goes down to 0.
4956  *
4957  * Removal of cgroup itself succeeds regardless of refs from swap.
4958  */
4959
4960 static void __mem_cgroup_free(struct mem_cgroup *mem)
4961 {
4962         int node;
4963
4964         mem_cgroup_remove_from_trees(mem);
4965         free_css_id(&mem_cgroup_subsys, &mem->css);
4966
4967         for_each_node_state(node, N_POSSIBLE)
4968                 free_mem_cgroup_per_zone_info(mem, node);
4969
4970         free_percpu(mem->stat);
4971         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4972                 kfree(mem);
4973         else
4974                 vfree(mem);
4975 }
4976
4977 static void mem_cgroup_get(struct mem_cgroup *mem)
4978 {
4979         atomic_inc(&mem->refcnt);
4980 }
4981
4982 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4983 {
4984         if (atomic_sub_and_test(count, &mem->refcnt)) {
4985                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4986                 __mem_cgroup_free(mem);
4987                 if (parent)
4988                         mem_cgroup_put(parent);
4989         }
4990 }
4991
4992 static void mem_cgroup_put(struct mem_cgroup *mem)
4993 {
4994         __mem_cgroup_put(mem, 1);
4995 }
4996
4997 /*
4998  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4999  */
5000 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5001 {
5002         if (!mem->res.parent)
5003                 return NULL;
5004         return mem_cgroup_from_res_counter(mem->res.parent, res);
5005 }
5006
5007 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5008 static void __init enable_swap_cgroup(void)
5009 {
5010         if (!mem_cgroup_disabled() && really_do_swap_account)
5011                 do_swap_account = 1;
5012 }
5013 #else
5014 static void __init enable_swap_cgroup(void)
5015 {
5016 }
5017 #endif
5018
5019 static int mem_cgroup_soft_limit_tree_init(void)
5020 {
5021         struct mem_cgroup_tree_per_node *rtpn;
5022         struct mem_cgroup_tree_per_zone *rtpz;
5023         int tmp, node, zone;
5024
5025         for_each_node_state(node, N_POSSIBLE) {
5026                 tmp = node;
5027                 if (!node_state(node, N_NORMAL_MEMORY))
5028                         tmp = -1;
5029                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5030                 if (!rtpn)
5031                         return 1;
5032
5033                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5034
5035                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5036                         rtpz = &rtpn->rb_tree_per_zone[zone];
5037                         rtpz->rb_root = RB_ROOT;
5038                         spin_lock_init(&rtpz->lock);
5039                 }
5040         }
5041         return 0;
5042 }
5043
5044 static struct cgroup_subsys_state * __ref
5045 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5046 {
5047         struct mem_cgroup *mem, *parent;
5048         long error = -ENOMEM;
5049         int node;
5050
5051         mem = mem_cgroup_alloc();
5052         if (!mem)
5053                 return ERR_PTR(error);
5054
5055         for_each_node_state(node, N_POSSIBLE)
5056                 if (alloc_mem_cgroup_per_zone_info(mem, node))
5057                         goto free_out;
5058
5059         /* root ? */
5060         if (cont->parent == NULL) {
5061                 int cpu;
5062                 enable_swap_cgroup();
5063                 parent = NULL;
5064                 root_mem_cgroup = mem;
5065                 if (mem_cgroup_soft_limit_tree_init())
5066                         goto free_out;
5067                 for_each_possible_cpu(cpu) {
5068                         struct memcg_stock_pcp *stock =
5069                                                 &per_cpu(memcg_stock, cpu);
5070                         INIT_WORK(&stock->work, drain_local_stock);
5071                 }
5072                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5073         } else {
5074                 parent = mem_cgroup_from_cont(cont->parent);
5075                 mem->use_hierarchy = parent->use_hierarchy;
5076                 mem->oom_kill_disable = parent->oom_kill_disable;
5077         }
5078
5079         if (parent && parent->use_hierarchy) {
5080                 res_counter_init(&mem->res, &parent->res);
5081                 res_counter_init(&mem->memsw, &parent->memsw);
5082                 /*
5083                  * We increment refcnt of the parent to ensure that we can
5084                  * safely access it on res_counter_charge/uncharge.
5085                  * This refcnt will be decremented when freeing this
5086                  * mem_cgroup(see mem_cgroup_put).
