mm: fix wrong vmap address calculations with odd NR_CPUS values
[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 static DEFINE_MUTEX(percpu_charge_mutex);
2095
2096 /*
2097  * Try to consume stocked charge on this cpu. If success, one page is consumed
2098  * from local stock and true is returned. If the stock is 0 or charges from a
2099  * cgroup which is not current target, returns false. This stock will be
2100  * refilled.
2101  */
2102 static bool consume_stock(struct mem_cgroup *mem)
2103 {
2104         struct memcg_stock_pcp *stock;
2105         bool ret = true;
2106
2107         stock = &get_cpu_var(memcg_stock);
2108         if (mem == stock->cached && stock->nr_pages)
2109                 stock->nr_pages--;
2110         else /* need to call res_counter_charge */
2111                 ret = false;
2112         put_cpu_var(memcg_stock);
2113         return ret;
2114 }
2115
2116 /*
2117  * Returns stocks cached in percpu to res_counter and reset cached information.
2118  */
2119 static void drain_stock(struct memcg_stock_pcp *stock)
2120 {
2121         struct mem_cgroup *old = stock->cached;
2122
2123         if (stock->nr_pages) {
2124                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2125
2126                 res_counter_uncharge(&old->res, bytes);
2127                 if (do_swap_account)
2128                         res_counter_uncharge(&old->memsw, bytes);
2129                 stock->nr_pages = 0;
2130         }
2131         stock->cached = NULL;
2132 }
2133
2134 /*
2135  * This must be called under preempt disabled or must be called by
2136  * a thread which is pinned to local cpu.
2137  */
2138 static void drain_local_stock(struct work_struct *dummy)
2139 {
2140         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2141         drain_stock(stock);
2142         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2143 }
2144
2145 /*
2146  * Cache charges(val) which is from res_counter, to local per_cpu area.
2147  * This will be consumed by consume_stock() function, later.
2148  */
2149 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2150 {
2151         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2152
2153         if (stock->cached != mem) { /* reset if necessary */
2154                 drain_stock(stock);
2155                 stock->cached = mem;
2156         }
2157         stock->nr_pages += nr_pages;
2158         put_cpu_var(memcg_stock);
2159 }
2160
2161 /*
2162  * Drains all per-CPU charge caches for given root_mem resp. subtree
2163  * of the hierarchy under it. sync flag says whether we should block
2164  * until the work is done.
2165  */
2166 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2167 {
2168         int cpu, curcpu;
2169
2170         /* Notify other cpus that system-wide "drain" is running */
2171         get_online_cpus();
2172         /*
2173          * Get a hint for avoiding draining charges on the current cpu,
2174          * which must be exhausted by our charging.  It is not required that
2175          * this be a precise check, so we use raw_smp_processor_id() instead of
2176          * getcpu()/putcpu().
2177          */
2178         curcpu = raw_smp_processor_id();
2179         for_each_online_cpu(cpu) {
2180                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2181                 struct mem_cgroup *mem;
2182
2183                 mem = stock->cached;
2184                 if (!mem || !stock->nr_pages)
2185                         continue;
2186                 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2187                         continue;
2188                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2189                         if (cpu == curcpu)
2190                                 drain_local_stock(&stock->work);
2191                         else
2192                                 schedule_work_on(cpu, &stock->work);
2193                 }
2194         }
2195
2196         if (!sync)
2197                 goto out;
2198
2199         for_each_online_cpu(cpu) {
2200                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2201                 if (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         /*
2217          * If someone calls draining, avoid adding more kworker runs.
2218          */
2219         if (!mutex_trylock(&percpu_charge_mutex))
2220                 return;
2221         drain_all_stock(root_mem, false);
2222         mutex_unlock(&percpu_charge_mutex);
2223 }
2224
2225 /* This is a synchronous drain interface. */
2226 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2227 {
2228         /* called when force_empty is called */
2229         mutex_lock(&percpu_charge_mutex);
2230         drain_all_stock(root_mem, true);
2231         mutex_unlock(&percpu_charge_mutex);
2232 }
2233
2234 /*
2235  * This function drains percpu counter value from DEAD cpu and
2236  * move it to local cpu. Note that this function can be preempted.
2237  */
2238 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2239 {
2240         int i;
2241
2242         spin_lock(&mem->pcp_counter_lock);
2243         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2244                 long x = per_cpu(mem->stat->count[i], cpu);
2245
2246                 per_cpu(mem->stat->count[i], cpu) = 0;
2247                 mem->nocpu_base.count[i] += x;
2248         }
2249         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2250                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2251
2252                 per_cpu(mem->stat->events[i], cpu) = 0;
2253                 mem->nocpu_base.events[i] += x;
2254         }
2255         /* need to clear ON_MOVE value, works as a kind of lock. */
2256         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2257         spin_unlock(&mem->pcp_counter_lock);
2258 }
2259
2260 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2261 {
2262         int idx = MEM_CGROUP_ON_MOVE;
2263
2264         spin_lock(&mem->pcp_counter_lock);
2265         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2266         spin_unlock(&mem->pcp_counter_lock);
2267 }
2268
2269 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2270                                         unsigned long action,
2271                                         void *hcpu)
2272 {
2273         int cpu = (unsigned long)hcpu;
2274         struct memcg_stock_pcp *stock;
2275         struct mem_cgroup *iter;
2276
2277         if ((action == CPU_ONLINE)) {
2278                 for_each_mem_cgroup_all(iter)
2279                         synchronize_mem_cgroup_on_move(iter, cpu);
2280                 return NOTIFY_OK;
2281         }
2282
2283         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2284                 return NOTIFY_OK;
2285
2286         for_each_mem_cgroup_all(iter)
2287                 mem_cgroup_drain_pcp_counter(iter, cpu);
2288
2289         stock = &per_cpu(memcg_stock, cpu);
2290         drain_stock(stock);
2291         return NOTIFY_OK;
2292 }
2293
2294
2295 /* See __mem_cgroup_try_charge() for details */
2296 enum {
2297         CHARGE_OK,              /* success */
2298         CHARGE_RETRY,           /* need to retry but retry is not bad */
2299         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2300         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2301         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2302 };
2303
2304 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2305                                 unsigned int nr_pages, bool oom_check)
2306 {
2307         unsigned long csize = nr_pages * PAGE_SIZE;
2308         struct mem_cgroup *mem_over_limit;
2309         struct res_counter *fail_res;
2310         unsigned long flags = 0;
2311         int ret;
2312
2313         ret = res_counter_charge(&mem->res, csize, &fail_res);
2314
2315         if (likely(!ret)) {
2316                 if (!do_swap_account)
2317                         return CHARGE_OK;
2318                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2319                 if (likely(!ret))
2320                         return CHARGE_OK;
2321
2322                 res_counter_uncharge(&mem->res, csize);
2323                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2324                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2325         } else
2326                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2327         /*
2328          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2329          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2330          *
2331          * Never reclaim on behalf of optional batching, retry with a
2332          * single page instead.
2333          */
2334         if (nr_pages == CHARGE_BATCH)
2335                 return CHARGE_RETRY;
2336
2337         if (!(gfp_mask & __GFP_WAIT))
2338                 return CHARGE_WOULDBLOCK;
2339
2340         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2341                                               gfp_mask, flags, NULL);
2342         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2343                 return CHARGE_RETRY;
2344         /*
2345          * Even though the limit is exceeded at this point, reclaim
2346          * may have been able to free some pages.  Retry the charge
2347          * before killing the task.
2348          *
2349          * Only for regular pages, though: huge pages are rather
2350          * unlikely to succeed so close to the limit, and we fall back
2351          * to regular pages anyway in case of failure.
2352          */
2353         if (nr_pages == 1 && ret)
2354                 return CHARGE_RETRY;
2355
2356         /*
2357          * At task move, charge accounts can be doubly counted. So, it's
2358          * better to wait until the end of task_move if something is going on.
2359          */
2360         if (mem_cgroup_wait_acct_move(mem_over_limit))
2361                 return CHARGE_RETRY;
2362
2363         /* If we don't need to call oom-killer at el, return immediately */
2364         if (!oom_check)
2365                 return CHARGE_NOMEM;
2366         /* check OOM */
2367         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2368                 return CHARGE_OOM_DIE;
2369
2370         return CHARGE_RETRY;
2371 }
2372
2373 /*
2374  * Unlike exported interface, "oom" parameter is added. if oom==true,
2375  * oom-killer can be invoked.
2376  */
2377 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2378                                    gfp_t gfp_mask,
2379                                    unsigned int nr_pages,
2380                                    struct mem_cgroup **memcg,
2381                                    bool oom)
2382 {
2383         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2384         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2385         struct mem_cgroup *mem = NULL;
2386         int ret;
2387
2388         /*
2389          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2390          * in system level. So, allow to go ahead dying process in addition to
2391          * MEMDIE process.
2392          */
2393         if (unlikely(test_thread_flag(TIF_MEMDIE)
2394                      || fatal_signal_pending(current)))
2395                 goto bypass;
2396
2397         /*
2398          * We always charge the cgroup the mm_struct belongs to.
2399          * The mm_struct's mem_cgroup changes on task migration if the
2400          * thread group leader migrates. It's possible that mm is not
2401          * set, if so charge the init_mm (happens for pagecache usage).
2402          */
2403         if (!*memcg && !mm)
2404                 goto bypass;
2405 again:
2406         if (*memcg) { /* css should be a valid one */
2407                 mem = *memcg;
2408                 VM_BUG_ON(css_is_removed(&mem->css));
2409                 if (mem_cgroup_is_root(mem))
2410                         goto done;
2411                 if (nr_pages == 1 && consume_stock(mem))
2412                         goto done;
2413                 css_get(&mem->css);
2414         } else {
2415                 struct task_struct *p;
2416
2417                 rcu_read_lock();
2418                 p = rcu_dereference(mm->owner);
2419                 /*
2420                  * Because we don't have task_lock(), "p" can exit.
2421                  * In that case, "mem" can point to root or p can be NULL with
2422                  * race with swapoff. Then, we have small risk of mis-accouning.
2423                  * But such kind of mis-account by race always happens because
2424                  * we don't have cgroup_mutex(). It's overkill and we allo that
2425                  * small race, here.
2426                  * (*) swapoff at el will charge against mm-struct not against
2427                  * task-struct. So, mm->owner can be NULL.
2428                  */
2429                 mem = mem_cgroup_from_task(p);
2430                 if (!mem || mem_cgroup_is_root(mem)) {
2431                         rcu_read_unlock();
2432                         goto done;
2433                 }
2434                 if (nr_pages == 1 && consume_stock(mem)) {
2435                         /*
2436                          * It seems dagerous to access memcg without css_get().
2437                          * But considering how consume_stok works, it's not
2438                          * necessary. If consume_stock success, some charges
2439                          * from this memcg are cached on this cpu. So, we
2440                          * don't need to call css_get()/css_tryget() before
2441                          * calling consume_stock().
2442                          */
2443                         rcu_read_unlock();
2444                         goto done;
2445                 }
2446                 /* after here, we may be blocked. we need to get refcnt */
2447                 if (!css_tryget(&mem->css)) {
2448                         rcu_read_unlock();
2449                         goto again;
2450                 }
2451                 rcu_read_unlock();
2452         }
2453
2454         do {
2455                 bool oom_check;
2456
2457                 /* If killed, bypass charge */
2458                 if (fatal_signal_pending(current)) {
2459                         css_put(&mem->css);
2460                         goto bypass;
2461                 }
2462
2463                 oom_check = false;
2464                 if (oom && !nr_oom_retries) {
2465                         oom_check = true;
2466                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2467                 }
2468
2469                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2470                 switch (ret) {
2471                 case CHARGE_OK:
2472                         break;
2473                 case CHARGE_RETRY: /* not in OOM situation but retry */
2474                         batch = nr_pages;
2475                         css_put(&mem->css);
2476                         mem = NULL;
2477                         goto again;
2478                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2479                         css_put(&mem->css);
2480                         goto nomem;
2481                 case CHARGE_NOMEM: /* OOM routine works */
2482                         if (!oom) {
2483                                 css_put(&mem->css);
2484                                 goto nomem;
2485                         }
2486                         /* If oom, we never return -ENOMEM */
2487                         nr_oom_retries--;
2488                         break;
2489                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2490                         css_put(&mem->css);
2491                         goto bypass;
2492                 }
2493         } while (ret != CHARGE_OK);
2494
2495         if (batch > nr_pages)
2496                 refill_stock(mem, batch - nr_pages);
2497         css_put(&mem->css);
2498 done:
2499         *memcg = mem;
2500         return 0;
2501 nomem:
2502         *memcg = NULL;
2503         return -ENOMEM;
2504 bypass:
2505         *memcg = NULL;
2506         return 0;
2507 }
2508
2509 /*
2510  * Somemtimes we have to undo a charge we got by try_charge().
2511  * This function is for that and do uncharge, put css's refcnt.
2512  * gotten by try_charge().
2513  */
2514 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2515                                        unsigned int nr_pages)
2516 {
2517         if (!mem_cgroup_is_root(mem)) {
2518                 unsigned long bytes = nr_pages * PAGE_SIZE;
2519
2520                 res_counter_uncharge(&mem->res, bytes);
2521                 if (do_swap_account)
2522                         res_counter_uncharge(&mem->memsw, bytes);
2523         }
2524 }
2525
2526 /*
2527  * A helper function to get mem_cgroup from ID. must be called under
2528  * rcu_read_lock(). The caller must check css_is_removed() or some if
2529  * it's concern. (dropping refcnt from swap can be called against removed
2530  * memcg.)
