memcg: add mem_cgroup_same_or_subtree() helper
[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/shmem_fs.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
45 #include <linux/fs.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
52 #include "internal.h"
53
54 #include <asm/uaccess.h>
55
56 #include <trace/events/vmscan.h>
57
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES      5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
61
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
65
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
69 #else
70 static int really_do_swap_account __initdata = 0;
71 #endif
72
73 #else
74 #define do_swap_account         (0)
75 #endif
76
77
78 /*
79  * Statistics for memory cgroup.
80  */
81 enum mem_cgroup_stat_index {
82         /*
83          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
84          */
85         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
86         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
87         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
88         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
91         MEM_CGROUP_STAT_NSTATS,
92 };
93
94 enum mem_cgroup_events_index {
95         MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
96         MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
97         MEM_CGROUP_EVENTS_COUNT,        /* # of pages paged in/out */
98         MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
99         MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
100         MEM_CGROUP_EVENTS_NSTATS,
101 };
102 /*
103  * Per memcg event counter is incremented at every pagein/pageout. With THP,
104  * it will be incremated by the number of pages. This counter is used for
105  * for trigger some periodic events. This is straightforward and better
106  * than using jiffies etc. to handle periodic memcg event.
107  */
108 enum mem_cgroup_events_target {
109         MEM_CGROUP_TARGET_THRESH,
110         MEM_CGROUP_TARGET_SOFTLIMIT,
111         MEM_CGROUP_TARGET_NUMAINFO,
112         MEM_CGROUP_NTARGETS,
113 };
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET  (1024)
117
118 struct mem_cgroup_stat_cpu {
119         long count[MEM_CGROUP_STAT_NSTATS];
120         unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121         unsigned long targets[MEM_CGROUP_NTARGETS];
122 };
123
124 /*
125  * per-zone information in memory controller.
126  */
127 struct mem_cgroup_per_zone {
128         /*
129          * spin_lock to protect the per cgroup LRU
130          */
131         struct list_head        lists[NR_LRU_LISTS];
132         unsigned long           count[NR_LRU_LISTS];
133
134         struct zone_reclaim_stat reclaim_stat;
135         struct rb_node          tree_node;      /* RB tree node */
136         unsigned long long      usage_in_excess;/* Set to the value by which */
137                                                 /* the soft limit is exceeded*/
138         bool                    on_tree;
139         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
140                                                 /* use container_of        */
141 };
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
144
145 struct mem_cgroup_per_node {
146         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
147 };
148
149 struct mem_cgroup_lru_info {
150         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
151 };
152
153 /*
154  * Cgroups above their limits are maintained in a RB-Tree, independent of
155  * their hierarchy representation
156  */
157
158 struct mem_cgroup_tree_per_zone {
159         struct rb_root rb_root;
160         spinlock_t lock;
161 };
162
163 struct mem_cgroup_tree_per_node {
164         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
165 };
166
167 struct mem_cgroup_tree {
168         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
169 };
170
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
172
173 struct mem_cgroup_threshold {
174         struct eventfd_ctx *eventfd;
175         u64 threshold;
176 };
177
178 /* For threshold */
179 struct mem_cgroup_threshold_ary {
180         /* An array index points to threshold just below usage. */
181         int current_threshold;
182         /* Size of entries[] */
183         unsigned int size;
184         /* Array of thresholds */
185         struct mem_cgroup_threshold entries[0];
186 };
187
188 struct mem_cgroup_thresholds {
189         /* Primary thresholds array */
190         struct mem_cgroup_threshold_ary *primary;
191         /*
192          * Spare threshold array.
193          * This is needed to make mem_cgroup_unregister_event() "never fail".
194          * It must be able to store at least primary->size - 1 entries.
195          */
196         struct mem_cgroup_threshold_ary *spare;
197 };
198
199 /* for OOM */
200 struct mem_cgroup_eventfd_list {
201         struct list_head list;
202         struct eventfd_ctx *eventfd;
203 };
204
205 static void mem_cgroup_threshold(struct mem_cgroup *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
207
208 enum {
209         SCAN_BY_LIMIT,
210         SCAN_BY_SYSTEM,
211         NR_SCAN_CONTEXT,
212         SCAN_BY_SHRINK, /* not recorded now */
213 };
214
215 enum {
216         SCAN,
217         SCAN_ANON,
218         SCAN_FILE,
219         ROTATE,
220         ROTATE_ANON,
221         ROTATE_FILE,
222         FREED,
223         FREED_ANON,
224         FREED_FILE,
225         ELAPSED,
226         NR_SCANSTATS,
227 };
228
229 struct scanstat {
230         spinlock_t      lock;
231         unsigned long   stats[NR_SCAN_CONTEXT][NR_SCANSTATS];
232         unsigned long   rootstats[NR_SCAN_CONTEXT][NR_SCANSTATS];
233 };
234
235 const char *scanstat_string[NR_SCANSTATS] = {
236         "scanned_pages",
237         "scanned_anon_pages",
238         "scanned_file_pages",
239         "rotated_pages",
240         "rotated_anon_pages",
241         "rotated_file_pages",
242         "freed_pages",
243         "freed_anon_pages",
244         "freed_file_pages",
245         "elapsed_ns",
246 };
247 #define SCANSTAT_WORD_LIMIT     "_by_limit"
248 #define SCANSTAT_WORD_SYSTEM    "_by_system"
249 #define SCANSTAT_WORD_HIERARCHY "_under_hierarchy"
250
251
252 /*
253  * The memory controller data structure. The memory controller controls both
254  * page cache and RSS per cgroup. We would eventually like to provide
255  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
256  * to help the administrator determine what knobs to tune.
257  *
258  * TODO: Add a water mark for the memory controller. Reclaim will begin when
259  * we hit the water mark. May be even add a low water mark, such that
260  * no reclaim occurs from a cgroup at it's low water mark, this is
261  * a feature that will be implemented much later in the future.
262  */
263 struct mem_cgroup {
264         struct cgroup_subsys_state css;
265         /*
266          * the counter to account for memory usage
267          */
268         struct res_counter res;
269         /*
270          * the counter to account for mem+swap usage.
271          */
272         struct res_counter memsw;
273         /*
274          * Per cgroup active and inactive list, similar to the
275          * per zone LRU lists.
276          */
277         struct mem_cgroup_lru_info info;
278         /*
279          * While reclaiming in a hierarchy, we cache the last child we
280          * reclaimed from.
281          */
282         int last_scanned_child;
283         int last_scanned_node;
284 #if MAX_NUMNODES > 1
285         nodemask_t      scan_nodes;
286         atomic_t        numainfo_events;
287         atomic_t        numainfo_updating;
288 #endif
289         /*
290          * Should the accounting and control be hierarchical, per subtree?
291          */
292         bool use_hierarchy;
293
294         bool            oom_lock;
295         atomic_t        under_oom;
296
297         atomic_t        refcnt;
298
299         int     swappiness;
300         /* OOM-Killer disable */
301         int             oom_kill_disable;
302
303         /* set when res.limit == memsw.limit */
304         bool            memsw_is_minimum;
305
306         /* protect arrays of thresholds */
307         struct mutex thresholds_lock;
308
309         /* thresholds for memory usage. RCU-protected */
310         struct mem_cgroup_thresholds thresholds;
311
312         /* thresholds for mem+swap usage. RCU-protected */
313         struct mem_cgroup_thresholds memsw_thresholds;
314
315         /* For oom notifier event fd */
316         struct list_head oom_notify;
317         /* For recording LRU-scan statistics */
318         struct scanstat scanstat;
319         /*
320          * Should we move charges of a task when a task is moved into this
321          * mem_cgroup ? And what type of charges should we move ?
322          */
323         unsigned long   move_charge_at_immigrate;
324         /*
325          * percpu counter.
326          */
327         struct mem_cgroup_stat_cpu *stat;
328         /*
329          * used when a cpu is offlined or other synchronizations
330          * See mem_cgroup_read_stat().
331          */
332         struct mem_cgroup_stat_cpu nocpu_base;
333         spinlock_t pcp_counter_lock;
334 };
335
336 /* Stuffs for move charges at task migration. */
337 /*
338  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
339  * left-shifted bitmap of these types.
340  */
341 enum move_type {
342         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
343         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
344         NR_MOVE_TYPE,
345 };
346
347 /* "mc" and its members are protected by cgroup_mutex */
348 static struct move_charge_struct {
349         spinlock_t        lock; /* for from, to */
350         struct mem_cgroup *from;
351         struct mem_cgroup *to;
352         unsigned long precharge;
353         unsigned long moved_charge;
354         unsigned long moved_swap;
355         struct task_struct *moving_task;        /* a task moving charges */
356         wait_queue_head_t waitq;                /* a waitq for other context */
357 } mc = {
358         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
359         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
360 };
361
362 static bool move_anon(void)
363 {
364         return test_bit(MOVE_CHARGE_TYPE_ANON,
365                                         &mc.to->move_charge_at_immigrate);
366 }
367
368 static bool move_file(void)
369 {
370         return test_bit(MOVE_CHARGE_TYPE_FILE,
371                                         &mc.to->move_charge_at_immigrate);
372 }
373
374 /*
375  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
376  * limit reclaim to prevent infinite loops, if they ever occur.
377  */
378 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
379 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
380
381 enum charge_type {
382         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
383         MEM_CGROUP_CHARGE_TYPE_MAPPED,
384         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
385         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
386         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
387         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
388         NR_CHARGE_TYPE,
389 };
390
391 /* for encoding cft->private value on file */
392 #define _MEM                    (0)
393 #define _MEMSWAP                (1)
394 #define _OOM_TYPE               (2)
395 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
396 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
397 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
398 /* Used for OOM nofiier */
399 #define OOM_CONTROL             (0)
400
401 /*
402  * Reclaim flags for mem_cgroup_hierarchical_reclaim
403  */
404 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
405 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
406 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
407 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
408 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
409 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
410
411 static void mem_cgroup_get(struct mem_cgroup *mem);
412 static void mem_cgroup_put(struct mem_cgroup *mem);
413 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
414 static void drain_all_stock_async(struct mem_cgroup *mem);
415
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
418 {
419         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
420 }
421
422 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
423 {
424         return &mem->css;
425 }
426
427 static struct mem_cgroup_per_zone *
428 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
429 {
430         int nid = page_to_nid(page);
431         int zid = page_zonenum(page);
432
433         return mem_cgroup_zoneinfo(mem, nid, zid);
434 }
435
436 static struct mem_cgroup_tree_per_zone *
437 soft_limit_tree_node_zone(int nid, int zid)
438 {
439         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
440 }
441
442 static struct mem_cgroup_tree_per_zone *
443 soft_limit_tree_from_page(struct page *page)
444 {
445         int nid = page_to_nid(page);
446         int zid = page_zonenum(page);
447
448         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
449 }
450
451 static void
452 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
453                                 struct mem_cgroup_per_zone *mz,
454                                 struct mem_cgroup_tree_per_zone *mctz,
455                                 unsigned long long new_usage_in_excess)
456 {
457         struct rb_node **p = &mctz->rb_root.rb_node;
458         struct rb_node *parent = NULL;
459         struct mem_cgroup_per_zone *mz_node;
460
461         if (mz->on_tree)
462                 return;
463
464         mz->usage_in_excess = new_usage_in_excess;
465         if (!mz->usage_in_excess)
466                 return;
467         while (*p) {
468                 parent = *p;
469                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
470                                         tree_node);
471                 if (mz->usage_in_excess < mz_node->usage_in_excess)
472                         p = &(*p)->rb_left;
473                 /*
474                  * We can't avoid mem cgroups that are over their soft
475                  * limit by the same amount
476                  */
477                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
478                         p = &(*p)->rb_right;
479         }
480         rb_link_node(&mz->tree_node, parent, p);
481         rb_insert_color(&mz->tree_node, &mctz->rb_root);
482         mz->on_tree = true;
483 }
484
485 static void
486 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
487                                 struct mem_cgroup_per_zone *mz,
488                                 struct mem_cgroup_tree_per_zone *mctz)
489 {
490         if (!mz->on_tree)
491                 return;
492         rb_erase(&mz->tree_node, &mctz->rb_root);
493         mz->on_tree = false;
494 }
495
496 static void
497 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
498                                 struct mem_cgroup_per_zone *mz,
499                                 struct mem_cgroup_tree_per_zone *mctz)
500 {
501         spin_lock(&mctz->lock);
502         __mem_cgroup_remove_exceeded(mem, mz, mctz);
503         spin_unlock(&mctz->lock);
504 }
505
506
507 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
508 {
509         unsigned long long excess;
510         struct mem_cgroup_per_zone *mz;
511         struct mem_cgroup_tree_per_zone *mctz;
512         int nid = page_to_nid(page);
513         int zid = page_zonenum(page);
514         mctz = soft_limit_tree_from_page(page);
515
516         /*
517          * Necessary to update all ancestors when hierarchy is used.