5087                  */
5088                 mem_cgroup_get(parent);
5089         } else {
5090                 res_counter_init(&mem->res, NULL);
5091                 res_counter_init(&mem->memsw, NULL);
5092         }
5093         mem->last_scanned_child = 0;
5094         mem->last_scanned_node = MAX_NUMNODES;
5095         INIT_LIST_HEAD(&mem->oom_notify);
5096
5097         if (parent)
5098                 mem->swappiness = mem_cgroup_swappiness(parent);
5099         atomic_set(&mem->refcnt, 1);
5100         mem->move_charge_at_immigrate = 0;
5101         mutex_init(&mem->thresholds_lock);
5102         spin_lock_init(&mem->scanstat.lock);
5103         return &mem->css;
5104 free_out:
5105         __mem_cgroup_free(mem);
5106         root_mem_cgroup = NULL;
5107         return ERR_PTR(error);
5108 }
5109
5110 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5111                                         struct cgroup *cont)
5112 {
5113         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5114
5115         return mem_cgroup_force_empty(mem, false);
5116 }
5117
5118 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5119                                 struct cgroup *cont)
5120 {
5121         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5122
5123         mem_cgroup_put(mem);
5124 }
5125
5126 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5127                                 struct cgroup *cont)
5128 {
5129         int ret;
5130
5131         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5132                                 ARRAY_SIZE(mem_cgroup_files));
5133
5134         if (!ret)
5135                 ret = register_memsw_files(cont, ss);
5136         return ret;
5137 }
5138
5139 #ifdef CONFIG_MMU
5140 /* Handlers for move charge at task migration. */
5141 #define PRECHARGE_COUNT_AT_ONCE 256
5142 static int mem_cgroup_do_precharge(unsigned long count)
5143 {
5144         int ret = 0;
5145         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5146         struct mem_cgroup *mem = mc.to;
5147
5148         if (mem_cgroup_is_root(mem)) {
5149                 mc.precharge += count;
5150                 /* we don't need css_get for root */
5151                 return ret;
5152         }
5153         /* try to charge at once */
5154         if (count > 1) {
5155                 struct res_counter *dummy;
5156                 /*
5157                  * "mem" cannot be under rmdir() because we've already checked
5158                  * by cgroup_lock_live_cgroup() that it is not removed and we
5159                  * are still under the same cgroup_mutex. So we can postpone
5160                  * css_get().
5161                  */
5162                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5163                         goto one_by_one;
5164                 if (do_swap_account && res_counter_charge(&mem->memsw,
5165                                                 PAGE_SIZE * count, &dummy)) {
5166                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5167                         goto one_by_one;
5168                 }
5169                 mc.precharge += count;
5170                 return ret;
5171         }
5172 one_by_one:
5173         /* fall back to one by one charge */
5174         while (count--) {
5175                 if (signal_pending(current)) {
5176                         ret = -EINTR;
5177                         break;
5178                 }
5179                 if (!batch_count--) {
5180                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5181                         cond_resched();
5182                 }
5183                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5184                 if (ret || !mem)
5185                         /* mem_cgroup_clear_mc() will do uncharge later */
5186                         return -ENOMEM;
5187                 mc.precharge++;
5188         }
5189         return ret;
5190 }
5191
5192 /**
5193  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5194  * @vma: the vma the pte to be checked belongs
5195  * @addr: the address corresponding to the pte to be checked
5196  * @ptent: the pte to be checked
5197  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5198  *
5199  * Returns
5200  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5201  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5202  *     move charge. if @target is not NULL, the page is stored in target->page
5203  *     with extra refcnt got(Callers should handle it).
5204  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5205  *     target for charge migration. if @target is not NULL, the entry is stored
5206  *     in target->ent.
5207  *
5208  * Called with pte lock held.