2531  */
2532 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2533 {
2534         struct cgroup_subsys_state *css;
2535
2536         /* ID 0 is unused ID */
2537         if (!id)
2538                 return NULL;
2539         css = css_lookup(&mem_cgroup_subsys, id);
2540         if (!css)
2541                 return NULL;
2542         return container_of(css, struct mem_cgroup, css);
2543 }
2544
2545 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2546 {
2547         struct mem_cgroup *mem = NULL;
2548         struct page_cgroup *pc;
2549         unsigned short id;
2550         swp_entry_t ent;
2551
2552         VM_BUG_ON(!PageLocked(page));
2553
2554         pc = lookup_page_cgroup(page);
2555         lock_page_cgroup(pc);
2556         if (PageCgroupUsed(pc)) {
2557                 mem = pc->mem_cgroup;
2558                 if (mem && !css_tryget(&mem->css))
2559                         mem = NULL;
2560         } else if (PageSwapCache(page)) {
2561                 ent.val = page_private(page);
2562                 id = lookup_swap_cgroup(ent);
2563                 rcu_read_lock();
2564                 mem = mem_cgroup_lookup(id);
2565                 if (mem && !css_tryget(&mem->css))
2566                         mem = NULL;
2567                 rcu_read_unlock();
2568         }
2569         unlock_page_cgroup(pc);
2570         return mem;
2571 }
2572
2573 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2574                                        struct page *page,
2575                                        unsigned int nr_pages,
2576                                        struct page_cgroup *pc,
2577                                        enum charge_type ctype)
2578 {
2579         lock_page_cgroup(pc);
2580         if (unlikely(PageCgroupUsed(pc))) {
2581                 unlock_page_cgroup(pc);
2582                 __mem_cgroup_cancel_charge(mem, nr_pages);
2583                 return;
2584         }
2585         /*
2586          * we don't need page_cgroup_lock about tail pages, becase they are not
2587          * accessed by any other context at this point.
2588          */
2589         pc->mem_cgroup = mem;
2590         /*
2591          * We access a page_cgroup asynchronously without lock_page_cgroup().
2592          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2593          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2594          * before USED bit, we need memory barrier here.
2595          * See mem_cgroup_add_lru_list(), etc.
2596          */
2597         smp_wmb();
2598         switch (ctype) {
2599         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2600         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2601                 SetPageCgroupCache(pc);
2602                 SetPageCgroupUsed(pc);
2603                 break;
2604         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2605                 ClearPageCgroupCache(pc);
2606                 SetPageCgroupUsed(pc);
2607                 break;
2608         default:
2609                 break;
2610         }
2611
2612         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2613         unlock_page_cgroup(pc);
2614         /*
2615          * "charge_statistics" updated event counter. Then, check it.
2616          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2617          * if they exceeds softlimit.
2618          */
2619         memcg_check_events(mem, page);
2620 }
2621
2622 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2623
2624 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2625                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2626 /*
2627  * Because tail pages are not marked as "used", set it. We're under
2628  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2629  */
2630 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2631 {
2632         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2633         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2634         unsigned long flags;
2635
2636         if (mem_cgroup_disabled())
2637                 return;
2638         /*
2639          * We have no races with charge/uncharge but will have races with
2640          * page state accounting.
2641          */
2642         move_lock_page_cgroup(head_pc, &flags);
2643
2644         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2645         smp_wmb(); /* see __commit_charge() */
2646         if (PageCgroupAcctLRU(head_pc)) {
2647                 enum lru_list lru;
2648                 struct mem_cgroup_per_zone *mz;
2649
2650                 /*
2651                  * LRU flags cannot be copied because we need to add tail
2652                  *.page to LRU by generic call and our hook will be called.
2653                  * We hold lru_lock, then, reduce counter directly.
2654                  */
2655                 lru = page_lru(head);
2656                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2657                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2658         }
2659         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2660         move_unlock_page_cgroup(head_pc, &flags);
2661 }
2662 #endif
2663
2664 /**
2665  * mem_cgroup_move_account - move account of the page
2666  * @page: the page
2667  * @nr_pages: number of regular pages (>1 for huge pages)
2668  * @pc: page_cgroup of the page.
2669  * @from: mem_cgroup which the page is moved from.
2670  * @to: mem_cgroup which the page is moved to. @from != @to.
2671  * @uncharge: whether we should call uncharge and css_put against @from.
2672  *
2673  * The caller must confirm following.
2674  * - page is not on LRU (isolate_page() is useful.)
2675  * - compound_lock is held when nr_pages > 1
2676  *
2677  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2678  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2679  * true, this function does "uncharge" from old cgroup, but it doesn't if
2680  * @uncharge is false, so a caller should do "uncharge".
2681  */
2682 static int mem_cgroup_move_account(struct page *page,
2683                                    unsigned int nr_pages,
2684                                    struct page_cgroup *pc,
2685                                    struct mem_cgroup *from,
2686                                    struct mem_cgroup *to,
2687                                    bool uncharge)
2688 {
2689         unsigned long flags;
2690         int ret;
2691
2692         VM_BUG_ON(from == to);
2693         VM_BUG_ON(PageLRU(page));
2694         /*
2695          * The page is isolated from LRU. So, collapse function
2696          * will not handle this page. But page splitting can happen.
2697          * Do this check under compound_page_lock(). The caller should
2698          * hold it.
2699          */
2700         ret = -EBUSY;
2701         if (nr_pages > 1 && !PageTransHuge(page))
2702                 goto out;
2703
2704         lock_page_cgroup(pc);
2705
2706         ret = -EINVAL;
2707         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2708                 goto unlock;
2709
2710         move_lock_page_cgroup(pc, &flags);
2711
2712         if (PageCgroupFileMapped(pc)) {
2713                 /* Update mapped_file data for mem_cgroup */
2714                 preempt_disable();
2715                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2716                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2717                 preempt_enable();
2718         }
2719         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2720         if (uncharge)
2721                 /* This is not "cancel", but cancel_charge does all we need. */
2722                 __mem_cgroup_cancel_charge(from, nr_pages);
2723
2724         /* caller should have done css_get */
2725         pc->mem_cgroup = to;
2726         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2727         /*
2728          * We charges against "to" which may not have any tasks. Then, "to"
2729          * can be under rmdir(). But in current implementation, caller of
2730          * this function is just force_empty() and move charge, so it's
2731          * guaranteed that "to" is never removed. So, we don't check rmdir
2732          * status here.
2733          */
2734         move_unlock_page_cgroup(pc, &flags);
2735         ret = 0;
2736 unlock:
2737         unlock_page_cgroup(pc);
2738         /*
2739          * check events
2740          */
2741         memcg_check_events(to, page);
2742         memcg_check_events(from, page);
2743 out:
2744         return ret;
2745 }
2746
2747 /*
2748  * move charges to its parent.
2749  */
2750
2751 static int mem_cgroup_move_parent(struct page *page,
2752                                   struct page_cgroup *pc,
2753                                   struct mem_cgroup *child,
2754                                   gfp_t gfp_mask)
2755 {
2756         struct cgroup *cg = child->css.cgroup;
2757         struct cgroup *pcg = cg->parent;
2758         struct mem_cgroup *parent;
2759         unsigned int nr_pages;
2760         unsigned long uninitialized_var(flags);
2761         int ret;
2762
2763         /* Is ROOT ? */
2764         if (!pcg)
2765                 return -EINVAL;
2766
2767         ret = -EBUSY;
2768         if (!get_page_unless_zero(page))
2769                 goto out;
2770         if (isolate_lru_page(page))
2771                 goto put;
2772
2773         nr_pages = hpage_nr_pages(page);
2774
2775         parent = mem_cgroup_from_cont(pcg);
2776         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2777         if (ret || !parent)
2778                 goto put_back;
2779
2780         if (nr_pages > 1)
2781                 flags = compound_lock_irqsave(page);
2782
2783         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2784         if (ret)
2785                 __mem_cgroup_cancel_charge(parent, nr_pages);
2786
2787         if (nr_pages > 1)
2788                 compound_unlock_irqrestore(page, flags);
2789 put_back:
2790         putback_lru_page(page);
2791 put:
2792         put_page(page);
2793 out:
2794         return ret;
2795 }
2796
2797 /*
2798  * Charge the memory controller for page usage.
2799  * Return
2800  * 0 if the charge was successful
2801  * < 0 if the cgroup is over its limit
2802  */
2803 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2804                                 gfp_t gfp_mask, enum charge_type ctype)
2805 {
2806         struct mem_cgroup *mem = NULL;
2807         unsigned int nr_pages = 1;
2808         struct page_cgroup *pc;
2809         bool oom = true;
2810         int ret;
2811
2812         if (PageTransHuge(page)) {
2813                 nr_pages <<= compound_order(page);
2814                 VM_BUG_ON(!PageTransHuge(page));
2815                 /*
2816                  * Never OOM-kill a process for a huge page.  The
2817                  * fault handler will fall back to regular pages.
2818                  */
2819                 oom = false;
2820         }
2821
2822         pc = lookup_page_cgroup(page);
2823         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2824
2825         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2826         if (ret || !mem)
2827                 return ret;
2828
2829         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2830         return 0;
2831 }
2832
2833 int mem_cgroup_newpage_charge(struct page *page,
2834                               struct mm_struct *mm, gfp_t gfp_mask)
2835 {
2836         if (mem_cgroup_disabled())
2837                 return 0;
2838         /*
2839          * If already mapped, we don't have to account.
2840          * If page cache, page->mapping has address_space.
2841          * But page->mapping may have out-of-use anon_vma pointer,
2842          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2843          * is NULL.
2844          */
2845         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2846                 return 0;
2847         if (unlikely(!mm))
2848                 mm = &init_mm;
2849         return mem_cgroup_charge_common(page, mm, gfp_mask,
2850                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2851 }
2852
2853 static void
2854 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2855                                         enum charge_type ctype);
2856
2857 static void
2858 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2859                                         enum charge_type ctype)
2860 {
2861         struct page_cgroup *pc = lookup_page_cgroup(page);
2862         /*
2863          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2864          * is already on LRU. It means the page may on some other page_cgroup's
2865          * LRU. Take care of it.
2866          */
2867         mem_cgroup_lru_del_before_commit(page);
2868         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2869         mem_cgroup_lru_add_after_commit(page);
2870         return;
2871 }
2872
2873 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2874                                 gfp_t gfp_mask)
2875 {
2876         struct mem_cgroup *mem = NULL;
2877         int ret;
2878
2879         if (mem_cgroup_disabled())
2880                 return 0;
2881         if (PageCompound(page))
2882                 return 0;
2883
2884         if (unlikely(!mm))
2885                 mm = &init_mm;
2886
2887         if (page_is_file_cache(page)) {
2888                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2889                 if (ret || !mem)
2890                         return ret;
2891
2892                 /*
2893                  * FUSE reuses pages without going through the final
2894                  * put that would remove them from the LRU list, make
2895                  * sure that they get relinked properly.
2896                  */
2897                 __mem_cgroup_commit_charge_lrucare(page, mem,
2898                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2899                 return ret;
2900         }
2901         /* shmem */
2902         if (PageSwapCache(page)) {
2903                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2904                 if (!ret)
2905                         __mem_cgroup_commit_charge_swapin(page, mem,
2906                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2907         } else
2908                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2909                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2910
2911         return ret;
2912 }
2913
2914 /*
2915  * While swap-in, try_charge -> commit or cancel, the page is locked.
2916  * And when try_charge() successfully returns, one refcnt to memcg without
2917  * struct page_cgroup is acquired. This refcnt will be consumed by
2918  * "commit()" or removed by "cancel()"
2919  */
2920 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2921                                  struct page *page,
2922                                  gfp_t mask, struct mem_cgroup **ptr)
2923 {
2924         struct mem_cgroup *mem;
2925         int ret;
2926
2927         *ptr = NULL;
2928
2929         if (mem_cgroup_disabled())
2930                 return 0;
2931
2932         if (!do_swap_account)
2933                 goto charge_cur_mm;
2934         /*
2935          * A racing thread's fault, or swapoff, may have already updated
2936          * the pte, and even removed page from swap cache: in those cases
2937          * do_swap_page()'s pte_same() test will fail; but there's also a
2938          * KSM case which does need to charge the page.
2939          */
2940         if (!PageSwapCache(page))
2941                 goto charge_cur_mm;
2942         mem = try_get_mem_cgroup_from_page(page);
2943         if (!mem)
2944                 goto charge_cur_mm;
2945         *ptr = mem;
2946         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2947         css_put(&mem->css);
2948         return ret;
2949 charge_cur_mm:
2950         if (unlikely(!mm))
2951                 mm = &init_mm;
2952         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2953 }
2954
2955 static void
2956 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2957                                         enum charge_type ctype)
2958 {
2959         if (mem_cgroup_disabled())
2960                 return;
2961         if (!ptr)
2962                 return;
2963         cgroup_exclude_rmdir(&ptr->css);
2964
2965         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2966         /*
2967          * Now swap is on-memory. This means this page may be
2968          * counted both as mem and swap....double count.
2969          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2970          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2971          * may call delete_from_swap_cache() before reach here.
2972          */
2973         if (do_swap_account && PageSwapCache(page)) {
2974                 swp_entry_t ent = {.val = page_private(page)};
2975                 unsigned short id;
2976                 struct mem_cgroup *memcg;
2977
2978                 id = swap_cgroup_record(ent, 0);
2979                 rcu_read_lock();
2980                 memcg = mem_cgroup_lookup(id);
2981                 if (memcg) {
2982                         /*
2983                          * This recorded memcg can be obsolete one. So, avoid
2984                          * calling css_tryget
2985                          */
2986                         if (!mem_cgroup_is_root(memcg))
2987                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2988                         mem_cgroup_swap_statistics(memcg, false);
2989                         mem_cgroup_put(memcg);
2990                 }
2991                 rcu_read_unlock();
2992         }
2993         /*
2994          * At swapin, we may charge account against cgroup which has no tasks.