518          * because their event counter is not touched.
519          */
520         for (; mem; mem = parent_mem_cgroup(mem)) {
521                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
522                 excess = res_counter_soft_limit_excess(&mem->res);
523                 /*
524                  * We have to update the tree if mz is on RB-tree or
525                  * mem is over its softlimit.
526                  */
527                 if (excess || mz->on_tree) {
528                         spin_lock(&mctz->lock);
529                         /* if on-tree, remove it */
530                         if (mz->on_tree)
531                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
532                         /*
533                          * Insert again. mz->usage_in_excess will be updated.
534                          * If excess is 0, no tree ops.
535                          */
536                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
537                         spin_unlock(&mctz->lock);
538                 }
539         }
540 }
541
542 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
543 {
544         int node, zone;
545         struct mem_cgroup_per_zone *mz;
546         struct mem_cgroup_tree_per_zone *mctz;
547
548         for_each_node_state(node, N_POSSIBLE) {
549                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
550                         mz = mem_cgroup_zoneinfo(mem, node, zone);
551                         mctz = soft_limit_tree_node_zone(node, zone);
552                         mem_cgroup_remove_exceeded(mem, mz, mctz);
553                 }
554         }
555 }
556
557 static struct mem_cgroup_per_zone *
558 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
559 {
560         struct rb_node *rightmost = NULL;
561         struct mem_cgroup_per_zone *mz;
562
563 retry:
564         mz = NULL;
565         rightmost = rb_last(&mctz->rb_root);
566         if (!rightmost)
567                 goto done;              /* Nothing to reclaim from */
568
569         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
570         /*
571          * Remove the node now but someone else can add it back,
572          * we will to add it back at the end of reclaim to its correct
573          * position in the tree.
574          */
575         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
576         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
577                 !css_tryget(&mz->mem->css))
578                 goto retry;
579 done:
580         return mz;
581 }
582
583 static struct mem_cgroup_per_zone *
584 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
585 {
586         struct mem_cgroup_per_zone *mz;
587
588         spin_lock(&mctz->lock);
589         mz = __mem_cgroup_largest_soft_limit_node(mctz);
590         spin_unlock(&mctz->lock);
591         return mz;
592 }
593
594 /*
595  * Implementation Note: reading percpu statistics for memcg.
596  *
597  * Both of vmstat[] and percpu_counter has threshold and do periodic
598  * synchronization to implement "quick" read. There are trade-off between
599  * reading cost and precision of value. Then, we may have a chance to implement
600  * a periodic synchronizion of counter in memcg's counter.
601  *
602  * But this _read() function is used for user interface now. The user accounts
603  * memory usage by memory cgroup and he _always_ requires exact value because
604  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
605  * have to visit all online cpus and make sum. So, for now, unnecessary
606  * synchronization is not implemented. (just implemented for cpu hotplug)
607  *
608  * If there are kernel internal actions which can make use of some not-exact
609  * value, and reading all cpu value can be performance bottleneck in some
610  * common workload, threashold and synchonization as vmstat[] should be
611  * implemented.
612  */
613 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
614                                  enum mem_cgroup_stat_index idx)
615 {
616         long val = 0;
617         int cpu;
618
619         get_online_cpus();
620         for_each_online_cpu(cpu)
621                 val += per_cpu(mem->stat->count[idx], cpu);
622 #ifdef CONFIG_HOTPLUG_CPU
623         spin_lock(&mem->pcp_counter_lock);
624         val += mem->nocpu_base.count[idx];
625         spin_unlock(&mem->pcp_counter_lock);
626 #endif
627         put_online_cpus();
628         return val;
629 }
630
631 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
632                                          bool charge)
633 {
634         int val = (charge) ? 1 : -1;
635         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
636 }
637
638 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
639 {
640         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
641 }
642
643 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
644 {
645         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
646 }
647
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
649                                             enum mem_cgroup_events_index idx)
650 {
651         unsigned long val = 0;
652         int cpu;
653
654         for_each_online_cpu(cpu)
655                 val += per_cpu(mem->stat->events[idx], cpu);
656 #ifdef CONFIG_HOTPLUG_CPU
657         spin_lock(&mem->pcp_counter_lock);
658         val += mem->nocpu_base.events[idx];
659         spin_unlock(&mem->pcp_counter_lock);
660 #endif
661         return val;
662 }
663
664 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
665                                          bool file, int nr_pages)
666 {
667         preempt_disable();
668
669         if (file)
670                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
671         else
672                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
673
674         /* pagein of a big page is an event. So, ignore page size */
675         if (nr_pages > 0)
676                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
677         else {
678                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
679                 nr_pages = -nr_pages; /* for event */
680         }
681
682         __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
683
684         preempt_enable();
685 }
686
687 unsigned long
688 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
689                         unsigned int lru_mask)
690 {
691         struct mem_cgroup_per_zone *mz;
692         enum lru_list l;
693         unsigned long ret = 0;
694
695         mz = mem_cgroup_zoneinfo(mem, nid, zid);
696
697         for_each_lru(l) {
698                 if (BIT(l) & lru_mask)
699                         ret += MEM_CGROUP_ZSTAT(mz, l);
700         }
701         return ret;
702 }
703
704 static unsigned long
705 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
706                         int nid, unsigned int lru_mask)
707 {
708         u64 total = 0;
709         int zid;
710
711         for (zid = 0; zid < MAX_NR_ZONES; zid++)
712                 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
713
714         return total;
715 }
716
717 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
718                         unsigned int lru_mask)
719 {
720         int nid;
721         u64 total = 0;
722
723         for_each_node_state(nid, N_HIGH_MEMORY)
724                 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
725         return total;
726 }
727
728 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
729 {
730         unsigned long val, next;
731
732         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
733         next = this_cpu_read(mem->stat->targets[target]);
734         /* from time_after() in jiffies.h */
735         return ((long)next - (long)val < 0);
736 }
737
738 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
739 {
740         unsigned long val, next;
741
742         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
743
744         switch (target) {
745         case MEM_CGROUP_TARGET_THRESH:
746                 next = val + THRESHOLDS_EVENTS_TARGET;
747                 break;
748         case MEM_CGROUP_TARGET_SOFTLIMIT:
749                 next = val + SOFTLIMIT_EVENTS_TARGET;
750                 break;
751         case MEM_CGROUP_TARGET_NUMAINFO:
752                 next = val + NUMAINFO_EVENTS_TARGET;
753                 break;
754         default:
755                 return;
756         }
757
758         this_cpu_write(mem->stat->targets[target], next);
759 }
760
761 /*
762  * Check events in order.
763  *
764  */
765 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
766 {
767         /* threshold event is triggered in finer grain than soft limit */
768         if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
769                 mem_cgroup_threshold(mem);
770                 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
771                 if (unlikely(__memcg_event_check(mem,
772                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
773                         mem_cgroup_update_tree(mem, page);
774                         __mem_cgroup_target_update(mem,
775                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
776                 }
777 #if MAX_NUMNODES > 1
778                 if (unlikely(__memcg_event_check(mem,
779                         MEM_CGROUP_TARGET_NUMAINFO))) {
780                         atomic_inc(&mem->numainfo_events);
781                         __mem_cgroup_target_update(mem,
782                                 MEM_CGROUP_TARGET_NUMAINFO);
783                 }
784 #endif
785         }
786 }
787
788 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
789 {
790         return container_of(cgroup_subsys_state(cont,
791                                 mem_cgroup_subsys_id), struct mem_cgroup,
792                                 css);
793 }
794
795 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
796 {
797         /*
798          * mm_update_next_owner() may clear mm->owner to NULL
799          * if it races with swapoff, page migration, etc.
800          * So this can be called with p == NULL.
801          */
802         if (unlikely(!p))
803                 return NULL;
804
805         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
806                                 struct mem_cgroup, css);
807 }
808
809 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
810 {
811         struct mem_cgroup *mem = NULL;
812
813         if (!mm)
814                 return NULL;
815         /*
816          * Because we have no locks, mm->owner's may be being moved to other
817          * cgroup. We use css_tryget() here even if this looks
818          * pessimistic (rather than adding locks here).
819          */
820         rcu_read_lock();
821         do {
822                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
823                 if (unlikely(!mem))
824                         break;
825         } while (!css_tryget(&mem->css));
826         rcu_read_unlock();
827         return mem;
828 }
829
830 /* The caller has to guarantee "mem" exists before calling this */
831 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
832 {
833         struct cgroup_subsys_state *css;
834         int found;
835
836         if (!mem) /* ROOT cgroup has the smallest ID */
837                 return root_mem_cgroup; /*css_put/get against root is ignored*/
838         if (!mem->use_hierarchy) {
839                 if (css_tryget(&mem->css))
840                         return mem;
841                 return NULL;
842         }
843         rcu_read_lock();
844         /*
845          * searching a memory cgroup which has the smallest ID under given
846          * ROOT cgroup. (ID >= 1)
847          */
848         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
849         if (css && css_tryget(css))
850                 mem = container_of(css, struct mem_cgroup, css);
851         else
852                 mem = NULL;
853         rcu_read_unlock();
854         return mem;
855 }
856
857 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
858                                         struct mem_cgroup *root,
859                                         bool cond)
860 {
861         int nextid = css_id(&iter->css) + 1;
862         int found;
863         int hierarchy_used;
864         struct cgroup_subsys_state *css;
865
866         hierarchy_used = iter->use_hierarchy;
867
868         css_put(&iter->css);
869         /* If no ROOT, walk all, ignore hierarchy */
870         if (!cond || (root && !hierarchy_used))
871                 return NULL;
872
873         if (!root)
874                 root = root_mem_cgroup;
875
876         do {
877                 iter = NULL;
878                 rcu_read_lock();
879
880                 css = css_get_next(&mem_cgroup_subsys, nextid,
881                                 &root->css, &found);
882                 if (css && css_tryget(css))
883                         iter = container_of(css, struct mem_cgroup, css);
884                 rcu_read_unlock();
885                 /* If css is NULL, no more cgroups will be found */
886                 nextid = found + 1;
887         } while (css && !iter);
888
889         return iter;
890 }
891 /*
892  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
893  * be careful that "break" loop is not allowed. We have reference count.