5209  */
5210 union mc_target {
5211         struct page     *page;
5212         swp_entry_t     ent;
5213 };
5214
5215 enum mc_target_type {
5216         MC_TARGET_NONE, /* not used */
5217         MC_TARGET_PAGE,
5218         MC_TARGET_SWAP,
5219 };
5220
5221 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5222                                                 unsigned long addr, pte_t ptent)
5223 {
5224         struct page *page = vm_normal_page(vma, addr, ptent);
5225
5226         if (!page || !page_mapped(page))
5227                 return NULL;
5228         if (PageAnon(page)) {
5229                 /* we don't move shared anon */
5230                 if (!move_anon() || page_mapcount(page) > 2)
5231                         return NULL;
5232         } else if (!move_file())
5233                 /* we ignore mapcount for file pages */
5234                 return NULL;
5235         if (!get_page_unless_zero(page))
5236                 return NULL;
5237
5238         return page;
5239 }
5240
5241 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5242                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5243 {
5244         int usage_count;
5245         struct page *page = NULL;
5246         swp_entry_t ent = pte_to_swp_entry(ptent);
5247
5248         if (!move_anon() || non_swap_entry(ent))
5249                 return NULL;
5250         usage_count = mem_cgroup_count_swap_user(ent, &page);
5251         if (usage_count > 1) { /* we don't move shared anon */
5252                 if (page)
5253                         put_page(page);
5254                 return NULL;
5255         }
5256         if (do_swap_account)
5257                 entry->val = ent.val;
5258
5259         return page;
5260 }
5261
5262 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5263                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5264 {
5265         struct page *page = NULL;
5266         struct inode *inode;
5267         struct address_space *mapping;
5268         pgoff_t pgoff;
5269
5270         if (!vma->vm_file) /* anonymous vma */
5271                 return NULL;
5272         if (!move_file())
5273                 return NULL;
5274
5275         inode = vma->vm_file->f_path.dentry->d_inode;
5276         mapping = vma->vm_file->f_mapping;
5277         if (pte_none(ptent))
5278                 pgoff = linear_page_index(vma, addr);
5279         else /* pte_file(ptent) is true */
5280                 pgoff = pte_to_pgoff(ptent);
5281
5282         /* page is moved even if it's not RSS of this task(page-faulted). */
5283         page = find_get_page(mapping, pgoff);
5284
5285 #ifdef CONFIG_SWAP
5286         /* shmem/tmpfs may report page out on swap: account for that too. */
5287         if (radix_tree_exceptional_entry(page)) {
5288                 swp_entry_t swap = radix_to_swp_entry(page);
5289                 if (do_swap_account)
5290                         *entry = swap;
5291                 page = find_get_page(&swapper_space, swap.val);
5292         }
5293 #endif
5294         return page;
5295 }
5296
5297 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5298                 unsigned long addr, pte_t ptent, union mc_target *target)
5299 {
5300         struct page *page = NULL;
5301         struct page_cgroup *pc;
5302         int ret = 0;
5303         swp_entry_t ent = { .val = 0 };
5304
5305         if (pte_present(ptent))
5306                 page = mc_handle_present_pte(vma, addr, ptent);
5307         else if (is_swap_pte(ptent))
5308                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5309         else if (pte_none(ptent) || pte_file(ptent))
5310                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5311
5312         if (!page && !ent.val)
5313                 return 0;
5314         if (page) {
5315                 pc = lookup_page_cgroup(page);
5316                 /*
5317                  * Do only loose check w/o page_cgroup lock.
5318                  * mem_cgroup_move_account() checks the pc is valid or not under
5319                  * the lock.
5320                  */
5321                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5322                         ret = MC_TARGET_PAGE;
5323                         if (target)
5324                                 target->page = page;
5325                 }
5326                 if (!ret || !target)
5327                         put_page(page);
5328         }
5329         /* There is a swap entry and a page doesn't exist or isn't charged */
5330         if (ent.val && !ret &&
5331                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5332                 ret = MC_TARGET_SWAP;
5333                 if (target)
5334                         target->ent = ent;
5335         }
5336         return ret;
5337 }
5338
5339 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5340                                         unsigned long addr, unsigned long end,
5341                                         struct mm_walk *walk)
5342 {
5343         struct vm_area_struct *vma = walk->private;
5344         pte_t *pte;
5345         spinlock_t *ptl;
5346
5347         split_huge_page_pmd(walk->mm, pmd);
5348
5349         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5350         for (; addr != end; pte++, addr += PAGE_SIZE)
5351                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5352                         mc.precharge++; /* increment precharge temporarily */
5353         pte_unmap_unlock(pte - 1, ptl);
5354         cond_resched();
5355
5356         return 0;
5357 }
5358
5359 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5360 {
5361         unsigned long precharge;
5362         struct vm_area_struct *vma;
5363
5364         down_read(&mm->mmap_sem);
5365         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5366                 struct mm_walk mem_cgroup_count_precharge_walk = {
5367                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5368                         .mm = mm,
5369                         .private = vma,
5370                 };
5371                 if (is_vm_hugetlb_page(vma))
5372                         continue;
5373                 walk_page_range(vma->vm_start, vma->vm_end,
5374                                         &mem_cgroup_count_precharge_walk);
5375         }
5376         up_read(&mm->mmap_sem);
5377
5378         precharge = mc.precharge;
5379         mc.precharge = 0;
5380
5381         return precharge;
5382 }
5383
5384 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5385 {
5386         unsigned long precharge = mem_cgroup_count_precharge(mm);
5387
5388         VM_BUG_ON(mc.moving_task);
5389         mc.moving_task = current;
5390         return mem_cgroup_do_precharge(precharge);
5391 }
5392
5393 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5394 static void __mem_cgroup_clear_mc(void)
5395 {
5396         struct mem_cgroup *from = mc.from;
5397         struct mem_cgroup *to = mc.to;
5398
5399         /* we must uncharge all the leftover precharges from mc.to */
5400         if (mc.precharge) {
5401                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5402                 mc.precharge = 0;
5403         }
5404         /*
5405          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5406          * we must uncharge here.