2995          * So, rmdir()->pre_destroy() can be called while we do this charge.
2996          * In that case, we need to call pre_destroy() again. check it here.
2997          */
2998         cgroup_release_and_wakeup_rmdir(&ptr->css);
2999 }
3000
3001 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
3002 {
3003         __mem_cgroup_commit_charge_swapin(page, ptr,
3004                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
3005 }
3006
3007 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3008 {
3009         if (mem_cgroup_disabled())
3010                 return;
3011         if (!mem)
3012                 return;
3013         __mem_cgroup_cancel_charge(mem, 1);
3014 }
3015
3016 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3017                                    unsigned int nr_pages,
3018                                    const enum charge_type ctype)
3019 {
3020         struct memcg_batch_info *batch = NULL;
3021         bool uncharge_memsw = true;
3022
3023         /* If swapout, usage of swap doesn't decrease */
3024         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3025                 uncharge_memsw = false;
3026
3027         batch = &current->memcg_batch;
3028         /*
3029          * In usual, we do css_get() when we remember memcg pointer.
3030          * But in this case, we keep res->usage until end of a series of
3031          * uncharges. Then, it's ok to ignore memcg's refcnt.
3032          */
3033         if (!batch->memcg)
3034                 batch->memcg = mem;
3035         /*
3036          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3037          * In those cases, all pages freed continuously can be expected to be in
3038          * the same cgroup and we have chance to coalesce uncharges.
3039          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3040          * because we want to do uncharge as soon as possible.
3041          */
3042
3043         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3044                 goto direct_uncharge;
3045
3046         if (nr_pages > 1)
3047                 goto direct_uncharge;
3048
3049         /*
3050          * In typical case, batch->memcg == mem. This means we can
3051          * merge a series of uncharges to an uncharge of res_counter.
3052          * If not, we uncharge res_counter ony by one.
3053          */
3054         if (batch->memcg != mem)
3055                 goto direct_uncharge;
3056         /* remember freed charge and uncharge it later */
3057         batch->nr_pages++;
3058         if (uncharge_memsw)
3059                 batch->memsw_nr_pages++;
3060         return;
3061 direct_uncharge:
3062         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3063         if (uncharge_memsw)
3064                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3065         if (unlikely(batch->memcg != mem))
3066                 memcg_oom_recover(mem);
3067         return;
3068 }
3069
3070 /*
3071  * uncharge if !page_mapped(page)
3072  */
3073 static struct mem_cgroup *
3074 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3075 {
3076         struct mem_cgroup *mem = NULL;
3077         unsigned int nr_pages = 1;
3078         struct page_cgroup *pc;
3079
3080         if (mem_cgroup_disabled())
3081                 return NULL;
3082
3083         if (PageSwapCache(page))
3084                 return NULL;
3085
3086         if (PageTransHuge(page)) {
3087                 nr_pages <<= compound_order(page);
3088                 VM_BUG_ON(!PageTransHuge(page));
3089         }
3090         /*
3091          * Check if our page_cgroup is valid
3092          */
3093         pc = lookup_page_cgroup(page);
3094         if (unlikely(!pc || !PageCgroupUsed(pc)))
3095                 return NULL;
3096
3097         lock_page_cgroup(pc);
3098
3099         mem = pc->mem_cgroup;
3100
3101         if (!PageCgroupUsed(pc))
3102                 goto unlock_out;
3103
3104         switch (ctype) {
3105         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3106         case MEM_CGROUP_CHARGE_TYPE_DROP:
3107                 /* See mem_cgroup_prepare_migration() */
3108                 if (page_mapped(page) || PageCgroupMigration(pc))
3109                         goto unlock_out;
3110                 break;
3111         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3112                 if (!PageAnon(page)) {  /* Shared memory */
3113                         if (page->mapping && !page_is_file_cache(page))
3114                                 goto unlock_out;
3115                 } else if (page_mapped(page)) /* Anon */
3116                                 goto unlock_out;
3117                 break;
3118         default:
3119                 break;
3120         }
3121
3122         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3123
3124         ClearPageCgroupUsed(pc);
3125         /*
3126          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3127          * freed from LRU. This is safe because uncharged page is expected not
3128          * to be reused (freed soon). Exception is SwapCache, it's handled by
3129          * special functions.
3130          */
3131
3132         unlock_page_cgroup(pc);
3133         /*
3134          * even after unlock, we have mem->res.usage here and this memcg
3135          * will never be freed.
3136          */
3137         memcg_check_events(mem, page);
3138         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3139                 mem_cgroup_swap_statistics(mem, true);
3140                 mem_cgroup_get(mem);
3141         }
3142         if (!mem_cgroup_is_root(mem))
3143                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3144
3145         return mem;
3146
3147 unlock_out:
3148         unlock_page_cgroup(pc);
3149         return NULL;
3150 }
3151
3152 void mem_cgroup_uncharge_page(struct page *page)
3153 {
3154         /* early check. */
3155         if (page_mapped(page))
3156                 return;
3157         if (page->mapping && !PageAnon(page))
3158                 return;
3159         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3160 }
3161
3162 void mem_cgroup_uncharge_cache_page(struct page *page)
3163 {
3164         VM_BUG_ON(page_mapped(page));
3165         VM_BUG_ON(page->mapping);
3166         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3167 }
3168
3169 /*
3170  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3171  * In that cases, pages are freed continuously and we can expect pages
3172  * are in the same memcg. All these calls itself limits the number of
3173  * pages freed at once, then uncharge_start/end() is called properly.
3174  * This may be called prural(2) times in a context,
3175  */
3176
3177 void mem_cgroup_uncharge_start(void)
3178 {
3179         current->memcg_batch.do_batch++;
3180         /* We can do nest. */
3181         if (current->memcg_batch.do_batch == 1) {
3182                 current->memcg_batch.memcg = NULL;
3183                 current->memcg_batch.nr_pages = 0;
3184                 current->memcg_batch.memsw_nr_pages = 0;
3185         }
3186 }
3187
3188 void mem_cgroup_uncharge_end(void)
3189 {
3190         struct memcg_batch_info *batch = &current->memcg_batch;
3191
3192         if (!batch->do_batch)
3193                 return;
3194
3195         batch->do_batch--;
3196         if (batch->do_batch) /* If stacked, do nothing. */
3197                 return;
3198
3199         if (!batch->memcg)
3200                 return;
3201         /*
3202          * This "batch->memcg" is valid without any css_get/put etc...
3203          * bacause we hide charges behind us.
3204          */
3205         if (batch->nr_pages)
3206                 res_counter_uncharge(&batch->memcg->res,
3207                                      batch->nr_pages * PAGE_SIZE);
3208         if (batch->memsw_nr_pages)
3209                 res_counter_uncharge(&batch->memcg->memsw,
3210                                      batch->memsw_nr_pages * PAGE_SIZE);
3211         memcg_oom_recover(batch->memcg);
3212         /* forget this pointer (for sanity check) */
3213         batch->memcg = NULL;
3214 }
3215
3216 #ifdef CONFIG_SWAP
3217 /*
3218  * called after __delete_from_swap_cache() and drop "page" account.
3219  * memcg information is recorded to swap_cgroup of "ent"
3220  */
3221 void
3222 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3223 {
3224         struct mem_cgroup *memcg;
3225         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3226
3227         if (!swapout) /* this was a swap cache but the swap is unused ! */
3228                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3229
3230         memcg = __mem_cgroup_uncharge_common(page, ctype);
3231
3232         /*
3233          * record memcg information,  if swapout && memcg != NULL,
3234          * mem_cgroup_get() was called in uncharge().
3235          */
3236         if (do_swap_account && swapout && memcg)
3237                 swap_cgroup_record(ent, css_id(&memcg->css));
3238 }
3239 #endif
3240
3241 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3242 /*
3243  * called from swap_entry_free(). remove record in swap_cgroup and
3244  * uncharge "memsw" account.
3245  */
3246 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3247 {
3248         struct mem_cgroup *memcg;
3249         unsigned short id;
3250
3251         if (!do_swap_account)
3252                 return;
3253
3254         id = swap_cgroup_record(ent, 0);
3255         rcu_read_lock();
3256         memcg = mem_cgroup_lookup(id);
3257         if (memcg) {
3258                 /*
3259                  * We uncharge this because swap is freed.
3260                  * This memcg can be obsolete one. We avoid calling css_tryget
3261                  */
3262                 if (!mem_cgroup_is_root(memcg))
3263                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3264                 mem_cgroup_swap_statistics(memcg, false);
3265                 mem_cgroup_put(memcg);
3266         }
3267         rcu_read_unlock();
3268 }
3269
3270 /**
3271  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3272  * @entry: swap entry to be moved
3273  * @from:  mem_cgroup which the entry is moved from
3274  * @to:  mem_cgroup which the entry is moved to
3275  * @need_fixup: whether we should fixup res_counters and refcounts.
3276  *
3277  * It succeeds only when the swap_cgroup's record for this entry is the same
3278  * as the mem_cgroup's id of @from.
3279  *
3280  * Returns 0 on success, -EINVAL on failure.
3281  *
3282  * The caller must have charged to @to, IOW, called res_counter_charge() about
3283  * both res and memsw, and called css_get().
3284  */
3285 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3286                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3287 {
3288         unsigned short old_id, new_id;
3289
3290         old_id = css_id(&from->css);
3291         new_id = css_id(&to->css);
3292
3293         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3294                 mem_cgroup_swap_statistics(from, false);
3295                 mem_cgroup_swap_statistics(to, true);
3296                 /*
3297                  * This function is only called from task migration context now.
3298                  * It postpones res_counter and refcount handling till the end
3299                  * of task migration(mem_cgroup_clear_mc()) for performance
3300                  * improvement. But we cannot postpone mem_cgroup_get(to)
3301                  * because if the process that has been moved to @to does
3302                  * swap-in, the refcount of @to might be decreased to 0.
3303                  */
3304                 mem_cgroup_get(to);
3305                 if (need_fixup) {
3306                         if (!mem_cgroup_is_root(from))
3307                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3308                         mem_cgroup_put(from);
3309                         /*
3310                          * we charged both to->res and to->memsw, so we should
3311                          * uncharge to->res.
3312                          */
3313                         if (!mem_cgroup_is_root(to))
3314                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3315                 }
3316                 return 0;
3317         }
3318         return -EINVAL;
3319 }
3320 #else
3321 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3322                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3323 {
3324         return -EINVAL;
3325 }
3326 #endif
3327
3328 /*
3329  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3330  * page belongs to.
3331  */
3332 int mem_cgroup_prepare_migration(struct page *page,
3333         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3334 {
3335         struct mem_cgroup *mem = NULL;
3336         struct page_cgroup *pc;
3337         enum charge_type ctype;
3338         int ret = 0;
3339
3340         *ptr = NULL;
3341
3342         VM_BUG_ON(PageTransHuge(page));
3343         if (mem_cgroup_disabled())
3344                 return 0;
3345
3346         pc = lookup_page_cgroup(page);
3347         lock_page_cgroup(pc);
3348         if (PageCgroupUsed(pc)) {
3349                 mem = pc->mem_cgroup;
3350                 css_get(&mem->css);
3351                 /*
3352                  * At migrating an anonymous page, its mapcount goes down
3353                  * to 0 and uncharge() will be called. But, even if it's fully
3354                  * unmapped, migration may fail and this page has to be
3355                  * charged again. We set MIGRATION flag here and delay uncharge
3356                  * until end_migration() is called
3357                  *
3358                  * Corner Case Thinking
3359                  * A)
3360                  * When the old page was mapped as Anon and it's unmap-and-freed
3361                  * while migration was ongoing.
3362                  * If unmap finds the old page, uncharge() of it will be delayed
3363                  * until end_migration(). If unmap finds a new page, it's
3364                  * uncharged when it make mapcount to be 1->0. If unmap code
3365                  * finds swap_migration_entry, the new page will not be mapped
3366                  * and end_migration() will find it(mapcount==0).
3367                  *
3368                  * B)
3369                  * When the old page was mapped but migraion fails, the kernel
3370                  * remaps it. A charge for it is kept by MIGRATION flag even
3371                  * if mapcount goes down to 0. We can do remap successfully
3372                  * without charging it again.
3373                  *
3374                  * C)
3375                  * The "old" page is under lock_page() until the end of
3376                  * migration, so, the old page itself will not be swapped-out.
3377                  * If the new page is swapped out before end_migraton, our
3378                  * hook to usual swap-out path will catch the event.
3379                  */
3380                 if (PageAnon(page))
3381                         SetPageCgroupMigration(pc);
3382         }
3383         unlock_page_cgroup(pc);
3384         /*
3385          * If the page is not charged at this point,
3386          * we return here.
3387          */
3388         if (!mem)
3389                 return 0;
3390
3391         *ptr = mem;
3392         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3393         css_put(&mem->css);/* drop extra refcnt */
3394         if (ret || *ptr == NULL) {
3395                 if (PageAnon(page)) {
3396                         lock_page_cgroup(pc);
3397                         ClearPageCgroupMigration(pc);
3398                         unlock_page_cgroup(pc);
3399                         /*
3400                          * The old page may be fully unmapped while we kept it.
3401                          */
3402                         mem_cgroup_uncharge_page(page);
3403                 }
3404                 return -ENOMEM;
3405         }
3406         /*
3407          * We charge new page before it's used/mapped. So, even if unlock_page()
3408          * is called before end_migration, we can catch all events on this new
3409          * page. In the case new page is migrated but not remapped, new page's
3410          * mapcount will be finally 0 and we call uncharge in end_migration().