894  * Instead of that modify "cond" to be false and "continue" to exit the loop.
895  */
896 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
897         for (iter = mem_cgroup_start_loop(root);\
898              iter != NULL;\
899              iter = mem_cgroup_get_next(iter, root, cond))
900
901 #define for_each_mem_cgroup_tree(iter, root) \
902         for_each_mem_cgroup_tree_cond(iter, root, true)
903
904 #define for_each_mem_cgroup_all(iter) \
905         for_each_mem_cgroup_tree_cond(iter, NULL, true)
906
907
908 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
909 {
910         return (mem == root_mem_cgroup);
911 }
912
913 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
914 {
915         struct mem_cgroup *mem;
916
917         if (!mm)
918                 return;
919
920         rcu_read_lock();
921         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
922         if (unlikely(!mem))
923                 goto out;
924
925         switch (idx) {
926         case PGMAJFAULT:
927                 mem_cgroup_pgmajfault(mem, 1);
928                 break;
929         case PGFAULT:
930                 mem_cgroup_pgfault(mem, 1);
931                 break;
932         default:
933                 BUG();
934         }
935 out:
936         rcu_read_unlock();
937 }
938 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
939
940 /*
941  * Following LRU functions are allowed to be used without PCG_LOCK.
942  * Operations are called by routine of global LRU independently from memcg.
943  * What we have to take care of here is validness of pc->mem_cgroup.
944  *
945  * Changes to pc->mem_cgroup happens when
946  * 1. charge
947  * 2. moving account
948  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
949  * It is added to LRU before charge.
950  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
951  * When moving account, the page is not on LRU. It's isolated.
952  */
953
954 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
955 {
956         struct page_cgroup *pc;
957         struct mem_cgroup_per_zone *mz;
958
959         if (mem_cgroup_disabled())
960                 return;
961         pc = lookup_page_cgroup(page);
962         /* can happen while we handle swapcache. */
963         if (!TestClearPageCgroupAcctLRU(pc))
964                 return;
965         VM_BUG_ON(!pc->mem_cgroup);
966         /*
967          * We don't check PCG_USED bit. It's cleared when the "page" is finally
968          * removed from global LRU.
969          */
970         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
971         /* huge page split is done under lru_lock. so, we have no races. */
972         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
973         if (mem_cgroup_is_root(pc->mem_cgroup))
974                 return;
975         VM_BUG_ON(list_empty(&pc->lru));
976         list_del_init(&pc->lru);
977 }
978
979 void mem_cgroup_del_lru(struct page *page)
980 {
981         mem_cgroup_del_lru_list(page, page_lru(page));
982 }
983
984 /*
985  * Writeback is about to end against a page which has been marked for immediate
986  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
987  * inactive list.
988  */
989 void mem_cgroup_rotate_reclaimable_page(struct page *page)
990 {
991         struct mem_cgroup_per_zone *mz;
992         struct page_cgroup *pc;
993         enum lru_list lru = page_lru(page);
994
995         if (mem_cgroup_disabled())
996                 return;
997
998         pc = lookup_page_cgroup(page);
999         /* unused or root page is not rotated. */
1000         if (!PageCgroupUsed(pc))
1001                 return;
1002         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1003         smp_rmb();
1004         if (mem_cgroup_is_root(pc->mem_cgroup))
1005                 return;
1006         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1007         list_move_tail(&pc->lru, &mz->lists[lru]);
1008 }
1009
1010 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1011 {
1012         struct mem_cgroup_per_zone *mz;
1013         struct page_cgroup *pc;
1014
1015         if (mem_cgroup_disabled())
1016                 return;
1017
1018         pc = lookup_page_cgroup(page);
1019         /* unused or root page is not rotated. */
1020         if (!PageCgroupUsed(pc))
1021                 return;
1022         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1023         smp_rmb();
1024         if (mem_cgroup_is_root(pc->mem_cgroup))
1025                 return;
1026         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1027         list_move(&pc->lru, &mz->lists[lru]);
1028 }
1029
1030 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1031 {
1032         struct page_cgroup *pc;
1033         struct mem_cgroup_per_zone *mz;
1034
1035         if (mem_cgroup_disabled())
1036                 return;
1037         pc = lookup_page_cgroup(page);
1038         VM_BUG_ON(PageCgroupAcctLRU(pc));
1039         if (!PageCgroupUsed(pc))
1040                 return;
1041         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1042         smp_rmb();
1043         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1044         /* huge page split is done under lru_lock. so, we have no races. */
1045         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1046         SetPageCgroupAcctLRU(pc);
1047         if (mem_cgroup_is_root(pc->mem_cgroup))
1048                 return;
1049         list_add(&pc->lru, &mz->lists[lru]);
1050 }
1051
1052 /*
1053  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1054  * while it's linked to lru because the page may be reused after it's fully
1055  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1056  * It's done under lock_page and expected that zone->lru_lock isnever held.
1057  */
1058 static void mem_cgroup_lru_del_before_commit(struct page *page)
1059 {
1060         unsigned long flags;
1061         struct zone *zone = page_zone(page);
1062         struct page_cgroup *pc = lookup_page_cgroup(page);
1063
1064         /*
1065          * Doing this check without taking ->lru_lock seems wrong but this
1066          * is safe. Because if page_cgroup's USED bit is unset, the page
1067          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1068          * set, the commit after this will fail, anyway.
1069          * This all charge/uncharge is done under some mutual execustion.
1070          * So, we don't need to taking care of changes in USED bit.
1071          */
1072         if (likely(!PageLRU(page)))
1073                 return;
1074
1075         spin_lock_irqsave(&zone->lru_lock, flags);
1076         /*
1077          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1078          * is guarded by lock_page() because the page is SwapCache.
1079          */
1080         if (!PageCgroupUsed(pc))
1081                 mem_cgroup_del_lru_list(page, page_lru(page));
1082         spin_unlock_irqrestore(&zone->lru_lock, flags);
1083 }
1084
1085 static void mem_cgroup_lru_add_after_commit(struct page *page)
1086 {
1087         unsigned long flags;
1088         struct zone *zone = page_zone(page);
1089         struct page_cgroup *pc = lookup_page_cgroup(page);
1090
1091         /* taking care of that the page is added to LRU while we commit it */
1092         if (likely(!PageLRU(page)))
1093                 return;
1094         spin_lock_irqsave(&zone->lru_lock, flags);
1095         /* link when the page is linked to LRU but page_cgroup isn't */
1096         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1097                 mem_cgroup_add_lru_list(page, page_lru(page));
1098         spin_unlock_irqrestore(&zone->lru_lock, flags);
1099 }
1100
1101
1102 void mem_cgroup_move_lists(struct page *page,
1103                            enum lru_list from, enum lru_list to)
1104 {
1105         if (mem_cgroup_disabled())
1106                 return;
1107         mem_cgroup_del_lru_list(page, from);
1108         mem_cgroup_add_lru_list(page, to);
1109 }
1110
1111 /*
1112  * Checks whether given mem is same or in the root_mem's
1113  * hierarchy subtree
1114  */
1115 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_mem,
1116                 struct mem_cgroup *mem)
1117 {
1118         if (root_mem != mem) {
1119                 return (root_mem->use_hierarchy &&
1120                         css_is_ancestor(&mem->css, &root_mem->css));
1121         }
1122
1123         return true;
1124 }
1125
1126 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1127 {
1128         int ret;
1129         struct mem_cgroup *curr = NULL;
1130         struct task_struct *p;
1131
1132         p = find_lock_task_mm(task);
1133         if (!p)
1134                 return 0;
1135         curr = try_get_mem_cgroup_from_mm(p->mm);
1136         task_unlock(p);
1137         if (!curr)
1138                 return 0;
1139         /*
1140          * We should check use_hierarchy of "mem" not "curr". Because checking
1141          * use_hierarchy of "curr" here make this function true if hierarchy is
1142          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1143          * hierarchy(even if use_hierarchy is disabled in "mem").
1144          */
1145         ret = mem_cgroup_same_or_subtree(mem, curr);
1146         css_put(&curr->css);
1147         return ret;
1148 }
1149
1150 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1151 {
1152         unsigned long active;
1153         unsigned long inactive;
1154         unsigned long gb;
1155         unsigned long inactive_ratio;
1156
1157         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1158         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1159
1160         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1161         if (gb)
1162                 inactive_ratio = int_sqrt(10 * gb);
1163         else
1164                 inactive_ratio = 1;
1165
1166         if (present_pages) {
1167                 present_pages[0] = inactive;
1168                 present_pages[1] = active;
1169         }
1170
1171         return inactive_ratio;
1172 }
1173
1174 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1175 {
1176         unsigned long active;
1177         unsigned long inactive;
1178         unsigned long present_pages[2];
1179         unsigned long inactive_ratio;
1180
1181         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1182
1183         inactive = present_pages[0];
1184         active = present_pages[1];
1185
1186         if (inactive * inactive_ratio < active)
1187                 return 1;
1188
1189         return 0;
1190 }
1191
1192 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1193 {
1194         unsigned long active;
1195         unsigned long inactive;
1196
1197         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1198         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1199
1200         return (active > inactive);
1201 }
1202
1203 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1204                                                       struct zone *zone)
1205 {
1206         int nid = zone_to_nid(zone);
1207         int zid = zone_idx(zone);
1208         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1209
1210         return &mz->reclaim_stat;
1211 }
1212
1213 struct zone_reclaim_stat *
1214 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1215 {
1216         struct page_cgroup *pc;
1217         struct mem_cgroup_per_zone *mz;
1218
1219         if (mem_cgroup_disabled())
1220                 return NULL;
1221
1222         pc = lookup_page_cgroup(page);
1223         if (!PageCgroupUsed(pc))
1224                 return NULL;
1225         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1226         smp_rmb();
1227         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1228         return &mz->reclaim_stat;
1229 }
1230
1231 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1232                                         struct list_head *dst,
1233                                         unsigned long *scanned, int order,
1234                                         int mode, struct zone *z,
1235                                         struct mem_cgroup *mem_cont,
1236                                         int active, int file)
1237 {
1238         unsigned long nr_taken = 0;
1239         struct page *page;
1240         unsigned long scan;
1241         LIST_HEAD(pc_list);
1242         struct list_head *src;
1243         struct page_cgroup *pc, *tmp;
1244         int nid = zone_to_nid(z);
1245         int zid = zone_idx(z);
1246         struct mem_cgroup_per_zone *mz;
1247         int lru = LRU_FILE * file + active;
1248         int ret;
1249
1250         BUG_ON(!mem_cont);
1251         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1252         src = &mz->lists[lru];
1253
1254         scan = 0;
1255         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1256                 if (scan >= nr_to_scan)
1257                         break;
1258
1259                 if (unlikely(!PageCgroupUsed(pc)))
1260                         continue;
1261
1262                 page = lookup_cgroup_page(pc);
1263
1264                 if (unlikely(!PageLRU(page)))
1265                         continue;
1266
1267                 scan++;
1268                 ret = __isolate_lru_page(page, mode, file);
1269                 switch (ret) {
1270                 case 0:
1271                         list_move(&page->lru, dst);
1272                         mem_cgroup_del_lru(page);
1273                         nr_taken += hpage_nr_pages(page);
1274                         break;
1275                 case -EBUSY:
1276                         /* we don't affect global LRU but rotate in our LRU */
1277                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1278                         break;
1279                 default:
1280                         break;
1281                 }
1282         }
1283
1284         *scanned = scan;
1285
1286         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1287                                       0, 0, 0, mode);
1288
1289         return nr_taken;
1290 }
1291
1292 #define mem_cgroup_from_res_counter(counter, member)    \
1293         container_of(counter, struct mem_cgroup, member)
1294
1295 /**
1296  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1297  * @mem: the memory cgroup
1298  *
1299  * Returns the maximum amount of memory @mem can be charged with, in
1300  * pages.