5407          */
5408         if (mc.moved_charge) {
5409                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5410                 mc.moved_charge = 0;
5411         }
5412         /* we must fixup refcnts and charges */
5413         if (mc.moved_swap) {
5414                 /* uncharge swap account from the old cgroup */
5415                 if (!mem_cgroup_is_root(mc.from))
5416                         res_counter_uncharge(&mc.from->memsw,
5417                                                 PAGE_SIZE * mc.moved_swap);
5418                 __mem_cgroup_put(mc.from, mc.moved_swap);
5419
5420                 if (!mem_cgroup_is_root(mc.to)) {
5421                         /*
5422                          * we charged both to->res and to->memsw, so we should
5423                          * uncharge to->res.
5424                          */
5425                         res_counter_uncharge(&mc.to->res,
5426                                                 PAGE_SIZE * mc.moved_swap);
5427                 }
5428                 /* we've already done mem_cgroup_get(mc.to) */
5429                 mc.moved_swap = 0;
5430         }
5431         memcg_oom_recover(from);
5432         memcg_oom_recover(to);
5433         wake_up_all(&mc.waitq);
5434 }
5435
5436 static void mem_cgroup_clear_mc(void)
5437 {
5438         struct mem_cgroup *from = mc.from;
5439
5440         /*
5441          * we must clear moving_task before waking up waiters at the end of
5442          * task migration.
5443          */
5444         mc.moving_task = NULL;
5445         __mem_cgroup_clear_mc();
5446         spin_lock(&mc.lock);
5447         mc.from = NULL;
5448         mc.to = NULL;
5449         spin_unlock(&mc.lock);
5450         mem_cgroup_end_move(from);
5451 }
5452
5453 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5454                                 struct cgroup *cgroup,
5455                                 struct task_struct *p)
5456 {
5457         int ret = 0;
5458         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5459
5460         if (mem->move_charge_at_immigrate) {
5461                 struct mm_struct *mm;
5462                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5463
5464                 VM_BUG_ON(from == mem);
5465
5466                 mm = get_task_mm(p);
5467                 if (!mm)
5468                         return 0;
5469                 /* We move charges only when we move a owner of the mm */
5470                 if (mm->owner == p) {
5471                         VM_BUG_ON(mc.from);
5472                         VM_BUG_ON(mc.to);
5473                         VM_BUG_ON(mc.precharge);
5474                         VM_BUG_ON(mc.moved_charge);
5475                         VM_BUG_ON(mc.moved_swap);
5476                         mem_cgroup_start_move(from);
5477                         spin_lock(&mc.lock);
5478                         mc.from = from;
5479                         mc.to = mem;
5480                         spin_unlock(&mc.lock);
5481                         /* We set mc.moving_task later */
5482
5483                         ret = mem_cgroup_precharge_mc(mm);
5484                         if (ret)
5485                                 mem_cgroup_clear_mc();
5486                 }
5487                 mmput(mm);
5488         }
5489         return ret;
5490 }
5491
5492 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5493                                 struct cgroup *cgroup,
5494                                 struct task_struct *p)
5495 {
5496         mem_cgroup_clear_mc();
5497 }
5498
5499 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5500                                 unsigned long addr, unsigned long end,
5501                                 struct mm_walk *walk)
5502 {
5503         int ret = 0;
5504         struct vm_area_struct *vma = walk->private;
5505         pte_t *pte;
5506         spinlock_t *ptl;
5507
5508         split_huge_page_pmd(walk->mm, pmd);
5509 retry:
5510         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5511         for (; addr != end; addr += PAGE_SIZE) {
5512                 pte_t ptent = *(pte++);
5513                 union mc_target target;
5514                 int type;
5515                 struct page *page;
5516                 struct page_cgroup *pc;
5517                 swp_entry_t ent;
5518
5519                 if (!mc.precharge)
5520                         break;
5521
5522                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5523                 switch (type) {
5524                 case MC_TARGET_PAGE:
5525                         page = target.page;
5526                         if (isolate_lru_page(page))
5527                                 goto put;
5528                         pc = lookup_page_cgroup(page);
5529                         if (!mem_cgroup_move_account(page, 1, pc,
5530                                                      mc.from, mc.to, false)) {
5531                                 mc.precharge--;
5532                                 /* we uncharge from mc.from later. */
5533                                 mc.moved_charge++;
5534                         }
5535                         putback_lru_page(page);
5536 put:                    /* is_target_pte_for_mc() gets the page */
5537                         put_page(page);
5538                         break;
5539                 case MC_TARGET_SWAP:
5540                         ent = target.ent;
5541                         if (!mem_cgroup_move_swap_account(ent,
5542                                                 mc.from, mc.to, false)) {
5543                                 mc.precharge--;
5544                                 /* we fixup refcnts and charges later. */
5545                                 mc.moved_swap++;
5546                         }
5547                         break;
5548                 default:
5549                         break;
5550                 }
5551         }
5552         pte_unmap_unlock(pte - 1, ptl);
5553         cond_resched();
5554
5555         if (addr != end) {
5556                 /*
5557                  * We have consumed all precharges we got in can_attach().