3411          */
3412         pc = lookup_page_cgroup(newpage);
3413         if (PageAnon(page))
3414                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3415         else if (page_is_file_cache(page))
3416                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3417         else
3418                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3419         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3420         return ret;
3421 }
3422
3423 /* remove redundant charge if migration failed*/
3424 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3425         struct page *oldpage, struct page *newpage, bool migration_ok)
3426 {
3427         struct page *used, *unused;
3428         struct page_cgroup *pc;
3429
3430         if (!mem)
3431                 return;
3432         /* blocks rmdir() */
3433         cgroup_exclude_rmdir(&mem->css);
3434         if (!migration_ok) {
3435                 used = oldpage;
3436                 unused = newpage;
3437         } else {
3438                 used = newpage;
3439                 unused = oldpage;
3440         }
3441         /*
3442          * We disallowed uncharge of pages under migration because mapcount
3443          * of the page goes down to zero, temporarly.
3444          * Clear the flag and check the page should be charged.
3445          */
3446         pc = lookup_page_cgroup(oldpage);
3447         lock_page_cgroup(pc);
3448         ClearPageCgroupMigration(pc);
3449         unlock_page_cgroup(pc);
3450
3451         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3452
3453         /*
3454          * If a page is a file cache, radix-tree replacement is very atomic
3455          * and we can skip this check. When it was an Anon page, its mapcount
3456          * goes down to 0. But because we added MIGRATION flage, it's not
3457          * uncharged yet. There are several case but page->mapcount check
3458          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3459          * check. (see prepare_charge() also)
3460          */
3461         if (PageAnon(used))
3462                 mem_cgroup_uncharge_page(used);
3463         /*
3464          * At migration, we may charge account against cgroup which has no
3465          * tasks.
3466          * So, rmdir()->pre_destroy() can be called while we do this charge.
3467          * In that case, we need to call pre_destroy() again. check it here.
3468          */
3469         cgroup_release_and_wakeup_rmdir(&mem->css);
3470 }
3471
3472 #ifdef CONFIG_DEBUG_VM
3473 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3474 {
3475         struct page_cgroup *pc;
3476
3477         pc = lookup_page_cgroup(page);
3478         if (likely(pc) && PageCgroupUsed(pc))
3479                 return pc;
3480         return NULL;
3481 }
3482
3483 bool mem_cgroup_bad_page_check(struct page *page)
3484 {
3485         if (mem_cgroup_disabled())
3486                 return false;
3487
3488         return lookup_page_cgroup_used(page) != NULL;
3489 }
3490
3491 void mem_cgroup_print_bad_page(struct page *page)
3492 {
3493         struct page_cgroup *pc;
3494
3495         pc = lookup_page_cgroup_used(page);
3496         if (pc) {
3497                 int ret = -1;
3498                 char *path;
3499
3500                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3501                        pc, pc->flags, pc->mem_cgroup);
3502
3503                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3504                 if (path) {
3505                         rcu_read_lock();
3506                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3507                                                         path, PATH_MAX);
3508                         rcu_read_unlock();
3509                 }
3510
3511                 printk(KERN_CONT "(%s)\n",
3512                                 (ret < 0) ? "cannot get the path" : path);
3513                 kfree(path);
3514         }
3515 }
3516 #endif
3517
3518 static DEFINE_MUTEX(set_limit_mutex);
3519
3520 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3521                                 unsigned long long val)
3522 {
3523         int retry_count;
3524         u64 memswlimit, memlimit;
3525         int ret = 0;
3526         int children = mem_cgroup_count_children(memcg);
3527         u64 curusage, oldusage;
3528         int enlarge;
3529
3530         /*
3531          * For keeping hierarchical_reclaim simple, how long we should retry
3532          * is depends on callers. We set our retry-count to be function
3533          * of # of children which we should visit in this loop.
3534          */
3535         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3536
3537         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3538
3539         enlarge = 0;
3540         while (retry_count) {
3541                 if (signal_pending(current)) {
3542                         ret = -EINTR;
3543                         break;
3544                 }
3545                 /*
3546                  * Rather than hide all in some function, I do this in
3547                  * open coded manner. You see what this really does.
3548                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3549                  */
3550                 mutex_lock(&set_limit_mutex);
3551                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3552                 if (memswlimit < val) {
3553                         ret = -EINVAL;
3554                         mutex_unlock(&set_limit_mutex);
3555                         break;
3556                 }
3557
3558                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3559                 if (memlimit < val)
3560                         enlarge = 1;
3561
3562                 ret = res_counter_set_limit(&memcg->res, val);
3563                 if (!ret) {
3564                         if (memswlimit == val)
3565                                 memcg->memsw_is_minimum = true;
3566                         else
3567                                 memcg->memsw_is_minimum = false;
3568                 }
3569                 mutex_unlock(&set_limit_mutex);
3570
3571                 if (!ret)
3572                         break;
3573
3574                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3575                                                 MEM_CGROUP_RECLAIM_SHRINK,
3576                                                 NULL);
3577                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3578                 /* Usage is reduced ? */
3579                 if (curusage >= oldusage)
3580                         retry_count--;
3581                 else
3582                         oldusage = curusage;
3583         }
3584         if (!ret && enlarge)
3585                 memcg_oom_recover(memcg);
3586
3587         return ret;
3588 }
3589
3590 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3591                                         unsigned long long val)
3592 {
3593         int retry_count;
3594         u64 memlimit, memswlimit, oldusage, curusage;
3595         int children = mem_cgroup_count_children(memcg);
3596         int ret = -EBUSY;
3597         int enlarge = 0;
3598
3599         /* see mem_cgroup_resize_res_limit */
3600         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3601         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3602         while (retry_count) {
3603                 if (signal_pending(current)) {
3604                         ret = -EINTR;
3605                         break;
3606                 }
3607                 /*
3608                  * Rather than hide all in some function, I do this in
3609                  * open coded manner. You see what this really does.
3610                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3611                  */
3612                 mutex_lock(&set_limit_mutex);
3613                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3614                 if (memlimit > val) {
3615                         ret = -EINVAL;
3616                         mutex_unlock(&set_limit_mutex);
3617                         break;
3618                 }
3619                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3620                 if (memswlimit < val)
3621                         enlarge = 1;
3622                 ret = res_counter_set_limit(&memcg->memsw, val);
3623                 if (!ret) {
3624                         if (memlimit == val)
3625                                 memcg->memsw_is_minimum = true;
3626                         else
3627                                 memcg->memsw_is_minimum = false;
3628                 }
3629                 mutex_unlock(&set_limit_mutex);
3630
3631                 if (!ret)
3632                         break;
3633
3634                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3635                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3636                                                 MEM_CGROUP_RECLAIM_SHRINK,
3637                                                 NULL);
3638                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3639                 /* Usage is reduced ? */
3640                 if (curusage >= oldusage)
3641                         retry_count--;
3642                 else
3643                         oldusage = curusage;
3644         }
3645         if (!ret && enlarge)
3646                 memcg_oom_recover(memcg);
3647         return ret;
3648 }
3649
3650 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3651                                             gfp_t gfp_mask,
3652                                             unsigned long *total_scanned)
3653 {
3654         unsigned long nr_reclaimed = 0;
3655         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3656         unsigned long reclaimed;
3657         int loop = 0;
3658         struct mem_cgroup_tree_per_zone *mctz;
3659         unsigned long long excess;
3660         unsigned long nr_scanned;
3661
3662         if (order > 0)
3663                 return 0;
3664
3665         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3666         /*
3667          * This loop can run a while, specially if mem_cgroup's continuously
3668          * keep exceeding their soft limit and putting the system under
3669          * pressure
3670          */
3671         do {
3672                 if (next_mz)
3673                         mz = next_mz;
3674                 else
3675                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3676                 if (!mz)
3677                         break;
3678
3679                 nr_scanned = 0;
3680                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3681                                                 gfp_mask,
3682                                                 MEM_CGROUP_RECLAIM_SOFT,
3683                                                 &nr_scanned);
3684                 nr_reclaimed += reclaimed;
3685                 *total_scanned += nr_scanned;
3686                 spin_lock(&mctz->lock);
3687
3688                 /*
3689                  * If we failed to reclaim anything from this memory cgroup
3690                  * it is time to move on to the next cgroup
3691                  */
3692                 next_mz = NULL;
3693                 if (!reclaimed) {
3694                         do {
3695                                 /*
3696                                  * Loop until we find yet another one.
3697                                  *
3698                                  * By the time we get the soft_limit lock
3699                                  * again, someone might have aded the
3700                                  * group back on the RB tree. Iterate to
3701                                  * make sure we get a different mem.
3702                                  * mem_cgroup_largest_soft_limit_node returns
3703                                  * NULL if no other cgroup is present on
3704                                  * the tree
3705                                  */
3706                                 next_mz =
3707                                 __mem_cgroup_largest_soft_limit_node(mctz);
3708                                 if (next_mz == mz)
3709                                         css_put(&next_mz->mem->css);
3710                                 else /* next_mz == NULL or other memcg */
3711                                         break;
3712                         } while (1);
3713                 }
3714                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3715                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3716                 /*
3717                  * One school of thought says that we should not add
3718                  * back the node to the tree if reclaim returns 0.
3719                  * But our reclaim could return 0, simply because due
3720                  * to priority we are exposing a smaller subset of
3721                  * memory to reclaim from. Consider this as a longer
3722                  * term TODO.
3723                  */
3724                 /* If excess == 0, no tree ops */
3725                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3726                 spin_unlock(&mctz->lock);
3727                 css_put(&mz->mem->css);
3728                 loop++;
3729                 /*
3730                  * Could not reclaim anything and there are no more
3731                  * mem cgroups to try or we seem to be looping without
3732                  * reclaiming anything.
3733                  */
3734                 if (!nr_reclaimed &&
3735                         (next_mz == NULL ||
3736                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3737                         break;
3738         } while (!nr_reclaimed);
3739         if (next_mz)
3740                 css_put(&next_mz->mem->css);
3741         return nr_reclaimed;
3742 }
3743
3744 /*
3745  * This routine traverse page_cgroup in given list and drop them all.
3746  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3747  */
3748 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3749                                 int node, int zid, enum lru_list lru)
3750 {
3751         struct zone *zone;
3752         struct mem_cgroup_per_zone *mz;
3753         struct page_cgroup *pc, *busy;
3754         unsigned long flags, loop;
3755         struct list_head *list;
3756         int ret = 0;
3757
3758         zone = &NODE_DATA(node)->node_zones[zid];
3759         mz = mem_cgroup_zoneinfo(mem, node, zid);
3760         list = &mz->lists[lru];
3761
3762         loop = MEM_CGROUP_ZSTAT(mz, lru);
3763         /* give some margin against EBUSY etc...*/
3764         loop += 256;
3765         busy = NULL;
3766         while (loop--) {
3767                 struct page *page;
3768
3769                 ret = 0;
3770                 spin_lock_irqsave(&zone->lru_lock, flags);
3771                 if (list_empty(list)) {
3772                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3773                         break;
3774                 }
3775                 pc = list_entry(list->prev, struct page_cgroup, lru);
3776                 if (busy == pc) {
3777                         list_move(&pc->lru, list);
3778                         busy = NULL;
3779                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3780                         continue;
3781                 }
3782                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3783
3784                 page = lookup_cgroup_page(pc);
3785
3786                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3787                 if (ret == -ENOMEM)
3788                         break;
3789
3790                 if (ret == -EBUSY || ret == -EINVAL) {
3791                         /* found lock contention or "pc" is obsolete. */
3792                         busy = pc;
3793                         cond_resched();
3794                 } else
3795                         busy = NULL;
3796         }
3797
3798         if (!ret && !list_empty(list))
3799                 return -EBUSY;
3800         return ret;
3801 }
3802
3803 /*
3804  * make mem_cgroup's charge to be 0 if there is no task.
3805  * This enables deleting this mem_cgroup.