1301  */
1302 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1303 {
1304         unsigned long long margin;
1305
1306         margin = res_counter_margin(&mem->res);
1307         if (do_swap_account)
1308                 margin = min(margin, res_counter_margin(&mem->memsw));
1309         return margin >> PAGE_SHIFT;
1310 }
1311
1312 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1313 {
1314         struct cgroup *cgrp = memcg->css.cgroup;
1315
1316         /* root ? */
1317         if (cgrp->parent == NULL)
1318                 return vm_swappiness;
1319
1320         return memcg->swappiness;
1321 }
1322
1323 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1324 {
1325         int cpu;
1326
1327         get_online_cpus();
1328         spin_lock(&mem->pcp_counter_lock);
1329         for_each_online_cpu(cpu)
1330                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1331         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1332         spin_unlock(&mem->pcp_counter_lock);
1333         put_online_cpus();
1334
1335         synchronize_rcu();
1336 }
1337
1338 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1339 {
1340         int cpu;
1341
1342         if (!mem)
1343                 return;
1344         get_online_cpus();
1345         spin_lock(&mem->pcp_counter_lock);
1346         for_each_online_cpu(cpu)
1347                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1348         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1349         spin_unlock(&mem->pcp_counter_lock);
1350         put_online_cpus();
1351 }
1352 /*
1353  * 2 routines for checking "mem" is under move_account() or not.
1354  *
1355  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1356  *                        for avoiding race in accounting. If true,
1357  *                        pc->mem_cgroup may be overwritten.
1358  *
1359  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1360  *                        under hierarchy of moving cgroups. This is for
1361  *                        waiting at hith-memory prressure caused by "move".
1362  */
1363
1364 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1365 {
1366         VM_BUG_ON(!rcu_read_lock_held());
1367         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1368 }
1369
1370 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1371 {
1372         struct mem_cgroup *from;
1373         struct mem_cgroup *to;
1374         bool ret = false;
1375         /*
1376          * Unlike task_move routines, we access mc.to, mc.from not under
1377          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1378          */
1379         spin_lock(&mc.lock);
1380         from = mc.from;
1381         to = mc.to;
1382         if (!from)
1383                 goto unlock;
1384
1385         ret = mem_cgroup_same_or_subtree(mem, from)
1386                 || mem_cgroup_same_or_subtree(mem, to);
1387 unlock:
1388         spin_unlock(&mc.lock);
1389         return ret;
1390 }
1391
1392 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1393 {
1394         if (mc.moving_task && current != mc.moving_task) {
1395                 if (mem_cgroup_under_move(mem)) {
1396                         DEFINE_WAIT(wait);
1397                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1398                         /* moving charge context might have finished. */
1399                         if (mc.moving_task)
1400                                 schedule();
1401                         finish_wait(&mc.waitq, &wait);
1402                         return true;
1403                 }
1404         }
1405         return false;
1406 }
1407
1408 /**
1409  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1410  * @memcg: The memory cgroup that went over limit
1411  * @p: Task that is going to be killed
1412  *
1413  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1414  * enabled
1415  */
1416 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1417 {
1418         struct cgroup *task_cgrp;
1419         struct cgroup *mem_cgrp;
1420         /*
1421          * Need a buffer in BSS, can't rely on allocations. The code relies
1422          * on the assumption that OOM is serialized for memory controller.
1423          * If this assumption is broken, revisit this code.
1424          */
1425         static char memcg_name[PATH_MAX];
1426         int ret;
1427
1428         if (!memcg || !p)
1429                 return;
1430
1431
1432         rcu_read_lock();
1433
1434         mem_cgrp = memcg->css.cgroup;
1435         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1436
1437         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1438         if (ret < 0) {
1439                 /*
1440                  * Unfortunately, we are unable to convert to a useful name
1441                  * But we'll still print out the usage information
1442                  */
1443                 rcu_read_unlock();
1444                 goto done;
1445         }
1446         rcu_read_unlock();
1447
1448         printk(KERN_INFO "Task in %s killed", memcg_name);
1449
1450         rcu_read_lock();
1451         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1452         if (ret < 0) {
1453                 rcu_read_unlock();
1454                 goto done;
1455         }
1456         rcu_read_unlock();
1457
1458         /*
1459          * Continues from above, so we don't need an KERN_ level
1460          */
1461         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1462 done:
1463
1464         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1465                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1466                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1467                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1468         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1469                 "failcnt %llu\n",
1470                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1471                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1472                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1473 }
1474
1475 /*
1476  * This function returns the number of memcg under hierarchy tree. Returns
1477  * 1(self count) if no children.
1478  */
1479 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1480 {
1481         int num = 0;
1482         struct mem_cgroup *iter;
1483
1484         for_each_mem_cgroup_tree(iter, mem)
1485                 num++;
1486         return num;
1487 }
1488
1489 /*
1490  * Return the memory (and swap, if configured) limit for a memcg.
1491  */
1492 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1493 {
1494         u64 limit;
1495         u64 memsw;
1496
1497         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1498         limit += total_swap_pages << PAGE_SHIFT;
1499
1500         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1501         /*
1502          * If memsw is finite and limits the amount of swap space available
1503          * to this memcg, return that limit.
1504          */
1505         return min(limit, memsw);
1506 }
1507
1508 /*
1509  * Visit the first child (need not be the first child as per the ordering
1510  * of the cgroup list, since we track last_scanned_child) of @mem and use
1511  * that to reclaim free pages from.
1512  */
1513 static struct mem_cgroup *
1514 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1515 {
1516         struct mem_cgroup *ret = NULL;
1517         struct cgroup_subsys_state *css;
1518         int nextid, found;
1519
1520         if (!root_mem->use_hierarchy) {
1521                 css_get(&root_mem->css);
1522                 ret = root_mem;
1523         }
1524
1525         while (!ret) {
1526                 rcu_read_lock();
1527                 nextid = root_mem->last_scanned_child + 1;
1528                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1529                                    &found);
1530                 if (css && css_tryget(css))
1531                         ret = container_of(css, struct mem_cgroup, css);
1532
1533                 rcu_read_unlock();
1534                 /* Updates scanning parameter */
1535                 if (!css) {
1536                         /* this means start scan from ID:1 */
1537                         root_mem->last_scanned_child = 0;
1538                 } else
1539                         root_mem->last_scanned_child = found;
1540         }
1541
1542         return ret;
1543 }
1544
1545 /**
1546  * test_mem_cgroup_node_reclaimable
1547  * @mem: the target memcg
1548  * @nid: the node ID to be checked.
1549  * @noswap : specify true here if the user wants flle only information.
1550  *
1551  * This function returns whether the specified memcg contains any
1552  * reclaimable pages on a node. Returns true if there are any reclaimable
1553  * pages in the node.
1554  */
1555 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1556                 int nid, bool noswap)
1557 {
1558         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1559                 return true;
1560         if (noswap || !total_swap_pages)
1561                 return false;
1562         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1563                 return true;
1564         return false;
1565
1566 }
1567 #if MAX_NUMNODES > 1
1568
1569 /*
1570  * Always updating the nodemask is not very good - even if we have an empty
1571  * list or the wrong list here, we can start from some node and traverse all
1572  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1573  *
1574  */
1575 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1576 {
1577         int nid;
1578         /*
1579          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1580          * pagein/pageout changes since the last update.
1581          */
1582         if (!atomic_read(&mem->numainfo_events))
1583                 return;
1584         if (atomic_inc_return(&mem->numainfo_updating) > 1)
1585                 return;
1586
1587         /* make a nodemask where this memcg uses memory from */
1588         mem->scan_nodes = node_states[N_HIGH_MEMORY];
1589
1590         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1591
1592                 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1593                         node_clear(nid, mem->scan_nodes);
1594         }
1595
1596         atomic_set(&mem->numainfo_events, 0);
1597         atomic_set(&mem->numainfo_updating, 0);
1598 }
1599
1600 /*
1601  * Selecting a node where we start reclaim from. Because what we need is just
1602  * reducing usage counter, start from anywhere is O,K. Considering
1603  * memory reclaim from current node, there are pros. and cons.
1604  *
1605  * Freeing memory from current node means freeing memory from a node which
1606  * we'll use or we've used. So, it may make LRU bad. And if several threads
1607  * hit limits, it will see a contention on a node. But freeing from remote
1608  * node means more costs for memory reclaim because of memory latency.
1609  *
1610  * Now, we use round-robin. Better algorithm is welcomed.
1611  */
1612 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1613 {
1614         int node;
1615
1616         mem_cgroup_may_update_nodemask(mem);
1617         node = mem->last_scanned_node;
1618
1619         node = next_node(node, mem->scan_nodes);
1620         if (node == MAX_NUMNODES)
1621                 node = first_node(mem->scan_nodes);
1622         /*
1623          * We call this when we hit limit, not when pages are added to LRU.
1624          * No LRU may hold pages because all pages are UNEVICTABLE or
1625          * memcg is too small and all pages are not on LRU. In that case,
1626          * we use curret node.
1627          */
1628         if (unlikely(node == MAX_NUMNODES))
1629                 node = numa_node_id();
1630
1631         mem->last_scanned_node = node;
1632         return node;
1633 }
1634
1635 /*
1636  * Check all nodes whether it contains reclaimable pages or not.
1637  * For quick scan, we make use of scan_nodes. This will allow us to skip
1638  * unused nodes. But scan_nodes is lazily updated and may not cotain
1639  * enough new information. We need to do double check.
1640  */
1641 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1642 {
1643         int nid;
1644
1645         /*
1646          * quick check...making use of scan_node.
1647          * We can skip unused nodes.