5558                  * We try charge one by one, but don't do any additional
5559                  * charges to mc.to if we have failed in charge once in attach()
5560                  * phase.
5561                  */
5562                 ret = mem_cgroup_do_precharge(1);
5563                 if (!ret)
5564                         goto retry;
5565         }
5566
5567         return ret;
5568 }
5569
5570 static void mem_cgroup_move_charge(struct mm_struct *mm)
5571 {
5572         struct vm_area_struct *vma;
5573
5574         lru_add_drain_all();
5575 retry:
5576         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5577                 /*
5578                  * Someone who are holding the mmap_sem might be waiting in
5579                  * waitq. So we cancel all extra charges, wake up all waiters,
5580                  * and retry. Because we cancel precharges, we might not be able
5581                  * to move enough charges, but moving charge is a best-effort
5582                  * feature anyway, so it wouldn't be a big problem.
5583                  */
5584                 __mem_cgroup_clear_mc();
5585                 cond_resched();
5586                 goto retry;
5587         }
5588         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5589                 int ret;
5590                 struct mm_walk mem_cgroup_move_charge_walk = {
5591                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5592                         .mm = mm,
5593                         .private = vma,
5594                 };
5595                 if (is_vm_hugetlb_page(vma))
5596                         continue;
5597                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5598                                                 &mem_cgroup_move_charge_walk);
5599                 if (ret)
5600                         /*
5601                          * means we have consumed all precharges and failed in
5602                          * doing additional charge. Just abandon here.
5603                          */
5604                         break;
5605         }
5606         up_read(&mm->mmap_sem);
5607 }
5608
5609 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5610                                 struct cgroup *cont,
5611                                 struct cgroup *old_cont,
5612                                 struct task_struct *p)
5613 {
5614         struct mm_struct *mm = get_task_mm(p);
5615
5616         if (mm) {
5617                 if (mc.to)
5618                         mem_cgroup_move_charge(mm);
5619                 put_swap_token(mm);
5620                 mmput(mm);
5621         }
5622         if (mc.to)
5623                 mem_cgroup_clear_mc();
5624 }
5625 #else   /* !CONFIG_MMU */
5626 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5627                                 struct cgroup *cgroup,
5628                                 struct task_struct *p)
5629 {
5630         return 0;
5631 }
5632 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5633                                 struct cgroup *cgroup,
5634                                 struct task_struct *p)
5635 {
5636 }
5637 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5638                                 struct cgroup *cont,
5639                                 struct cgroup *old_cont,
5640                                 struct task_struct *p)
5641 {
5642 }
5643 #endif
5644
5645 struct cgroup_subsys mem_cgroup_subsys = {
5646         .name = "memory",
5647         .subsys_id = mem_cgroup_subsys_id,
5648         .create = mem_cgroup_create,
5649         .pre_destroy = mem_cgroup_pre_destroy,
5650         .destroy = mem_cgroup_destroy,
5651         .populate = mem_cgroup_populate,
5652         .can_attach = mem_cgroup_can_attach,
5653         .cancel_attach = mem_cgroup_cancel_attach,
5654         .attach = mem_cgroup_move_task,
5655         .early_init = 0,
5656         .use_id = 1,
5657 };
5658
5659 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5660 static int __init enable_swap_account(char *s)
5661 {
5662         /* consider enabled if no parameter or 1 is given */
5663         if (!strcmp(s, "1"))
5664                 really_do_swap_account = 1;
5665         else if (!strcmp(s, "0"))
5666                 really_do_swap_account = 0;
5667         return 1;
5668 }
5669 __setup("swapaccount=", enable_swap_account);
5670
5671 #endif