3806  */
3807 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3808 {
3809         int ret;
3810         int node, zid, shrink;
3811         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3812         struct cgroup *cgrp = mem->css.cgroup;
3813
3814         css_get(&mem->css);
3815
3816         shrink = 0;
3817         /* should free all ? */
3818         if (free_all)
3819                 goto try_to_free;
3820 move_account:
3821         do {
3822                 ret = -EBUSY;
3823                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3824                         goto out;
3825                 ret = -EINTR;
3826                 if (signal_pending(current))
3827                         goto out;
3828                 /* This is for making all *used* pages to be on LRU. */
3829                 lru_add_drain_all();
3830                 drain_all_stock_sync(mem);
3831                 ret = 0;
3832                 mem_cgroup_start_move(mem);
3833                 for_each_node_state(node, N_HIGH_MEMORY) {
3834                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3835                                 enum lru_list l;
3836                                 for_each_lru(l) {
3837                                         ret = mem_cgroup_force_empty_list(mem,
3838                                                         node, zid, l);
3839                                         if (ret)
3840                                                 break;
3841                                 }
3842                         }
3843                         if (ret)
3844                                 break;
3845                 }
3846                 mem_cgroup_end_move(mem);
3847                 memcg_oom_recover(mem);
3848                 /* it seems parent cgroup doesn't have enough mem */
3849                 if (ret == -ENOMEM)
3850                         goto try_to_free;
3851                 cond_resched();
3852         /* "ret" should also be checked to ensure all lists are empty. */
3853         } while (mem->res.usage > 0 || ret);
3854 out:
3855         css_put(&mem->css);
3856         return ret;
3857
3858 try_to_free:
3859         /* returns EBUSY if there is a task or if we come here twice. */
3860         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3861                 ret = -EBUSY;
3862                 goto out;
3863         }
3864         /* we call try-to-free pages for make this cgroup empty */
3865         lru_add_drain_all();
3866         /* try to free all pages in this cgroup */
3867         shrink = 1;
3868         while (nr_retries && mem->res.usage > 0) {
3869                 struct memcg_scanrecord rec;
3870                 int progress;
3871
3872                 if (signal_pending(current)) {
3873                         ret = -EINTR;
3874                         goto out;
3875                 }
3876                 rec.context = SCAN_BY_SHRINK;
3877                 rec.mem = mem;
3878                 rec.root = mem;
3879                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3880                                                 false, &rec);
3881                 if (!progress) {
3882                         nr_retries--;
3883                         /* maybe some writeback is necessary */
3884                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3885                 }
3886
3887         }
3888         lru_add_drain();
3889         /* try move_account...there may be some *locked* pages. */
3890         goto move_account;
3891 }
3892
3893 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3894 {
3895         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3896 }
3897
3898
3899 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3900 {
3901         return mem_cgroup_from_cont(cont)->use_hierarchy;
3902 }
3903
3904 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3905                                         u64 val)
3906 {
3907         int retval = 0;
3908         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3909         struct cgroup *parent = cont->parent;
3910         struct mem_cgroup *parent_mem = NULL;
3911
3912         if (parent)
3913                 parent_mem = mem_cgroup_from_cont(parent);
3914
3915         cgroup_lock();
3916         /*
3917          * If parent's use_hierarchy is set, we can't make any modifications
3918          * in the child subtrees. If it is unset, then the change can
3919          * occur, provided the current cgroup has no children.
3920          *
3921          * For the root cgroup, parent_mem is NULL, we allow value to be
3922          * set if there are no children.
3923          */
3924         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3925                                 (val == 1 || val == 0)) {
3926                 if (list_empty(&cont->children))
3927                         mem->use_hierarchy = val;
3928                 else
3929                         retval = -EBUSY;
3930         } else
3931                 retval = -EINVAL;
3932         cgroup_unlock();
3933
3934         return retval;
3935 }
3936
3937
3938 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3939                                                enum mem_cgroup_stat_index idx)
3940 {
3941         struct mem_cgroup *iter;
3942         long val = 0;
3943
3944         /* Per-cpu values can be negative, use a signed accumulator */
3945         for_each_mem_cgroup_tree(iter, mem)
3946                 val += mem_cgroup_read_stat(iter, idx);
3947
3948         if (val < 0) /* race ? */
3949                 val = 0;
3950         return val;
3951 }
3952
3953 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3954 {
3955         u64 val;
3956
3957         if (!mem_cgroup_is_root(mem)) {
3958                 if (!swap)
3959                         return res_counter_read_u64(&mem->res, RES_USAGE);
3960                 else
3961                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3962         }
3963
3964         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3965         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3966
3967         if (swap)
3968                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3969
3970         return val << PAGE_SHIFT;
3971 }
3972
3973 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3974 {
3975         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3976         u64 val;
3977         int type, name;
3978
3979         type = MEMFILE_TYPE(cft->private);
3980         name = MEMFILE_ATTR(cft->private);
3981         switch (type) {
3982         case _MEM:
3983                 if (name == RES_USAGE)
3984                         val = mem_cgroup_usage(mem, false);
3985                 else
3986                         val = res_counter_read_u64(&mem->res, name);
3987                 break;
3988         case _MEMSWAP:
3989                 if (name == RES_USAGE)
3990                         val = mem_cgroup_usage(mem, true);
3991                 else
3992                         val = res_counter_read_u64(&mem->memsw, name);
3993                 break;
3994         default:
3995                 BUG();
3996                 break;
3997         }
3998         return val;
3999 }
4000 /*
4001  * The user of this function is...
4002  * RES_LIMIT.
4003  */
4004 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4005                             const char *buffer)
4006 {
4007         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4008         int type, name;
4009         unsigned long long val;
4010         int ret;
4011
4012         type = MEMFILE_TYPE(cft->private);
4013         name = MEMFILE_ATTR(cft->private);
4014         switch (name) {
4015         case RES_LIMIT:
4016                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4017                         ret = -EINVAL;
4018                         break;
4019                 }
4020                 /* This function does all necessary parse...reuse it */
4021                 ret = res_counter_memparse_write_strategy(buffer, &val);
4022                 if (ret)
4023                         break;
4024                 if (type == _MEM)
4025                         ret = mem_cgroup_resize_limit(memcg, val);
4026                 else
4027                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
4028                 break;
4029         case RES_SOFT_LIMIT:
4030                 ret = res_counter_memparse_write_strategy(buffer, &val);
4031                 if (ret)
4032                         break;
4033                 /*
4034                  * For memsw, soft limits are hard to implement in terms
4035                  * of semantics, for now, we support soft limits for
4036                  * control without swap
4037                  */
4038                 if (type == _MEM)
4039                         ret = res_counter_set_soft_limit(&memcg->res, val);
4040                 else
4041                         ret = -EINVAL;
4042                 break;
4043         default:
4044                 ret = -EINVAL; /* should be BUG() ? */
4045                 break;
4046         }
4047         return ret;
4048 }
4049
4050 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4051                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4052 {
4053         struct cgroup *cgroup;
4054         unsigned long long min_limit, min_memsw_limit, tmp;
4055
4056         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4057         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4058         cgroup = memcg->css.cgroup;
4059         if (!memcg->use_hierarchy)
4060                 goto out;
4061
4062         while (cgroup->parent) {
4063                 cgroup = cgroup->parent;
4064                 memcg = mem_cgroup_from_cont(cgroup);
4065                 if (!memcg->use_hierarchy)
4066                         break;
4067                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4068                 min_limit = min(min_limit, tmp);
4069                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4070                 min_memsw_limit = min(min_memsw_limit, tmp);
4071         }
4072 out:
4073         *mem_limit = min_limit;
4074         *memsw_limit = min_memsw_limit;
4075         return;
4076 }
4077
4078 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4079 {
4080         struct mem_cgroup *mem;
4081         int type, name;
4082
4083         mem = mem_cgroup_from_cont(cont);
4084         type = MEMFILE_TYPE(event);
4085         name = MEMFILE_ATTR(event);
4086         switch (name) {
4087         case RES_MAX_USAGE:
4088                 if (type == _MEM)
4089                         res_counter_reset_max(&mem->res);
4090                 else
4091                         res_counter_reset_max(&mem->memsw);
4092                 break;
4093         case RES_FAILCNT:
4094                 if (type == _MEM)
4095                         res_counter_reset_failcnt(&mem->res);
4096                 else
4097                         res_counter_reset_failcnt(&mem->memsw);
4098                 break;
4099         }
4100
4101         return 0;
4102 }
4103
4104 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4105                                         struct cftype *cft)
4106 {
4107         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4108 }
4109
4110 #ifdef CONFIG_MMU
4111 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4112                                         struct cftype *cft, u64 val)
4113 {
4114         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4115
4116         if (val >= (1 << NR_MOVE_TYPE))
4117                 return -EINVAL;
4118         /*
4119          * We check this value several times in both in can_attach() and
4120          * attach(), so we need cgroup lock to prevent this value from being
4121          * inconsistent.
4122          */
4123         cgroup_lock();
4124         mem->move_charge_at_immigrate = val;
4125         cgroup_unlock();
4126
4127         return 0;
4128 }
4129 #else
4130 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4131                                         struct cftype *cft, u64 val)
4132 {
4133         return -ENOSYS;
4134 }
4135 #endif
4136
4137
4138 /* For read statistics */
4139 enum {
4140         MCS_CACHE,
4141         MCS_RSS,
4142         MCS_FILE_MAPPED,
4143         MCS_PGPGIN,
4144         MCS_PGPGOUT,
4145         MCS_SWAP,
4146         MCS_PGFAULT,
4147         MCS_PGMAJFAULT,
4148         MCS_INACTIVE_ANON,
4149         MCS_ACTIVE_ANON,
4150         MCS_INACTIVE_FILE,
4151         MCS_ACTIVE_FILE,
4152         MCS_UNEVICTABLE,
4153         NR_MCS_STAT,
4154 };
4155
4156 struct mcs_total_stat {
4157         s64 stat[NR_MCS_STAT];
4158 };
4159
4160 struct {
4161         char *local_name;
4162         char *total_name;
4163 } memcg_stat_strings[NR_MCS_STAT] = {
4164         {"cache", "total_cache"},
4165         {"rss", "total_rss"},
4166         {"mapped_file", "total_mapped_file"},
4167         {"pgpgin", "total_pgpgin"},
4168         {"pgpgout", "total_pgpgout"},
4169         {"swap", "total_swap"},
4170         {"pgfault", "total_pgfault"},
4171         {"pgmajfault", "total_pgmajfault"},
4172         {"inactive_anon", "total_inactive_anon"},
4173         {"active_anon", "total_active_anon"},
4174         {"inactive_file", "total_inactive_file"},
4175         {"active_file", "total_active_file"},
4176         {"unevictable", "total_unevictable"}
4177 };
4178
4179
4180 static void
4181 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4182 {
4183         s64 val;
4184
4185         /* per cpu stat */
4186         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4187         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4188         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4189         s->stat[MCS_RSS] += val * PAGE_SIZE;
4190         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4191         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4192         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4193         s->stat[MCS_PGPGIN] += val;
4194         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4195         s->stat[MCS_PGPGOUT] += val;
4196         if (do_swap_account) {
4197                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4198                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4199         }
4200         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4201         s->stat[MCS_PGFAULT] += val;
4202         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4203         s->stat[MCS_PGMAJFAULT] += val;
4204
4205         /* per zone stat */
4206         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4207         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4208         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4209         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4210         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4211         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4212         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4213         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4214         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4215         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4216 }
4217
4218 static void
4219 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4220 {
4221         struct mem_cgroup *iter;
4222
4223         for_each_mem_cgroup_tree(iter, mem)
4224                 mem_cgroup_get_local_stat(iter, s);
4225 }
4226
4227 #ifdef CONFIG_NUMA
4228 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4229 {
4230         int nid;
4231         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4232         unsigned long node_nr;
4233         struct cgroup *cont = m->private;
4234         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4235
4236         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4237         seq_printf(m, "total=%lu", total_nr);
4238         for_each_node_state(nid, N_HIGH_MEMORY) {
4239                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4240                 seq_printf(m, " N%d=%lu", nid, node_nr);
4241         }
4242         seq_putc(m, '\n');
4243
4244         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4245         seq_printf(m, "file=%lu", file_nr);
4246         for_each_node_state(nid, N_HIGH_MEMORY) {
4247                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4248                                 LRU_ALL_FILE);
4249                 seq_printf(m, " N%d=%lu", nid, node_nr);
4250         }
4251         seq_putc(m, '\n');
4252
4253         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4254         seq_printf(m, "anon=%lu", anon_nr);
4255         for_each_node_state(nid, N_HIGH_MEMORY) {
4256                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4257                                 LRU_ALL_ANON);
4258                 seq_printf(m, " N%d=%lu", nid, node_nr);
4259         }
4260         seq_putc(m, '\n');
4261
4262         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4263         seq_printf(m, "unevictable=%lu", unevictable_nr);
4264         for_each_node_state(nid, N_HIGH_MEMORY) {
4265                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4266                                 BIT(LRU_UNEVICTABLE));
4267                 seq_printf(m, " N%d=%lu", nid, node_nr);
4268         }
4269         seq_putc(m, '\n');
4270         return 0;
4271 }
4272 #endif /* CONFIG_NUMA */
4273
4274 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4275                                  struct cgroup_map_cb *cb)
4276 {
4277         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4278         struct mcs_total_stat mystat;
4279         int i;
4280
4281         memset(&mystat, 0, sizeof(mystat));
4282         mem_cgroup_get_local_stat(mem_cont, &mystat);
4283
4284
4285         for (i = 0; i < NR_MCS_STAT; i++) {
4286                 if (i == MCS_SWAP && !do_swap_account)
4287                         continue;
4288                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4289         }
4290
4291         /* Hierarchical information */
4292         {
4293                 unsigned long long limit, memsw_limit;
4294                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4295                 cb->fill(cb, "hierarchical_memory_limit", limit);
4296                 if (do_swap_account)
4297                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4298         }
4299
4300         memset(&mystat, 0, sizeof(mystat));
4301         mem_cgroup_get_total_stat(mem_cont, &mystat);
4302         for (i = 0; i < NR_MCS_STAT; i++) {
4303                 if (i == MCS_SWAP && !do_swap_account)
4304                         continue;
4305                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4306         }
4307
4308 #ifdef CONFIG_DEBUG_VM
4309         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4310
4311         {
4312                 int nid, zid;
4313                 struct mem_cgroup_per_zone *mz;
4314                 unsigned long recent_rotated[2] = {0, 0};
4315                 unsigned long recent_scanned[2] = {0, 0};
4316
4317                 for_each_online_node(nid)
4318                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4319                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4320
4321                                 recent_rotated[0] +=
4322                                         mz->reclaim_stat.recent_rotated[0];
4323                                 recent_rotated[1] +=
4324                                         mz->reclaim_stat.recent_rotated[1];
4325                                 recent_scanned[0] +=
4326                                         mz->reclaim_stat.recent_scanned[0];
4327                                 recent_scanned[1] +=
4328                                         mz->reclaim_stat.recent_scanned[1];
4329                         }
4330                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4331                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4332                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4333                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4334         }
4335 #endif
4336
4337         return 0;
4338 }
4339
4340 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4341 {
4342         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4343
4344         return mem_cgroup_swappiness(memcg);
4345 }
4346
4347 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4348                                        u64 val)
4349 {
4350         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4351         struct mem_cgroup *parent;
4352
4353         if (val > 100)
4354                 return -EINVAL;
4355
4356         if (cgrp->parent == NULL)
4357                 return -EINVAL;
4358
4359         parent = mem_cgroup_from_cont(cgrp->parent);
4360
4361         cgroup_lock();
4362
4363         /* If under hierarchy, only empty-root can set this value */
4364         if ((parent->use_hierarchy) ||
4365             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4366                 cgroup_unlock();
4367                 return -EINVAL;
4368         }
4369
4370         memcg->swappiness = val;
4371
4372         cgroup_unlock();
4373
4374         return 0;
4375 }
4376
4377 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4378 {
4379         struct mem_cgroup_threshold_ary *t;
4380         u64 usage;
4381         int i;
4382
4383         rcu_read_lock();
4384         if (!swap)
4385                 t = rcu_dereference(memcg->thresholds.primary);
4386         else
4387                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4388
4389         if (!t)
4390                 goto unlock;
4391
4392         usage = mem_cgroup_usage(memcg, swap);
4393
4394         /*
4395          * current_threshold points to threshold just below usage.