1648          */
1649         if (!nodes_empty(mem->scan_nodes)) {
1650                 for (nid = first_node(mem->scan_nodes);
1651                      nid < MAX_NUMNODES;
1652                      nid = next_node(nid, mem->scan_nodes)) {
1653
1654                         if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1655                                 return true;
1656                 }
1657         }
1658         /*
1659          * Check rest of nodes.
1660          */
1661         for_each_node_state(nid, N_HIGH_MEMORY) {
1662                 if (node_isset(nid, mem->scan_nodes))
1663                         continue;
1664                 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1665                         return true;
1666         }
1667         return false;
1668 }
1669
1670 #else
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1672 {
1673         return 0;
1674 }
1675
1676 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1677 {
1678         return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1679 }
1680 #endif
1681
1682 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1683                            struct memcg_scanrecord *rec)
1684 {
1685
1686         stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1687         stats[SCAN_ANON] += rec->nr_scanned[0];
1688         stats[SCAN_FILE] += rec->nr_scanned[1];
1689
1690         stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1691         stats[ROTATE_ANON] += rec->nr_rotated[0];
1692         stats[ROTATE_FILE] += rec->nr_rotated[1];
1693
1694         stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1695         stats[FREED_ANON] += rec->nr_freed[0];
1696         stats[FREED_FILE] += rec->nr_freed[1];
1697
1698         stats[ELAPSED] += rec->elapsed;
1699 }
1700
1701 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1702 {
1703         struct mem_cgroup *mem;
1704         int context = rec->context;
1705
1706         if (context >= NR_SCAN_CONTEXT)
1707                 return;
1708
1709         mem = rec->mem;
1710         spin_lock(&mem->scanstat.lock);
1711         __mem_cgroup_record_scanstat(mem->scanstat.stats[context], rec);
1712         spin_unlock(&mem->scanstat.lock);
1713
1714         mem = rec->root;
1715         spin_lock(&mem->scanstat.lock);
1716         __mem_cgroup_record_scanstat(mem->scanstat.rootstats[context], rec);
1717         spin_unlock(&mem->scanstat.lock);
1718 }
1719
1720 /*
1721  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1722  * we reclaimed from, so that we don't end up penalizing one child extensively
1723  * based on its position in the children list.
1724  *
1725  * root_mem is the original ancestor that we've been reclaim from.
1726  *
1727  * We give up and return to the caller when we visit root_mem twice.
1728  * (other groups can be removed while we're walking....)
1729  *
1730  * If shrink==true, for avoiding to free too much, this returns immedieately.
1731  */
1732 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1733                                                 struct zone *zone,
1734                                                 gfp_t gfp_mask,
1735                                                 unsigned long reclaim_options,
1736                                                 unsigned long *total_scanned)
1737 {
1738         struct mem_cgroup *victim;
1739         int ret, total = 0;
1740         int loop = 0;
1741         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1742         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1743         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1744         struct memcg_scanrecord rec;
1745         unsigned long excess;
1746         unsigned long scanned;
1747
1748         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1749
1750         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1751         if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1752                 noswap = true;
1753
1754         if (shrink)
1755                 rec.context = SCAN_BY_SHRINK;
1756         else if (check_soft)
1757                 rec.context = SCAN_BY_SYSTEM;
1758         else
1759                 rec.context = SCAN_BY_LIMIT;
1760
1761         rec.root = root_mem;
1762
1763         while (1) {
1764                 victim = mem_cgroup_select_victim(root_mem);
1765                 if (victim == root_mem) {
1766                         loop++;
1767                         /*
1768                          * We are not draining per cpu cached charges during
1769                          * soft limit reclaim  because global reclaim doesn't
1770                          * care about charges. It tries to free some memory and
1771                          * charges will not give any.
1772                          */
1773                         if (!check_soft && loop >= 1)
1774                                 drain_all_stock_async(root_mem);
1775                         if (loop >= 2) {
1776                                 /*
1777                                  * If we have not been able to reclaim
1778                                  * anything, it might because there are
1779                                  * no reclaimable pages under this hierarchy
1780                                  */
1781                                 if (!check_soft || !total) {
1782                                         css_put(&victim->css);
1783                                         break;
1784                                 }
1785                                 /*
1786                                  * We want to do more targeted reclaim.
1787                                  * excess >> 2 is not to excessive so as to
1788                                  * reclaim too much, nor too less that we keep
1789                                  * coming back to reclaim from this cgroup
1790                                  */
1791                                 if (total >= (excess >> 2) ||
1792                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1793                                         css_put(&victim->css);
1794                                         break;
1795                                 }
1796                         }
1797                 }
1798                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1799                         /* this cgroup's local usage == 0 */
1800                         css_put(&victim->css);
1801                         continue;
1802                 }
1803                 rec.mem = victim;
1804                 rec.nr_scanned[0] = 0;
1805                 rec.nr_scanned[1] = 0;
1806                 rec.nr_rotated[0] = 0;
1807                 rec.nr_rotated[1] = 0;
1808                 rec.nr_freed[0] = 0;
1809                 rec.nr_freed[1] = 0;
1810                 rec.elapsed = 0;
1811                 /* we use swappiness of local cgroup */
1812                 if (check_soft) {
1813                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1814                                 noswap, zone, &rec, &scanned);
1815                         *total_scanned += scanned;
1816                 } else
1817                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1818                                                 noswap, &rec);
1819                 mem_cgroup_record_scanstat(&rec);
1820                 css_put(&victim->css);
1821                 /*
1822                  * At shrinking usage, we can't check we should stop here or
1823                  * reclaim more. It's depends on callers. last_scanned_child
1824                  * will work enough for keeping fairness under tree.
1825                  */
1826                 if (shrink)
1827                         return ret;
1828                 total += ret;
1829                 if (check_soft) {
1830                         if (!res_counter_soft_limit_excess(&root_mem->res))
1831                                 return total;
1832                 } else if (mem_cgroup_margin(root_mem))
1833                         return total;
1834         }
1835         return total;
1836 }
1837
1838 /*
1839  * Check OOM-Killer is already running under our hierarchy.
1840  * If someone is running, return false.
1841  * Has to be called with memcg_oom_lock
1842  */
1843 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1844 {
1845         int lock_count = -1;
1846         struct mem_cgroup *iter, *failed = NULL;
1847         bool cond = true;
1848
1849         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1850                 bool locked = iter->oom_lock;
1851
1852                 iter->oom_lock = true;
1853                 if (lock_count == -1)
1854                         lock_count = iter->oom_lock;
1855                 else if (lock_count != locked) {
1856                         /*
1857                          * this subtree of our hierarchy is already locked
1858                          * so we cannot give a lock.
1859                          */
1860                         lock_count = 0;
1861                         failed = iter;
1862                         cond = false;
1863                 }
1864         }
1865
1866         if (!failed)
1867                 goto done;
1868
1869         /*
1870          * OK, we failed to lock the whole subtree so we have to clean up
1871          * what we set up to the failing subtree
1872          */
1873         cond = true;
1874         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1875                 if (iter == failed) {
1876                         cond = false;
1877                         continue;
1878                 }
1879                 iter->oom_lock = false;
1880         }
1881 done:
1882         return lock_count;
1883 }
1884
1885 /*
1886  * Has to be called with memcg_oom_lock
1887  */
1888 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1889 {
1890         struct mem_cgroup *iter;
1891
1892         for_each_mem_cgroup_tree(iter, mem)
1893                 iter->oom_lock = false;
1894         return 0;
1895 }
1896
1897 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1898 {
1899         struct mem_cgroup *iter;
1900
1901         for_each_mem_cgroup_tree(iter, mem)
1902                 atomic_inc(&iter->under_oom);
1903 }
1904
1905 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1906 {
1907         struct mem_cgroup *iter;
1908
1909         /*
1910          * When a new child is created while the hierarchy is under oom,
1911          * mem_cgroup_oom_lock() may not be called. We have to use
1912          * atomic_add_unless() here.
1913          */
1914         for_each_mem_cgroup_tree(iter, mem)
1915                 atomic_add_unless(&iter->under_oom, -1, 0);
1916 }
1917
1918 static DEFINE_SPINLOCK(memcg_oom_lock);
1919 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1920
1921 struct oom_wait_info {
1922         struct mem_cgroup *mem;
1923         wait_queue_t    wait;
1924 };
1925
1926 static int memcg_oom_wake_function(wait_queue_t *wait,
1927         unsigned mode, int sync, void *arg)
1928 {
1929         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg,
1930                           *oom_wait_mem;
1931         struct oom_wait_info *oom_wait_info;
1932
1933         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1934         oom_wait_mem = oom_wait_info->mem;
1935
1936         /*
1937          * Both of oom_wait_info->mem and wake_mem are stable under us.
1938          * Then we can use css_is_ancestor without taking care of RCU.
1939          */
1940         if (!mem_cgroup_same_or_subtree(oom_wait_mem, wake_mem)
1941                         && !mem_cgroup_same_or_subtree(wake_mem, oom_wait_mem))
1942                 return 0;
1943         return autoremove_wake_function(wait, mode, sync, arg);
1944 }
1945
1946 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1947 {
1948         /* for filtering, pass "mem" as argument. */
1949         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1950 }
1951
1952 static void memcg_oom_recover(struct mem_cgroup *mem)
1953 {
1954         if (mem && atomic_read(&mem->under_oom))
1955                 memcg_wakeup_oom(mem);
1956 }
1957
1958 /*
1959  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1960  */
1961 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1962 {
1963         struct oom_wait_info owait;
1964         bool locked, need_to_kill;
1965
1966         owait.mem = mem;
1967         owait.wait.flags = 0;
1968         owait.wait.func = memcg_oom_wake_function;
1969         owait.wait.private = current;
1970         INIT_LIST_HEAD(&owait.wait.task_list);
1971         need_to_kill = true;
1972         mem_cgroup_mark_under_oom(mem);
1973
1974         /* At first, try to OOM lock hierarchy under mem.*/
1975         spin_lock(&memcg_oom_lock);
1976         locked = mem_cgroup_oom_lock(mem);
1977         /*
1978          * Even if signal_pending(), we can't quit charge() loop without
1979          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1980          * under OOM is always welcomed, use TASK_KILLABLE here.
1981          */
1982         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1983         if (!locked || mem->oom_kill_disable)
1984                 need_to_kill = false;
1985         if (locked)
1986                 mem_cgroup_oom_notify(mem);
1987         spin_unlock(&memcg_oom_lock);
1988
1989         if (need_to_kill) {
1990                 finish_wait(&memcg_oom_waitq, &owait.wait);
1991                 mem_cgroup_out_of_memory(mem, mask);
1992         } else {
1993                 schedule();
1994                 finish_wait(&memcg_oom_waitq, &owait.wait);
1995         }
1996         spin_lock(&memcg_oom_lock);
1997         if (locked)
1998                 mem_cgroup_oom_unlock(mem);
1999         memcg_wakeup_oom(mem);
2000         spin_unlock(&memcg_oom_lock);
2001
2002         mem_cgroup_unmark_under_oom(mem);
2003
2004         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2005                 return false;
2006         /* Give chance to dying process */
2007         schedule_timeout(1);
2008         return true;
2009 }
2010
2011 /*
2012  * Currently used to update mapped file statistics, but the routine can be
2013  * generalized to update other statistics as well.
2014  *
2015  * Notes: Race condition
2016  *
2017  * We usually use page_cgroup_lock() for accessing page_cgroup member but
2018  * it tends to be costly. But considering some conditions, we doesn't need
2019  * to do so _always_.
2020  *
2021  * Considering "charge", lock_page_cgroup() is not required because all
2022  * file-stat operations happen after a page is attached to radix-tree. There
2023  * are no race with "charge".
2024  *
2025  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2026  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2027  * if there are race with "uncharge". Statistics itself is properly handled
2028  * by flags.
2029  *
2030  * Considering "move", this is an only case we see a race. To make the race
2031  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2032  * possibility of race condition. If there is, we take a lock.
2033  */
2034
2035 void mem_cgroup_update_page_stat(struct page *page,
2036                                  enum mem_cgroup_page_stat_item idx, int val)
2037 {
2038         struct mem_cgroup *mem;
2039         struct page_cgroup *pc = lookup_page_cgroup(page);
2040         bool need_unlock = false;
2041         unsigned long uninitialized_var(flags);
2042
2043         if (unlikely(!pc))
2044                 return;
2045
2046         rcu_read_lock();
2047         mem = pc->mem_cgroup;
2048         if (unlikely(!mem || !PageCgroupUsed(pc)))
2049                 goto out;
2050         /* pc->mem_cgroup is unstable ? */
2051         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
2052                 /* take a lock against to access pc->mem_cgroup */
2053                 move_lock_page_cgroup(pc, &flags);
2054                 need_unlock = true;
2055                 mem = pc->mem_cgroup;
2056                 if (!mem || !PageCgroupUsed(pc))
2057                         goto out;
2058         }
2059
2060         switch (idx) {
2061         case MEMCG_NR_FILE_MAPPED:
2062                 if (val > 0)
2063                         SetPageCgroupFileMapped(pc);
2064                 else if (!page_mapped(page))
2065                         ClearPageCgroupFileMapped(pc);
2066                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2067                 break;
2068         default:
2069                 BUG();
2070         }
2071
2072         this_cpu_add(mem->stat->count[idx], val);
2073
2074 out:
2075         if (unlikely(need_unlock))
2076                 move_unlock_page_cgroup(pc, &flags);
2077         rcu_read_unlock();
2078         return;
2079 }
2080 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2081
2082 /*
2083  * size of first charge trial. "32" comes from vmscan.c's magic value.
2084  * TODO: maybe necessary to use big numbers in big irons.
2085  */
2086 #define CHARGE_BATCH    32U
2087 struct memcg_stock_pcp {
2088         struct mem_cgroup *cached; /* this never be root cgroup */
2089         unsigned int nr_pages;
2090         struct work_struct work;
2091         unsigned long flags;
2092 #define FLUSHING_CACHED_CHARGE  (0)
2093 };
2094 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2095 static DEFINE_MUTEX(percpu_charge_mutex);
2096
2097 /*
2098  * Try to consume stocked charge on this cpu. If success, one page is consumed
2099  * from local stock and true is returned. If the stock is 0 or charges from a
2100  * cgroup which is not current target, returns false. This stock will be
2101  * refilled.
2102  */
2103 static bool consume_stock(struct mem_cgroup *mem)
2104 {
2105         struct memcg_stock_pcp *stock;
2106         bool ret = true;
2107
2108         stock = &get_cpu_var(memcg_stock);
2109         if (mem == stock->cached && stock->nr_pages)
2110                 stock->nr_pages--;
2111         else /* need to call res_counter_charge */
2112                 ret = false;
2113         put_cpu_var(memcg_stock);
2114         return ret;
2115 }
2116
2117 /*
2118  * Returns stocks cached in percpu to res_counter and reset cached information.
2119  */
2120 static void drain_stock(struct memcg_stock_pcp *stock)
2121 {
2122         struct mem_cgroup *old = stock->cached;
2123
2124         if (stock->nr_pages) {
2125                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2126
2127                 res_counter_uncharge(&old->res, bytes);
2128                 if (do_swap_account)
2129                         res_counter_uncharge(&old->memsw, bytes);
2130                 stock->nr_pages = 0;
2131         }
2132         stock->cached = NULL;
2133 }
2134
2135 /*
2136  * This must be called under preempt disabled or must be called by
2137  * a thread which is pinned to local cpu.
2138  */
2139 static void drain_local_stock(struct work_struct *dummy)
2140 {
2141         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2142         drain_stock(stock);
2143         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2144 }
2145
2146 /*
2147  * Cache charges(val) which is from res_counter, to local per_cpu area.
2148  * This will be consumed by consume_stock() function, later.
2149  */
2150 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2151 {
2152         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2153
2154         if (stock->cached != mem) { /* reset if necessary */
2155                 drain_stock(stock);
2156                 stock->cached = mem;
2157         }
2158         stock->nr_pages += nr_pages;
2159         put_cpu_var(memcg_stock);
2160 }
2161
2162 /*
2163  * Drains all per-CPU charge caches for given root_mem resp. subtree
2164  * of the hierarchy under it. sync flag says whether we should block
2165  * until the work is done.
2166  */
2167 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2168 {
2169         int cpu, curcpu;
2170
2171         /* Notify other cpus that system-wide "drain" is running */
2172         get_online_cpus();
2173         /*
2174          * Get a hint for avoiding draining charges on the current cpu,
2175          * which must be exhausted by our charging.  It is not required that
2176          * this be a precise check, so we use raw_smp_processor_id() instead of
2177          * getcpu()/putcpu().
2178          */
2179         curcpu = raw_smp_processor_id();
2180         for_each_online_cpu(cpu) {
2181                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2182                 struct mem_cgroup *mem;
2183
2184                 mem = stock->cached;
2185                 if (!mem || !stock->nr_pages)
2186                         continue;
2187                 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2188                         continue;
2189                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2190                         if (cpu == curcpu)
2191                                 drain_local_stock(&stock->work);
2192                         else
2193                                 schedule_work_on(cpu, &stock->work);
2194                 }
2195         }
2196
2197         if (!sync)
2198                 goto out;
2199
2200         for_each_online_cpu(cpu) {
2201                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2202                 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2203                         flush_work(&stock->work);
2204         }
2205 out:
2206         put_online_cpus();
2207 }
2208
2209 /*
2210  * Tries to drain stocked charges in other cpus. This function is asynchronous
2211  * and just put a work per cpu for draining localy on each cpu. Caller can
2212  * expects some charges will be back to res_counter later but cannot wait for
2213  * it.
2214  */
2215 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2216 {
2217         /*
2218          * If someone calls draining, avoid adding more kworker runs.
2219          */
2220         if (!mutex_trylock(&percpu_charge_mutex))
2221                 return;
2222         drain_all_stock(root_mem, false);
2223         mutex_unlock(&percpu_charge_mutex);
2224 }
2225
2226 /* This is a synchronous drain interface. */
2227 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2228 {
2229         /* called when force_empty is called */
2230         mutex_lock(&percpu_charge_mutex);
2231         drain_all_stock(root_mem, true);
2232         mutex_unlock(&percpu_charge_mutex);
2233 }
2234
2235 /*
2236  * This function drains percpu counter value from DEAD cpu and
2237  * move it to local cpu. Note that this function can be preempted.
2238  */
2239 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2240 {
2241         int i;
2242
2243         spin_lock(&mem->pcp_counter_lock);
2244         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2245                 long x = per_cpu(mem->stat->count[i], cpu);
2246
2247                 per_cpu(mem->stat->count[i], cpu) = 0;
2248                 mem->nocpu_base.count[i] += x;
2249         }
2250         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2251                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2252
2253                 per_cpu(mem->stat->events[i], cpu) = 0;
2254                 mem->nocpu_base.events[i] += x;
2255         }
2256         /* need to clear ON_MOVE value, works as a kind of lock. */
2257         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2258         spin_unlock(&mem->pcp_counter_lock);
2259 }
2260
2261 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2262 {
2263         int idx = MEM_CGROUP_ON_MOVE;
2264
2265         spin_lock(&mem->pcp_counter_lock);
2266         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2267         spin_unlock(&mem->pcp_counter_lock);
2268 }
2269
2270 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2271                                         unsigned long action,
2272                                         void *hcpu)
2273 {
2274         int cpu = (unsigned long)hcpu;
2275         struct memcg_stock_pcp *stock;
2276         struct mem_cgroup *iter;
2277
2278         if ((action == CPU_ONLINE)) {
2279                 for_each_mem_cgroup_all(iter)
2280                         synchronize_mem_cgroup_on_move(iter, cpu);
2281                 return NOTIFY_OK;
2282         }
2283
2284         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2285                 return NOTIFY_OK;
2286
2287         for_each_mem_cgroup_all(iter)
2288                 mem_cgroup_drain_pcp_counter(iter, cpu);
2289
2290         stock = &per_cpu(memcg_stock, cpu);
2291         drain_stock(stock);
2292         return NOTIFY_OK;
2293 }
2294
2295
2296 /* See __mem_cgroup_try_charge() for details */
2297 enum {
2298         CHARGE_OK,              /* success */
2299         CHARGE_RETRY,           /* need to retry but retry is not bad */
2300         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2301         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2302         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2303 };
2304
2305 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2306                                 unsigned int nr_pages, bool oom_check)
2307 {
2308         unsigned long csize = nr_pages * PAGE_SIZE;
2309         struct mem_cgroup *mem_over_limit;
2310         struct res_counter *fail_res;
2311         unsigned long flags = 0;
2312         int ret;
2313
2314         ret = res_counter_charge(&mem->res, csize, &fail_res);
2315
2316         if (likely(!ret)) {
2317                 if (!do_swap_account)
2318                         return CHARGE_OK;
2319                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2320                 if (likely(!ret))
2321                         return CHARGE_OK;
2322
2323                 res_counter_uncharge(&mem->res, csize);
2324                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2325                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2326         } else
2327                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2328         /*
2329          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2330          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2331          *
2332          * Never reclaim on behalf of optional batching, retry with a
2333          * single page instead.
2334          */
2335         if (nr_pages == CHARGE_BATCH)
2336                 return CHARGE_RETRY;
2337
2338         if (!(gfp_mask & __GFP_WAIT))
2339                 return CHARGE_WOULDBLOCK;
2340
2341         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2342                                               gfp_mask, flags, NULL);
2343         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2344                 return CHARGE_RETRY;
2345         /*
2346          * Even though the limit is exceeded at this point, reclaim
2347          * may have been able to free some pages.  Retry the charge
2348          * before killing the task.
2349          *
2350          * Only for regular pages, though: huge pages are rather
2351          * unlikely to succeed so close to the limit, and we fall back
2352          * to regular pages anyway in case of failure.
2353          */
2354         if (nr_pages == 1 && ret)
2355                 return CHARGE_RETRY;
2356
2357         /*
2358          * At task move, charge accounts can be doubly counted. So, it's
2359          * better to wait until the end of task_move if something is going on.