4396          * If it's not true, a threshold was crossed after last
4397          * call of __mem_cgroup_threshold().
4398          */
4399         i = t->current_threshold;
4400
4401         /*
4402          * Iterate backward over array of thresholds starting from
4403          * current_threshold and check if a threshold is crossed.
4404          * If none of thresholds below usage is crossed, we read
4405          * only one element of the array here.
4406          */
4407         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4408                 eventfd_signal(t->entries[i].eventfd, 1);
4409
4410         /* i = current_threshold + 1 */
4411         i++;
4412
4413         /*
4414          * Iterate forward over array of thresholds starting from
4415          * current_threshold+1 and check if a threshold is crossed.
4416          * If none of thresholds above usage is crossed, we read
4417          * only one element of the array here.
4418          */
4419         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4420                 eventfd_signal(t->entries[i].eventfd, 1);
4421
4422         /* Update current_threshold */
4423         t->current_threshold = i - 1;
4424 unlock:
4425         rcu_read_unlock();
4426 }
4427
4428 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4429 {
4430         while (memcg) {
4431                 __mem_cgroup_threshold(memcg, false);
4432                 if (do_swap_account)
4433                         __mem_cgroup_threshold(memcg, true);
4434
4435                 memcg = parent_mem_cgroup(memcg);
4436         }
4437 }
4438
4439 static int compare_thresholds(const void *a, const void *b)
4440 {
4441         const struct mem_cgroup_threshold *_a = a;
4442         const struct mem_cgroup_threshold *_b = b;
4443
4444         return _a->threshold - _b->threshold;
4445 }
4446
4447 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4448 {
4449         struct mem_cgroup_eventfd_list *ev;
4450
4451         list_for_each_entry(ev, &mem->oom_notify, list)
4452                 eventfd_signal(ev->eventfd, 1);
4453         return 0;
4454 }
4455
4456 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4457 {
4458         struct mem_cgroup *iter;
4459
4460         for_each_mem_cgroup_tree(iter, mem)
4461                 mem_cgroup_oom_notify_cb(iter);
4462 }
4463
4464 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4465         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4466 {
4467         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4468         struct mem_cgroup_thresholds *thresholds;
4469         struct mem_cgroup_threshold_ary *new;
4470         int type = MEMFILE_TYPE(cft->private);
4471         u64 threshold, usage;
4472         int i, size, ret;
4473
4474         ret = res_counter_memparse_write_strategy(args, &threshold);
4475         if (ret)
4476                 return ret;
4477
4478         mutex_lock(&memcg->thresholds_lock);
4479
4480         if (type == _MEM)
4481                 thresholds = &memcg->thresholds;
4482         else if (type == _MEMSWAP)
4483                 thresholds = &memcg->memsw_thresholds;
4484         else
4485                 BUG();
4486
4487         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4488
4489         /* Check if a threshold crossed before adding a new one */
4490         if (thresholds->primary)
4491                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4492
4493         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4494
4495         /* Allocate memory for new array of thresholds */
4496         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4497                         GFP_KERNEL);
4498         if (!new) {
4499                 ret = -ENOMEM;
4500                 goto unlock;
4501         }
4502         new->size = size;
4503
4504         /* Copy thresholds (if any) to new array */
4505         if (thresholds->primary) {
4506                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4507                                 sizeof(struct mem_cgroup_threshold));
4508         }
4509
4510         /* Add new threshold */
4511         new->entries[size - 1].eventfd = eventfd;
4512         new->entries[size - 1].threshold = threshold;
4513
4514         /* Sort thresholds. Registering of new threshold isn't time-critical */
4515         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4516                         compare_thresholds, NULL);
4517
4518         /* Find current threshold */
4519         new->current_threshold = -1;
4520         for (i = 0; i < size; i++) {
4521                 if (new->entries[i].threshold < usage) {
4522                         /*
4523                          * new->current_threshold will not be used until
4524                          * rcu_assign_pointer(), so it's safe to increment
4525                          * it here.
4526                          */
4527                         ++new->current_threshold;
4528                 }
4529         }
4530
4531         /* Free old spare buffer and save old primary buffer as spare */
4532         kfree(thresholds->spare);
4533         thresholds->spare = thresholds->primary;
4534
4535         rcu_assign_pointer(thresholds->primary, new);
4536
4537         /* To be sure that nobody uses thresholds */
4538         synchronize_rcu();
4539
4540 unlock:
4541         mutex_unlock(&memcg->thresholds_lock);
4542
4543         return ret;
4544 }
4545
4546 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4547         struct cftype *cft, struct eventfd_ctx *eventfd)
4548 {
4549         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4550         struct mem_cgroup_thresholds *thresholds;
4551         struct mem_cgroup_threshold_ary *new;
4552         int type = MEMFILE_TYPE(cft->private);
4553         u64 usage;
4554         int i, j, size;
4555
4556         mutex_lock(&memcg->thresholds_lock);
4557         if (type == _MEM)
4558                 thresholds = &memcg->thresholds;
4559         else if (type == _MEMSWAP)
4560                 thresholds = &memcg->memsw_thresholds;
4561         else
4562                 BUG();
4563
4564         /*
4565          * Something went wrong if we trying to unregister a threshold
4566          * if we don't have thresholds
4567          */
4568         BUG_ON(!thresholds);
4569
4570         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4571
4572         /* Check if a threshold crossed before removing */
4573         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4574
4575         /* Calculate new number of threshold */
4576         size = 0;
4577         for (i = 0; i < thresholds->primary->size; i++) {
4578                 if (thresholds->primary->entries[i].eventfd != eventfd)
4579                         size++;
4580         }
4581
4582         new = thresholds->spare;
4583
4584         /* Set thresholds array to NULL if we don't have thresholds */
4585         if (!size) {
4586                 kfree(new);
4587                 new = NULL;
4588                 goto swap_buffers;
4589         }
4590
4591         new->size = size;
4592
4593         /* Copy thresholds and find current threshold */
4594         new->current_threshold = -1;
4595         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4596                 if (thresholds->primary->entries[i].eventfd == eventfd)
4597                         continue;
4598
4599                 new->entries[j] = thresholds->primary->entries[i];
4600                 if (new->entries[j].threshold < usage) {
4601                         /*
4602                          * new->current_threshold will not be used
4603                          * until rcu_assign_pointer(), so it's safe to increment
4604                          * it here.
4605                          */
4606                         ++new->current_threshold;
4607                 }
4608                 j++;
4609         }
4610
4611 swap_buffers:
4612         /* Swap primary and spare array */
4613         thresholds->spare = thresholds->primary;
4614         rcu_assign_pointer(thresholds->primary, new);
4615
4616         /* To be sure that nobody uses thresholds */
4617         synchronize_rcu();
4618
4619         mutex_unlock(&memcg->thresholds_lock);
4620 }
4621
4622 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4623         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4624 {
4625         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4626         struct mem_cgroup_eventfd_list *event;
4627         int type = MEMFILE_TYPE(cft->private);
4628
4629         BUG_ON(type != _OOM_TYPE);
4630         event = kmalloc(sizeof(*event), GFP_KERNEL);
4631         if (!event)
4632                 return -ENOMEM;
4633
4634         spin_lock(&memcg_oom_lock);
4635
4636         event->eventfd = eventfd;
4637         list_add(&event->list, &memcg->oom_notify);
4638
4639         /* already in OOM ? */
4640         if (atomic_read(&memcg->under_oom))
4641                 eventfd_signal(eventfd, 1);
4642         spin_unlock(&memcg_oom_lock);
4643
4644         return 0;
4645 }
4646
4647 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4648         struct cftype *cft, struct eventfd_ctx *eventfd)
4649 {
4650         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4651         struct mem_cgroup_eventfd_list *ev, *tmp;
4652         int type = MEMFILE_TYPE(cft->private);
4653
4654         BUG_ON(type != _OOM_TYPE);
4655
4656         spin_lock(&memcg_oom_lock);
4657
4658         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4659                 if (ev->eventfd == eventfd) {
4660                         list_del(&ev->list);
4661                         kfree(ev);
4662                 }
4663         }
4664
4665         spin_unlock(&memcg_oom_lock);
4666 }
4667
4668 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4669         struct cftype *cft,  struct cgroup_map_cb *cb)
4670 {
4671         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4672
4673         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4674
4675         if (atomic_read(&mem->under_oom))
4676                 cb->fill(cb, "under_oom", 1);
4677         else
4678                 cb->fill(cb, "under_oom", 0);
4679         return 0;
4680 }
4681
4682 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4683         struct cftype *cft, u64 val)
4684 {
4685         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4686         struct mem_cgroup *parent;
4687
4688         /* cannot set to root cgroup and only 0 and 1 are allowed */
4689         if (!cgrp->parent || !((val == 0) || (val == 1)))
4690                 return -EINVAL;
4691
4692         parent = mem_cgroup_from_cont(cgrp->parent);
4693
4694         cgroup_lock();
4695         /* oom-kill-disable is a flag for subhierarchy. */
4696         if ((parent->use_hierarchy) ||
4697             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4698                 cgroup_unlock();
4699                 return -EINVAL;
4700         }
4701         mem->oom_kill_disable = val;
4702         if (!val)
4703                 memcg_oom_recover(mem);
4704         cgroup_unlock();
4705         return 0;
4706 }
4707
4708 #ifdef CONFIG_NUMA
4709 static const struct file_operations mem_control_numa_stat_file_operations = {
4710         .read = seq_read,
4711         .llseek = seq_lseek,
4712         .release = single_release,
4713 };
4714
4715 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4716 {
4717         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4718
4719         file->f_op = &mem_control_numa_stat_file_operations;
4720         return single_open(file, mem_control_numa_stat_show, cont);
4721 }
4722 #endif /* CONFIG_NUMA */
4723
4724 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4725                                 struct cftype *cft,
4726                                 struct cgroup_map_cb *cb)
4727 {
4728         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4729         char string[64];
4730         int i;
4731
4732         for (i = 0; i < NR_SCANSTATS; i++) {
4733                 strcpy(string, scanstat_string[i]);
4734                 strcat(string, SCANSTAT_WORD_LIMIT);
4735                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4736         }
4737
4738         for (i = 0; i < NR_SCANSTATS; i++) {
4739                 strcpy(string, scanstat_string[i]);
4740                 strcat(string, SCANSTAT_WORD_SYSTEM);
4741                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4742         }
4743
4744         for (i = 0; i < NR_SCANSTATS; i++) {
4745                 strcpy(string, scanstat_string[i]);
4746                 strcat(string, SCANSTAT_WORD_LIMIT);
4747                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4748                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4749         }
4750         for (i = 0; i < NR_SCANSTATS; i++) {
4751                 strcpy(string, scanstat_string[i]);
4752                 strcat(string, SCANSTAT_WORD_SYSTEM);
4753                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4754                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4755         }
4756         return 0;
4757 }
4758
4759 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4760                                 unsigned int event)
4761 {
4762         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4763
4764         spin_lock(&mem->scanstat.lock);
4765         memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4766         memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4767         spin_unlock(&mem->scanstat.lock);
4768         return 0;
4769 }
4770
4771
4772 static struct cftype mem_cgroup_files[] = {
4773         {
4774                 .name = "usage_in_bytes",
4775                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4776                 .read_u64 = mem_cgroup_read,
4777                 .register_event = mem_cgroup_usage_register_event,
4778                 .unregister_event = mem_cgroup_usage_unregister_event,
4779         },
4780         {
4781                 .name = "max_usage_in_bytes",
4782                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4783                 .trigger = mem_cgroup_reset,
4784                 .read_u64 = mem_cgroup_read,
4785         },
4786         {
4787                 .name = "limit_in_bytes",
4788                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4789                 .write_string = mem_cgroup_write,
4790                 .read_u64 = mem_cgroup_read,
4791         },
4792         {
4793                 .name = "soft_limit_in_bytes",
4794                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4795                 .write_string = mem_cgroup_write,
4796                 .read_u64 = mem_cgroup_read,
4797         },
4798         {
4799                 .name = "failcnt",
4800                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4801                 .trigger = mem_cgroup_reset,
4802                 .read_u64 = mem_cgroup_read,
4803         },
4804         {
4805                 .name = "stat",
4806                 .read_map = mem_control_stat_show,
4807         },
4808         {
4809                 .name = "force_empty",
4810                 .trigger = mem_cgroup_force_empty_write,
4811         },
4812         {
4813                 .name = "use_hierarchy",
4814                 .write_u64 = mem_cgroup_hierarchy_write,
4815                 .read_u64 = mem_cgroup_hierarchy_read,
4816         },
4817         {
4818                 .name = "swappiness",
4819                 .read_u64 = mem_cgroup_swappiness_read,
4820                 .write_u64 = mem_cgroup_swappiness_write,
4821         },
4822         {
4823                 .name = "move_charge_at_immigrate",
4824                 .read_u64 = mem_cgroup_move_charge_read,
4825                 .write_u64 = mem_cgroup_move_charge_write,
4826         },
4827         {
4828                 .name = "oom_control",
4829                 .read_map = mem_cgroup_oom_control_read,
4830                 .write_u64 = mem_cgroup_oom_control_write,
4831                 .register_event = mem_cgroup_oom_register_event,
4832                 .unregister_event = mem_cgroup_oom_unregister_event,
4833                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4834         },
4835 #ifdef CONFIG_NUMA
4836         {
4837                 .name = "numa_stat",
4838                 .open = mem_control_numa_stat_open,
4839                 .mode = S_IRUGO,
4840         },
4841 #endif
4842         {
4843                 .name = "vmscan_stat",
4844                 .read_map = mem_cgroup_vmscan_stat_read,
4845                 .trigger = mem_cgroup_reset_vmscan_stat,
4846         },
4847 };
4848
4849 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4850 static struct cftype memsw_cgroup_files[] = {
4851         {
4852                 .name = "memsw.usage_in_bytes",
4853                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4854                 .read_u64 = mem_cgroup_read,
4855                 .register_event = mem_cgroup_usage_register_event,
4856                 .unregister_event = mem_cgroup_usage_unregister_event,
4857         },
4858         {
4859                 .name = "memsw.max_usage_in_bytes",
4860                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4861                 .trigger = mem_cgroup_reset,
4862                 .read_u64 = mem_cgroup_read,
4863         },
4864         {
4865                 .name = "memsw.limit_in_bytes",
4866                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4867                 .write_string = mem_cgroup_write,
4868                 .read_u64 = mem_cgroup_read,
4869         },
4870         {
4871                 .name = "memsw.failcnt",
4872                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4873                 .trigger = mem_cgroup_reset,
4874                 .read_u64 = mem_cgroup_read,
4875         },
4876 };
4877
4878 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4879 {
4880         if (!do_swap_account)
4881                 return 0;
4882         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4883                                 ARRAY_SIZE(memsw_cgroup_files));
4884 };
4885 #else
4886 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4887 {
4888         return 0;
4889 }
4890 #endif
4891
4892 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4893 {
4894         struct mem_cgroup_per_node *pn;
4895         struct mem_cgroup_per_zone *mz;
4896         enum lru_list l;
4897         int zone, tmp = node;
4898         /*
4899          * This routine is called against possible nodes.