2360          */
2361         if (mem_cgroup_wait_acct_move(mem_over_limit))
2362                 return CHARGE_RETRY;
2363
2364         /* If we don't need to call oom-killer at el, return immediately */
2365         if (!oom_check)
2366                 return CHARGE_NOMEM;
2367         /* check OOM */
2368         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2369                 return CHARGE_OOM_DIE;
2370
2371         return CHARGE_RETRY;
2372 }
2373
2374 /*
2375  * Unlike exported interface, "oom" parameter is added. if oom==true,
2376  * oom-killer can be invoked.
2377  */
2378 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2379                                    gfp_t gfp_mask,
2380                                    unsigned int nr_pages,
2381                                    struct mem_cgroup **memcg,
2382                                    bool oom)
2383 {
2384         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2385         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2386         struct mem_cgroup *mem = NULL;
2387         int ret;
2388
2389         /*
2390          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2391          * in system level. So, allow to go ahead dying process in addition to
2392          * MEMDIE process.
2393          */
2394         if (unlikely(test_thread_flag(TIF_MEMDIE)
2395                      || fatal_signal_pending(current)))
2396                 goto bypass;
2397
2398         /*
2399          * We always charge the cgroup the mm_struct belongs to.
2400          * The mm_struct's mem_cgroup changes on task migration if the
2401          * thread group leader migrates. It's possible that mm is not
2402          * set, if so charge the init_mm (happens for pagecache usage).
2403          */
2404         if (!*memcg && !mm)
2405                 goto bypass;
2406 again:
2407         if (*memcg) { /* css should be a valid one */
2408                 mem = *memcg;
2409                 VM_BUG_ON(css_is_removed(&mem->css));
2410                 if (mem_cgroup_is_root(mem))
2411                         goto done;
2412                 if (nr_pages == 1 && consume_stock(mem))
2413                         goto done;
2414                 css_get(&mem->css);
2415         } else {
2416                 struct task_struct *p;
2417
2418                 rcu_read_lock();
2419                 p = rcu_dereference(mm->owner);
2420                 /*
2421                  * Because we don't have task_lock(), "p" can exit.
2422                  * In that case, "mem" can point to root or p can be NULL with
2423                  * race with swapoff. Then, we have small risk of mis-accouning.
2424                  * But such kind of mis-account by race always happens because
2425                  * we don't have cgroup_mutex(). It's overkill and we allo that
2426                  * small race, here.
2427                  * (*) swapoff at el will charge against mm-struct not against
2428                  * task-struct. So, mm->owner can be NULL.
2429                  */
2430                 mem = mem_cgroup_from_task(p);
2431                 if (!mem || mem_cgroup_is_root(mem)) {
2432                         rcu_read_unlock();
2433                         goto done;
2434                 }
2435                 if (nr_pages == 1 && consume_stock(mem)) {
2436                         /*
2437                          * It seems dagerous to access memcg without css_get().
2438                          * But considering how consume_stok works, it's not
2439                          * necessary. If consume_stock success, some charges
2440                          * from this memcg are cached on this cpu. So, we
2441                          * don't need to call css_get()/css_tryget() before
2442                          * calling consume_stock().
2443                          */
2444                         rcu_read_unlock();
2445                         goto done;
2446                 }
2447                 /* after here, we may be blocked. we need to get refcnt */
2448                 if (!css_tryget(&mem->css)) {
2449                         rcu_read_unlock();
2450                         goto again;
2451                 }
2452                 rcu_read_unlock();
2453         }
2454
2455         do {
2456                 bool oom_check;
2457
2458                 /* If killed, bypass charge */
2459                 if (fatal_signal_pending(current)) {
2460                         css_put(&mem->css);
2461                         goto bypass;
2462                 }
2463
2464                 oom_check = false;
2465                 if (oom && !nr_oom_retries) {
2466                         oom_check = true;
2467                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2468                 }
2469
2470                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2471                 switch (ret) {
2472                 case CHARGE_OK:
2473                         break;
2474                 case CHARGE_RETRY: /* not in OOM situation but retry */
2475                         batch = nr_pages;
2476                         css_put(&mem->css);
2477                         mem = NULL;
2478                         goto again;
2479                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2480                         css_put(&mem->css);
2481                         goto nomem;
2482                 case CHARGE_NOMEM: /* OOM routine works */
2483                         if (!oom) {
2484                                 css_put(&mem->css);
2485                                 goto nomem;
2486                         }
2487                         /* If oom, we never return -ENOMEM */
2488                         nr_oom_retries--;
2489                         break;
2490                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2491                         css_put(&mem->css);
2492                         goto bypass;
2493                 }
2494         } while (ret != CHARGE_OK);
2495
2496         if (batch > nr_pages)
2497                 refill_stock(mem, batch - nr_pages);
2498         css_put(&mem->css);
2499 done:
2500         *memcg = mem;
2501         return 0;
2502 nomem:
2503         *memcg = NULL;
2504         return -ENOMEM;
2505 bypass:
2506         *memcg = NULL;
2507         return 0;
2508 }
2509
2510 /*
2511  * Somemtimes we have to undo a charge we got by try_charge().
2512  * This function is for that and do uncharge, put css's refcnt.
2513  * gotten by try_charge().
2514  */
2515 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2516                                        unsigned int nr_pages)
2517 {
2518         if (!mem_cgroup_is_root(mem)) {
2519                 unsigned long bytes = nr_pages * PAGE_SIZE;
2520
2521                 res_counter_uncharge(&mem->res, bytes);
2522                 if (do_swap_account)
2523                         res_counter_uncharge(&mem->memsw, bytes);
2524         }
2525 }
2526
2527 /*
2528  * A helper function to get mem_cgroup from ID. must be called under
2529  * rcu_read_lock(). The caller must check css_is_removed() or some if
2530  * it's concern. (dropping refcnt from swap can be called against removed
2531  * memcg.)
2532  */
2533 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2534 {
2535         struct cgroup_subsys_state *css;
2536
2537         /* ID 0 is unused ID */
2538         if (!id)
2539                 return NULL;
2540         css = css_lookup(&mem_cgroup_subsys, id);
2541         if (!css)
2542                 return NULL;
2543         return container_of(css, struct mem_cgroup, css);
2544 }
2545
2546 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2547 {
2548         struct mem_cgroup *mem = NULL;
2549         struct page_cgroup *pc;
2550         unsigned short id;
2551         swp_entry_t ent;
2552
2553         VM_BUG_ON(!PageLocked(page));
2554
2555         pc = lookup_page_cgroup(page);
2556         lock_page_cgroup(pc);
2557         if (PageCgroupUsed(pc)) {
2558                 mem = pc->mem_cgroup;
2559                 if (mem && !css_tryget(&mem->css))
2560                         mem = NULL;
2561         } else if (PageSwapCache(page)) {
2562                 ent.val = page_private(page);
2563                 id = lookup_swap_cgroup(ent);
2564                 rcu_read_lock();
2565                 mem = mem_cgroup_lookup(id);
2566                 if (mem && !css_tryget(&mem->css))
2567                         mem = NULL;
2568                 rcu_read_unlock();
2569         }
2570         unlock_page_cgroup(pc);
2571         return mem;
2572 }
2573
2574 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2575                                        struct page *page,
2576                                        unsigned int nr_pages,
2577                                        struct page_cgroup *pc,
2578                                        enum charge_type ctype)
2579 {
2580         lock_page_cgroup(pc);
2581         if (unlikely(PageCgroupUsed(pc))) {
2582                 unlock_page_cgroup(pc);
2583                 __mem_cgroup_cancel_charge(mem, nr_pages);
2584                 return;
2585         }
2586         /*
2587          * we don't need page_cgroup_lock about tail pages, becase they are not
2588          * accessed by any other context at this point.
2589          */
2590         pc->mem_cgroup = mem;
2591         /*
2592          * We access a page_cgroup asynchronously without lock_page_cgroup().
2593          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2594          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2595          * before USED bit, we need memory barrier here.
2596          * See mem_cgroup_add_lru_list(), etc.
2597          */
2598         smp_wmb();
2599         switch (ctype) {
2600         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2601         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2602                 SetPageCgroupCache(pc);
2603                 SetPageCgroupUsed(pc);
2604                 break;
2605         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2606                 ClearPageCgroupCache(pc);
2607                 SetPageCgroupUsed(pc);
2608                 break;
2609         default:
2610                 break;
2611         }
2612
2613         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2614         unlock_page_cgroup(pc);
2615         /*
2616          * "charge_statistics" updated event counter. Then, check it.
2617          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2618          * if they exceeds softlimit.
2619          */
2620         memcg_check_events(mem, page);
2621 }
2622
2623 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2624
2625 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2626                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2627 /*
2628  * Because tail pages are not marked as "used", set it. We're under
2629  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2630  */
2631 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2632 {
2633         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2634         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2635         unsigned long flags;
2636
2637         if (mem_cgroup_disabled())
2638                 return;
2639         /*
2640          * We have no races with charge/uncharge but will have races with
2641          * page state accounting.
2642          */
2643         move_lock_page_cgroup(head_pc, &flags);
2644
2645         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2646         smp_wmb(); /* see __commit_charge() */
2647         if (PageCgroupAcctLRU(head_pc)) {
2648                 enum lru_list lru;
2649                 struct mem_cgroup_per_zone *mz;
2650
2651                 /*
2652                  * LRU flags cannot be copied because we need to add tail
2653                  *.page to LRU by generic call and our hook will be called.
2654                  * We hold lru_lock, then, reduce counter directly.
2655                  */
2656                 lru = page_lru(head);
2657                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2658                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2659         }
2660         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2661         move_unlock_page_cgroup(head_pc, &flags);
2662 }
2663 #endif
2664
2665 /**
2666  * mem_cgroup_move_account - move account of the page
2667  * @page: the page
2668  * @nr_pages: number of regular pages (>1 for huge pages)
2669  * @pc: page_cgroup of the page.
2670  * @from: mem_cgroup which the page is moved from.
2671  * @to: mem_cgroup which the page is moved to. @from != @to.
2672  * @uncharge: whether we should call uncharge and css_put against @from.
2673  *
2674  * The caller must confirm following.
2675  * - page is not on LRU (isolate_page() is useful.)
2676  * - compound_lock is held when nr_pages > 1
2677  *
2678  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2679  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2680  * true, this function does "uncharge" from old cgroup, but it doesn't if
2681  * @uncharge is false, so a caller should do "uncharge".
2682  */
2683 static int mem_cgroup_move_account(struct page *page,
2684                                    unsigned int nr_pages,
2685                                    struct page_cgroup *pc,
2686                                    struct mem_cgroup *from,
2687                                    struct mem_cgroup *to,
2688                                    bool uncharge)
2689 {
2690         unsigned long flags;
2691         int ret;
2692
2693         VM_BUG_ON(from == to);
2694         VM_BUG_ON(PageLRU(page));
2695         /*
2696          * The page is isolated from LRU. So, collapse function
2697          * will not handle this page. But page splitting can happen.
2698          * Do this check under compound_page_lock(). The caller should
2699          * hold it.