4900          * But it's BUG to call kmalloc() against offline node.
4901          *
4902          * TODO: this routine can waste much memory for nodes which will
4903          *       never be onlined. It's better to use memory hotplug callback
4904          *       function.
4905          */
4906         if (!node_state(node, N_NORMAL_MEMORY))
4907                 tmp = -1;
4908         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4909         if (!pn)
4910                 return 1;
4911
4912         mem->info.nodeinfo[node] = pn;
4913         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4914                 mz = &pn->zoneinfo[zone];
4915                 for_each_lru(l)
4916                         INIT_LIST_HEAD(&mz->lists[l]);
4917                 mz->usage_in_excess = 0;
4918                 mz->on_tree = false;
4919                 mz->mem = mem;
4920         }
4921         return 0;
4922 }
4923
4924 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4925 {
4926         kfree(mem->info.nodeinfo[node]);
4927 }
4928
4929 static struct mem_cgroup *mem_cgroup_alloc(void)
4930 {
4931         struct mem_cgroup *mem;
4932         int size = sizeof(struct mem_cgroup);
4933
4934         /* Can be very big if MAX_NUMNODES is very big */
4935         if (size < PAGE_SIZE)
4936                 mem = kzalloc(size, GFP_KERNEL);
4937         else
4938                 mem = vzalloc(size);
4939
4940         if (!mem)
4941                 return NULL;
4942
4943         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4944         if (!mem->stat)
4945                 goto out_free;
4946         spin_lock_init(&mem->pcp_counter_lock);
4947         return mem;
4948
4949 out_free:
4950         if (size < PAGE_SIZE)
4951                 kfree(mem);
4952         else
4953                 vfree(mem);
4954         return NULL;
4955 }
4956
4957 /*
4958  * At destroying mem_cgroup, references from swap_cgroup can remain.
4959  * (scanning all at force_empty is too costly...)
4960  *
4961  * Instead of clearing all references at force_empty, we remember
4962  * the number of reference from swap_cgroup and free mem_cgroup when
4963  * it goes down to 0.
4964  *
4965  * Removal of cgroup itself succeeds regardless of refs from swap.
4966  */
4967
4968 static void __mem_cgroup_free(struct mem_cgroup *mem)
4969 {
4970         int node;
4971
4972         mem_cgroup_remove_from_trees(mem);
4973         free_css_id(&mem_cgroup_subsys, &mem->css);
4974
4975         for_each_node_state(node, N_POSSIBLE)
4976                 free_mem_cgroup_per_zone_info(mem, node);
4977
4978         free_percpu(mem->stat);
4979         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4980                 kfree(mem);
4981         else
4982                 vfree(mem);
4983 }
4984
4985 static void mem_cgroup_get(struct mem_cgroup *mem)
4986 {
4987         atomic_inc(&mem->refcnt);
4988 }
4989
4990 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4991 {
4992         if (atomic_sub_and_test(count, &mem->refcnt)) {
4993                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4994                 __mem_cgroup_free(mem);
4995                 if (parent)
4996                         mem_cgroup_put(parent);
4997         }
4998 }
4999
5000 static void mem_cgroup_put(struct mem_cgroup *mem)
5001 {
5002         __mem_cgroup_put(mem, 1);
5003 }
5004
5005 /*
5006  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5007  */
5008 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5009 {
5010         if (!mem->res.parent)
5011                 return NULL;
5012         return mem_cgroup_from_res_counter(mem->res.parent, res);
5013 }
5014
5015 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5016 static void __init enable_swap_cgroup(void)
5017 {
5018         if (!mem_cgroup_disabled() && really_do_swap_account)
5019                 do_swap_account = 1;
5020 }
5021 #else
5022 static void __init enable_swap_cgroup(void)
5023 {
5024 }
5025 #endif
5026
5027 static int mem_cgroup_soft_limit_tree_init(void)
5028 {
5029         struct mem_cgroup_tree_per_node *rtpn;
5030         struct mem_cgroup_tree_per_zone *rtpz;
5031         int tmp, node, zone;
5032
5033         for_each_node_state(node, N_POSSIBLE) {
5034                 tmp = node;
5035                 if (!node_state(node, N_NORMAL_MEMORY))
5036                         tmp = -1;
5037                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5038                 if (!rtpn)
5039                         return 1;
5040
5041                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5042
5043                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5044                         rtpz = &rtpn->rb_tree_per_zone[zone];
5045                         rtpz->rb_root = RB_ROOT;
5046                         spin_lock_init(&rtpz->lock);
5047                 }
5048         }
5049         return 0;
5050 }
5051
5052 static struct cgroup_subsys_state * __ref
5053 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5054 {
5055         struct mem_cgroup *mem, *parent;
5056         long error = -ENOMEM;
5057         int node;
5058
5059         mem = mem_cgroup_alloc();
5060         if (!mem)
5061                 return ERR_PTR(error);
5062
5063         for_each_node_state(node, N_POSSIBLE)
5064                 if (alloc_mem_cgroup_per_zone_info(mem, node))
5065                         goto free_out;
5066
5067         /* root ? */
5068         if (cont->parent == NULL) {
5069                 int cpu;
5070                 enable_swap_cgroup();
5071                 parent = NULL;
5072                 root_mem_cgroup = mem;
5073                 if (mem_cgroup_soft_limit_tree_init())
5074                         goto free_out;
5075                 for_each_possible_cpu(cpu) {
5076                         struct memcg_stock_pcp *stock =
5077                                                 &per_cpu(memcg_stock, cpu);
5078                         INIT_WORK(&stock->work, drain_local_stock);
5079                 }
5080                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5081         } else {
5082                 parent = mem_cgroup_from_cont(cont->parent);
5083                 mem->use_hierarchy = parent->use_hierarchy;
5084                 mem->oom_kill_disable = parent->oom_kill_disable;
5085         }
5086
5087         if (parent && parent->use_hierarchy) {
5088                 res_counter_init(&mem->res, &parent->res);
5089                 res_counter_init(&mem->memsw, &parent->memsw);
5090                 /*
5091                  * We increment refcnt of the parent to ensure that we can
5092                  * safely access it on res_counter_charge/uncharge.
5093                  * This refcnt will be decremented when freeing this
5094                  * mem_cgroup(see mem_cgroup_put).
5095                  */
5096                 mem_cgroup_get(parent);
5097         } else {
5098                 res_counter_init(&mem->res, NULL);
5099                 res_counter_init(&mem->memsw, NULL);
5100         }
5101         mem->last_scanned_child = 0;
5102         mem->last_scanned_node = MAX_NUMNODES;
5103         INIT_LIST_HEAD(&mem->oom_notify);
5104
5105         if (parent)
5106                 mem->swappiness = mem_cgroup_swappiness(parent);
5107         atomic_set(&mem->refcnt, 1);
5108         mem->move_charge_at_immigrate = 0;
5109         mutex_init(&mem->thresholds_lock);
5110         spin_lock_init(&mem->scanstat.lock);
5111         return &mem->css;
5112 free_out:
5113         __mem_cgroup_free(mem);
5114         root_mem_cgroup = NULL;
5115         return ERR_PTR(error);
5116 }
5117
5118 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5119                                         struct cgroup *cont)
5120 {
5121         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5122
5123         return mem_cgroup_force_empty(mem, false);
5124 }
5125
5126 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5127                                 struct cgroup *cont)
5128 {
5129         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5130
5131         mem_cgroup_put(mem);
5132 }
5133
5134 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5135                                 struct cgroup *cont)
5136 {
5137         int ret;
5138
5139         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5140                                 ARRAY_SIZE(mem_cgroup_files));
5141
5142         if (!ret)
5143                 ret = register_memsw_files(cont, ss);
5144         return ret;
5145 }
5146
5147 #ifdef CONFIG_MMU
5148 /* Handlers for move charge at task migration. */
5149 #define PRECHARGE_COUNT_AT_ONCE 256
5150 static int mem_cgroup_do_precharge(unsigned long count)
5151 {
5152         int ret = 0;
5153         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5154         struct mem_cgroup *mem = mc.to;
5155
5156         if (mem_cgroup_is_root(mem)) {
5157                 mc.precharge += count;
5158                 /* we don't need css_get for root */
5159                 return ret;
5160         }
5161         /* try to charge at once */
5162         if (count > 1) {
5163                 struct res_counter *dummy;
5164                 /*
5165                  * "mem" cannot be under rmdir() because we've already checked
5166                  * by cgroup_lock_live_cgroup() that it is not removed and we
5167                  * are still under the same cgroup_mutex. So we can postpone
5168                  * css_get().
5169                  */
5170                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5171                         goto one_by_one;
5172                 if (do_swap_account && res_counter_charge(&mem->memsw,
5173                                                 PAGE_SIZE * count, &dummy)) {
5174                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5175                         goto one_by_one;
5176                 }
5177                 mc.precharge += count;
5178                 return ret;
5179         }
5180 one_by_one:
5181         /* fall back to one by one charge */
5182         while (count--) {
5183                 if (signal_pending(current)) {
5184                         ret = -EINTR;
5185                         break;
5186                 }
5187                 if (!batch_count--) {
5188                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5189                         cond_resched();
5190                 }
5191                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5192                 if (ret || !mem)
5193                         /* mem_cgroup_clear_mc() will do uncharge later */
5194                         return -ENOMEM;
5195                 mc.precharge++;
5196         }
5197         return ret;
5198 }
5199
5200 /**
5201  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5202  * @vma: the vma the pte to be checked belongs
5203  * @addr: the address corresponding to the pte to be checked
5204  * @ptent: the pte to be checked
5205  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5206  *
5207  * Returns
5208  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5209  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5210  *     move charge. if @target is not NULL, the page is stored in target->page
5211  *     with extra refcnt got(Callers should handle it).
5212  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5213  *     target for charge migration. if @target is not NULL, the entry is stored
5214  *     in target->ent.
5215  *
5216  * Called with pte lock held.
5217  */
5218 union mc_target {
5219         struct page     *page;
5220         swp_entry_t     ent;
5221 };
5222
5223 enum mc_target_type {
5224         MC_TARGET_NONE, /* not used */
5225         MC_TARGET_PAGE,
5226         MC_TARGET_SWAP,
5227 };
5228
5229 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5230                                                 unsigned long addr, pte_t ptent)
5231 {
5232         struct page *page = vm_normal_page(vma, addr, ptent);
5233
5234         if (!page || !page_mapped(page))
5235                 return NULL;
5236         if (PageAnon(page)) {
5237                 /* we don't move shared anon */
5238                 if (!move_anon() || page_mapcount(page) > 2)
5239                         return NULL;
5240         } else if (!move_file())
5241                 /* we ignore mapcount for file pages */
5242                 return NULL;
5243         if (!get_page_unless_zero(page))
5244                 return NULL;
5245
5246         return page;
5247 }
5248
5249 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5250                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5251 {
5252         int usage_count;
5253         struct page *page = NULL;
5254         swp_entry_t ent = pte_to_swp_entry(ptent);
5255
5256         if (!move_anon() || non_swap_entry(ent))
5257                 return NULL;
5258         usage_count = mem_cgroup_count_swap_user(ent, &page);
5259         if (usage_count > 1) { /* we don't move shared anon */
5260                 if (page)
5261                         put_page(page);
5262                 return NULL;
5263         }
5264         if (do_swap_account)
5265                 entry->val = ent.val;
5266
5267         return page;
5268 }
5269
5270 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5271                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5272 {
5273         struct page *page = NULL;
5274         struct inode *inode;
5275         struct address_space *mapping;
5276         pgoff_t pgoff;
5277
5278         if (!vma->vm_file) /* anonymous vma */
5279                 return NULL;
5280         if (!move_file())
5281                 return NULL;
5282
5283         inode = vma->vm_file->f_path.dentry->d_inode;
5284         mapping = vma->vm_file->f_mapping;
5285         if (pte_none(ptent))
5286                 pgoff = linear_page_index(vma, addr);
5287         else /* pte_file(ptent) is true */
5288                 pgoff = pte_to_pgoff(ptent);
5289
5290         /* page is moved even if it's not RSS of this task(page-faulted). */
5291         page = find_get_page(mapping, pgoff);
5292
5293 #ifdef CONFIG_SWAP
5294         /* shmem/tmpfs may report page out on swap: account for that too. */
5295         if (radix_tree_exceptional_entry(page)) {
5296                 swp_entry_t swap = radix_to_swp_entry(page);
5297                 if (do_swap_account)
5298                         *entry = swap;
5299                 page = find_get_page(&swapper_space, swap.val);
5300         }
5301 #endif
5302         return page;
5303 }
5304
5305 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5306                 unsigned long addr, pte_t ptent, union mc_target *target)
5307 {
5308         struct page *page = NULL;
5309         struct page_cgroup *pc;
5310         int ret = 0;
5311         swp_entry_t ent = { .val = 0 };
5312
5313         if (pte_present(ptent))
5314                 page = mc_handle_present_pte(vma, addr, ptent);
5315         else if (is_swap_pte(ptent))
5316                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5317         else if (pte_none(ptent) || pte_file(ptent))
5318                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5319
5320         if (!page && !ent.val)
5321                 return 0;
5322         if (page) {
5323                 pc = lookup_page_cgroup(page);
5324                 /*
5325                  * Do only loose check w/o page_cgroup lock.
5326                  * mem_cgroup_move_account() checks the pc is valid or not under
5327                  * the lock.
5328                  */
5329                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5330                         ret = MC_TARGET_PAGE;
5331                         if (target)
5332                                 target->page = page;
5333                 }
5334                 if (!ret || !target)
5335                         put_page(page);
5336         }
5337         /* There is a swap entry and a page doesn't exist or isn't charged */
5338         if (ent.val && !ret &&
5339                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5340                 ret = MC_TARGET_SWAP;
5341                 if (target)
5342                         target->ent = ent;
5343         }
5344         return ret;
5345 }
5346
5347 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5348                                         unsigned long addr, unsigned long end,
5349                                         struct mm_walk *walk)
5350 {
5351         struct vm_area_struct *vma = walk->private;
5352         pte_t *pte;
5353         spinlock_t *ptl;
5354
5355         split_huge_page_pmd(walk->mm, pmd);
5356
5357         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5358         for (; addr != end; pte++, addr += PAGE_SIZE)
5359                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5360                         mc.precharge++; /* increment precharge temporarily */
5361         pte_unmap_unlock(pte - 1, ptl);
5362         cond_resched();
5363
5364         return 0;
5365 }
5366
5367 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5368 {
5369         unsigned long precharge;
5370         struct vm_area_struct *vma;
5371
5372         down_read(&mm->mmap_sem);
5373         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5374                 struct mm_walk mem_cgroup_count_precharge_walk = {
5375                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5376                         .mm = mm,
5377                         .private = vma,
5378                 };
5379                 if (is_vm_hugetlb_page(vma))
5380                         continue;
5381                 walk_page_range(vma->vm_start, vma->vm_end,
5382                                         &mem_cgroup_count_precharge_walk);
5383         }
5384         up_read(&mm->mmap_sem);
5385
5386         precharge = mc.precharge;
5387         mc.precharge = 0;
5388
5389         return precharge;
5390 }
5391
5392 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5393 {
5394         unsigned long precharge = mem_cgroup_count_precharge(mm);
5395
5396         VM_BUG_ON(mc.moving_task);
5397         mc.moving_task = current;
5398         return mem_cgroup_do_precharge(precharge);
5399 }
5400
5401 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5402 static void __mem_cgroup_clear_mc(void)
5403 {
5404         struct mem_cgroup *from = mc.from;
5405         struct mem_cgroup *to = mc.to;
5406
5407         /* we must uncharge all the leftover precharges from mc.to */
5408         if (mc.precharge) {
5409                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5410                 mc.precharge = 0;
5411         }
5412         /*
5413          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5414          * we must uncharge here.
5415          */
5416         if (mc.moved_charge) {
5417                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5418                 mc.moved_charge = 0;
5419         }
5420         /* we must fixup refcnts and charges */
5421         if (mc.moved_swap) {
5422                 /* uncharge swap account from the old cgroup */
5423                 if (!mem_cgroup_is_root(mc.from))
5424                         res_counter_uncharge(&mc.from->memsw,
5425                                                 PAGE_SIZE * mc.moved_swap);
5426                 __mem_cgroup_put(mc.from, mc.moved_swap);
5427
5428                 if (!mem_cgroup_is_root(mc.to)) {
5429                         /*
5430                          * we charged both to->res and to->memsw, so we should
5431                          * uncharge to->res.
5432                          */
5433                         res_counter_uncharge(&mc.to->res,
5434                                                 PAGE_SIZE * mc.moved_swap);
5435                 }
5436                 /* we've already done mem_cgroup_get(mc.to) */
5437                 mc.moved_swap = 0;
5438         }
5439         memcg_oom_recover(from);
5440         memcg_oom_recover(to);
5441         wake_up_all(&mc.waitq);
5442 }
5443
5444 static void mem_cgroup_clear_mc(void)
5445 {
5446         struct mem_cgroup *from = mc.from;
5447
5448         /*
5449          * we must clear moving_task before waking up waiters at the end of
5450          * task migration.
5451          */
5452         mc.moving_task = NULL;
5453         __mem_cgroup_clear_mc();
5454         spin_lock(&mc.lock);
5455         mc.from = NULL;
5456         mc.to = NULL;
5457         spin_unlock(&mc.lock);
5458         mem_cgroup_end_move(from);
5459 }
5460
5461 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5462                                 struct cgroup *cgroup,
5463                                 struct task_struct *p)
5464 {
5465         int ret = 0;
5466         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5467
5468         if (mem->move_charge_at_immigrate) {
5469                 struct mm_struct *mm;
5470                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5471
5472                 VM_BUG_ON(from == mem);
5473
5474                 mm = get_task_mm(p);
5475                 if (!mm)
5476                         return 0;
5477                 /* We move charges only when we move a owner of the mm */
5478                 if (mm->owner == p) {
5479                         VM_BUG_ON(mc.from);
5480                         VM_BUG_ON(mc.to);
5481                         VM_BUG_ON(mc.precharge);
5482                         VM_BUG_ON(mc.moved_charge);
5483                         VM_BUG_ON(mc.moved_swap);
5484                         mem_cgroup_start_move(from);
5485                         spin_lock(&mc.lock);
5486                         mc.from = from;
5487                         mc.to = mem;
5488                         spin_unlock(&mc.lock);
5489                         /* We set mc.moving_task later */
5490
5491                         ret = mem_cgroup_precharge_mc(mm);
5492                         if (ret)
5493                                 mem_cgroup_clear_mc();
5494                 }
5495                 mmput(mm);
5496         }
5497         return ret;
5498 }
5499
5500 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5501                                 struct cgroup *cgroup,
5502                                 struct task_struct *p)
5503 {
5504         mem_cgroup_clear_mc();
5505 }
5506
5507 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5508                                 unsigned long addr, unsigned long end,
5509                                 struct mm_walk *walk)
5510 {
5511         int ret = 0;
5512         struct vm_area_struct *vma = walk->private;
5513         pte_t *pte;
5514         spinlock_t *ptl;
5515
5516         split_huge_page_pmd(walk->mm, pmd);
5517 retry:
5518         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5519         for (; addr != end; addr += PAGE_SIZE) {
5520                 pte_t ptent = *(pte++);
5521                 union mc_target target;
5522                 int type;
5523                 struct page *page;
5524                 struct page_cgroup *pc;
5525                 swp_entry_t ent;
5526
5527                 if (!mc.precharge)
5528                         break;
5529
5530                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5531                 switch (type) {
5532                 case MC_TARGET_PAGE:
5533                         page = target.page;
5534                         if (isolate_lru_page(page))
5535                                 goto put;
5536                         pc = lookup_page_cgroup(page);
5537                         if (!mem_cgroup_move_account(page, 1, pc,
5538                                                      mc.from, mc.to, false)) {
5539                                 mc.precharge--;
5540                                 /* we uncharge from mc.from later. */
5541                                 mc.moved_charge++;
5542                         }
5543                         putback_lru_page(page);
5544 put:                    /* is_target_pte_for_mc() gets the page */
5545                         put_page(page);
5546                         break;
5547                 case MC_TARGET_SWAP:
5548                         ent = target.ent;
5549                         if (!mem_cgroup_move_swap_account(ent,
5550                                                 mc.from, mc.to, false)) {
5551                                 mc.precharge--;
5552                                 /* we fixup refcnts and charges later. */
5553                                 mc.moved_swap++;
5554                         }
5555                         break;
5556                 default:
5557                         break;
5558                 }
5559         }
5560         pte_unmap_unlock(pte - 1, ptl);
5561         cond_resched();
5562
5563         if (addr != end) {
5564                 /*
5565                  * We have consumed all precharges we got in can_attach().
5566                  * We try charge one by one, but don't do any additional
5567                  * charges to mc.to if we have failed in charge once in attach()
5568                  * phase.
5569                  */
5570                 ret = mem_cgroup_do_precharge(1);
5571                 if (!ret)
5572                         goto retry;
5573         }
5574
5575         return ret;
5576 }
5577
5578 static void mem_cgroup_move_charge(struct mm_struct *mm)
5579 {
5580         struct vm_area_struct *vma;
5581
5582         lru_add_drain_all();
5583 retry:
5584         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5585                 /*
5586                  * Someone who are holding the mmap_sem might be waiting in
5587                  * waitq. So we cancel all extra charges, wake up all waiters,
5588                  * and retry. Because we cancel precharges, we might not be able
5589                  * to move enough charges, but moving charge is a best-effort
5590                  * feature anyway, so it wouldn't be a big problem.
5591                  */
5592                 __mem_cgroup_clear_mc();
5593                 cond_resched();
5594                 goto retry;
5595         }
5596         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5597                 int ret;
5598                 struct mm_walk mem_cgroup_move_charge_walk = {
5599                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5600                         .mm = mm,
5601                         .private = vma,
5602                 };
5603                 if (is_vm_hugetlb_page(vma))
5604                         continue;
5605                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5606                                                 &mem_cgroup_move_charge_walk);
5607                 if (ret)
5608                         /*
5609                          * means we have consumed all precharges and failed in
5610                          * doing additional charge. Just abandon here.
5611                          */
5612                         break;
5613         }
5614         up_read(&mm->mmap_sem);
5615 }
5616
5617 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5618                                 struct cgroup *cont,
5619                                 struct cgroup *old_cont,
5620                                 struct task_struct *p)
5621 {
5622         struct mm_struct *mm = get_task_mm(p);
5623
5624         if (mm) {
5625                 if (mc.to)
5626                         mem_cgroup_move_charge(mm);
5627                 put_swap_token(mm);
5628                 mmput(mm);
5629         }
5630         if (mc.to)
5631                 mem_cgroup_clear_mc();
5632 }
5633 #else   /* !CONFIG_MMU */
5634 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5635                                 struct cgroup *cgroup,
5636                                 struct task_struct *p)
5637 {
5638         return 0;
5639 }
5640 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5641                                 struct cgroup *cgroup,
5642                                 struct task_struct *p)
5643 {
5644 }
5645 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5646                                 struct cgroup *cont,
5647                                 struct cgroup *old_cont,
5648                                 struct task_struct *p)
5649 {
5650 }
5651 #endif
5652
5653 struct cgroup_subsys mem_cgroup_subsys = {
5654         .name = "memory",
5655         .subsys_id = mem_cgroup_subsys_id,
5656         .create = mem_cgroup_create,
5657         .pre_destroy = mem_cgroup_pre_destroy,
5658         .destroy = mem_cgroup_destroy,
5659         .populate = mem_cgroup_populate,
5660         .can_attach = mem_cgroup_can_attach,
5661         .cancel_attach = mem_cgroup_cancel_attach,
5662         .attach = mem_cgroup_move_task,
5663         .early_init = 0,
5664         .use_id = 1,
5665 };
5666
5667 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5668 static int __init enable_swap_account(char *s)
5669 {
5670         /* consider enabled if no parameter or 1 is given */
5671         if (!strcmp(s, "1"))
5672                 really_do_swap_account = 1;
5673         else if (!strcmp(s, "0"))
5674                 really_do_swap_account = 0;
5675         return 1;
5676 }
5677 __setup("swapaccount=", enable_swap_account);
5678
5679 #endif