2700          */
2701         ret = -EBUSY;
2702         if (nr_pages > 1 && !PageTransHuge(page))
2703                 goto out;
2704
2705         lock_page_cgroup(pc);
2706
2707         ret = -EINVAL;
2708         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2709                 goto unlock;
2710
2711         move_lock_page_cgroup(pc, &flags);
2712
2713         if (PageCgroupFileMapped(pc)) {
2714                 /* Update mapped_file data for mem_cgroup */
2715                 preempt_disable();
2716                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2717                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2718                 preempt_enable();
2719         }
2720         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2721         if (uncharge)
2722                 /* This is not "cancel", but cancel_charge does all we need. */
2723                 __mem_cgroup_cancel_charge(from, nr_pages);
2724
2725         /* caller should have done css_get */
2726         pc->mem_cgroup = to;
2727         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2728         /*
2729          * We charges against "to" which may not have any tasks. Then, "to"
2730          * can be under rmdir(). But in current implementation, caller of
2731          * this function is just force_empty() and move charge, so it's
2732          * guaranteed that "to" is never removed. So, we don't check rmdir
2733          * status here.
2734          */
2735         move_unlock_page_cgroup(pc, &flags);
2736         ret = 0;
2737 unlock:
2738         unlock_page_cgroup(pc);
2739         /*
2740          * check events
2741          */
2742         memcg_check_events(to, page);
2743         memcg_check_events(from, page);
2744 out:
2745         return ret;
2746 }
2747
2748 /*
2749  * move charges to its parent.
2750  */
2751
2752 static int mem_cgroup_move_parent(struct page *page,
2753                                   struct page_cgroup *pc,
2754                                   struct mem_cgroup *child,
2755                                   gfp_t gfp_mask)
2756 {
2757         struct cgroup *cg = child->css.cgroup;
2758         struct cgroup *pcg = cg->parent;
2759         struct mem_cgroup *parent;
2760         unsigned int nr_pages;
2761         unsigned long uninitialized_var(flags);
2762         int ret;
2763
2764         /* Is ROOT ? */
2765         if (!pcg)
2766                 return -EINVAL;
2767
2768         ret = -EBUSY;
2769         if (!get_page_unless_zero(page))
2770                 goto out;
2771         if (isolate_lru_page(page))
2772                 goto put;
2773
2774         nr_pages = hpage_nr_pages(page);
2775
2776         parent = mem_cgroup_from_cont(pcg);
2777         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2778         if (ret || !parent)
2779                 goto put_back;
2780
2781         if (nr_pages > 1)
2782                 flags = compound_lock_irqsave(page);
2783
2784         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2785         if (ret)
2786                 __mem_cgroup_cancel_charge(parent, nr_pages);
2787
2788         if (nr_pages > 1)
2789                 compound_unlock_irqrestore(page, flags);
2790 put_back:
2791         putback_lru_page(page);
2792 put:
2793         put_page(page);
2794 out:
2795         return ret;
2796 }
2797
2798 /*
2799  * Charge the memory controller for page usage.
2800  * Return
2801  * 0 if the charge was successful
2802  * < 0 if the cgroup is over its limit
2803  */
2804 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2805                                 gfp_t gfp_mask, enum charge_type ctype)
2806 {
2807         struct mem_cgroup *mem = NULL;
2808         unsigned int nr_pages = 1;
2809         struct page_cgroup *pc;
2810         bool oom = true;
2811         int ret;
2812
2813         if (PageTransHuge(page)) {
2814                 nr_pages <<= compound_order(page);
2815                 VM_BUG_ON(!PageTransHuge(page));
2816                 /*
2817                  * Never OOM-kill a process for a huge page.  The
2818                  * fault handler will fall back to regular pages.
2819                  */
2820                 oom = false;
2821         }
2822
2823         pc = lookup_page_cgroup(page);
2824         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2825
2826         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2827         if (ret || !mem)
2828                 return ret;
2829
2830         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2831         return 0;
2832 }
2833
2834 int mem_cgroup_newpage_charge(struct page *page,
2835                               struct mm_struct *mm, gfp_t gfp_mask)
2836 {
2837         if (mem_cgroup_disabled())
2838                 return 0;
2839         /*
2840          * If already mapped, we don't have to account.
2841          * If page cache, page->mapping has address_space.
2842          * But page->mapping may have out-of-use anon_vma pointer,
2843          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2844          * is NULL.
2845          */
2846         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2847                 return 0;
2848         if (unlikely(!mm))
2849                 mm = &init_mm;
2850         return mem_cgroup_charge_common(page, mm, gfp_mask,
2851                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2852 }
2853
2854 static void
2855 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2856                                         enum charge_type ctype);
2857
2858 static void
2859 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2860                                         enum charge_type ctype)
2861 {
2862         struct page_cgroup *pc = lookup_page_cgroup(page);
2863         /*
2864          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2865          * is already on LRU. It means the page may on some other page_cgroup's
2866          * LRU. Take care of it.
2867          */
2868         mem_cgroup_lru_del_before_commit(page);
2869         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2870         mem_cgroup_lru_add_after_commit(page);
2871         return;
2872 }
2873
2874 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2875                                 gfp_t gfp_mask)
2876 {
2877         struct mem_cgroup *mem = NULL;
2878         int ret;
2879
2880         if (mem_cgroup_disabled())
2881                 return 0;
2882         if (PageCompound(page))
2883                 return 0;
2884         /*
2885          * Corner case handling. This is called from add_to_page_cache()
2886          * in usual. But some FS (shmem) precharges this page before calling it
2887          * and call add_to_page_cache() with GFP_NOWAIT.
2888          *
2889          * For GFP_NOWAIT case, the page may be pre-charged before calling
2890          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2891          * charge twice. (It works but has to pay a bit larger cost.)
2892          * And when the page is SwapCache, it should take swap information
2893          * into account. This is under lock_page() now.
2894          */
2895         if (!(gfp_mask & __GFP_WAIT)) {
2896                 struct page_cgroup *pc;
2897
2898                 pc = lookup_page_cgroup(page);
2899                 if (!pc)
2900                         return 0;
2901                 lock_page_cgroup(pc);
2902                 if (PageCgroupUsed(pc)) {
2903                         unlock_page_cgroup(pc);
2904                         return 0;
2905                 }
2906                 unlock_page_cgroup(pc);
2907         }
2908
2909         if (unlikely(!mm))
2910                 mm = &init_mm;
2911
2912         if (page_is_file_cache(page)) {
2913                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2914                 if (ret || !mem)
2915                         return ret;
2916
2917                 /*
2918                  * FUSE reuses pages without going through the final
2919                  * put that would remove them from the LRU list, make
2920                  * sure that they get relinked properly.
2921                  */
2922                 __mem_cgroup_commit_charge_lrucare(page, mem,
2923                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2924                 return ret;
2925         }
2926         /* shmem */
2927         if (PageSwapCache(page)) {
2928                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2929                 if (!ret)
2930                         __mem_cgroup_commit_charge_swapin(page, mem,
2931                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2932         } else
2933                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2934                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2935
2936         return ret;
2937 }
2938
2939 /*
2940  * While swap-in, try_charge -> commit or cancel, the page is locked.
2941  * And when try_charge() successfully returns, one refcnt to memcg without
2942  * struct page_cgroup is acquired. This refcnt will be consumed by
2943  * "commit()" or removed by "cancel()"
2944  */
2945 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2946                                  struct page *page,
2947                                  gfp_t mask, struct mem_cgroup **ptr)
2948 {
2949         struct mem_cgroup *mem;
2950         int ret;
2951
2952         *ptr = NULL;
2953
2954         if (mem_cgroup_disabled())
2955                 return 0;
2956
2957         if (!do_swap_account)
2958                 goto charge_cur_mm;
2959         /*
2960          * A racing thread's fault, or swapoff, may have already updated
2961          * the pte, and even removed page from swap cache: in those cases
2962          * do_swap_page()'s pte_same() test will fail; but there's also a
2963          * KSM case which does need to charge the page.
2964          */
2965         if (!PageSwapCache(page))
2966                 goto charge_cur_mm;
2967         mem = try_get_mem_cgroup_from_page(page);
2968         if (!mem)
2969                 goto charge_cur_mm;
2970         *ptr = mem;
2971         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2972         css_put(&mem->css);
2973         return ret;
2974 charge_cur_mm:
2975         if (unlikely(!mm))
2976                 mm = &init_mm;
2977         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2978 }
2979
2980 static void
2981 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2982                                         enum charge_type ctype)
2983 {
2984         if (mem_cgroup_disabled())
2985                 return;
2986         if (!ptr)
2987                 return;
2988         cgroup_exclude_rmdir(&ptr->css);
2989
2990         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2991         /*
2992          * Now swap is on-memory. This means this page may be
2993          * counted both as mem and swap....double count.
2994          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2995          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2996          * may call delete_from_swap_cache() before reach here.
2997          */
2998         if (do_swap_account && PageSwapCache(page)) {
2999                 swp_entry_t ent = {.val = page_private(page)};
3000                 unsigned short id;
3001                 struct mem_cgroup *memcg;
3002
3003                 id = swap_cgroup_record(ent, 0);
3004                 rcu_read_lock();
3005                 memcg = mem_cgroup_lookup(id);
3006                 if (memcg) {
3007                         /*
3008                          * This recorded memcg can be obsolete one. So, avoid
3009                          * calling css_tryget
3010                          */
3011                         if (!mem_cgroup_is_root(memcg))
3012                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3013                         mem_cgroup_swap_statistics(memcg, false);
3014                         mem_cgroup_put(memcg);
3015                 }
3016                 rcu_read_unlock();
3017         }
3018         /*
3019          * At swapin, we may charge account against cgroup which has no tasks.
3020          * So, rmdir()->pre_destroy() can be called while we do this charge.
3021          * In that case, we need to call pre_destroy() again. check it here.
3022          */
3023         cgroup_release_and_wakeup_rmdir(&ptr->css);
3024 }
3025
3026 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
3027 {
3028         __mem_cgroup_commit_charge_swapin(page, ptr,
3029                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
3030 }
3031
3032 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3033 {
3034         if (mem_cgroup_disabled())
3035                 return;
3036         if (!mem)
3037                 return;
3038         __mem_cgroup_cancel_charge(mem, 1);
3039 }
3040
3041 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3042                                    unsigned int nr_pages,
3043                                    const enum charge_type ctype)
3044 {
3045         struct memcg_batch_info *batch = NULL;
3046         bool uncharge_memsw = true;
3047
3048         /* If swapout, usage of swap doesn't decrease */
3049         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3050                 uncharge_memsw = false;
3051
3052         batch = &current->memcg_batch;
3053         /*
3054          * In usual, we do css_get() when we remember memcg pointer.
3055          * But in this case, we keep res->usage until end of a series of
3056          * uncharges. Then, it's ok to ignore memcg's refcnt.
3057          */
3058         if (!batch->memcg)
3059                 batch->memcg = mem;
3060         /*
3061          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3062          * In those cases, all pages freed continuously can be expected to be in
3063          * the same cgroup and we have chance to coalesce uncharges.
3064          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3065          * because we want to do uncharge as soon as possible.
3066          */
3067
3068         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3069                 goto direct_uncharge;
3070
3071         if (nr_pages > 1)
3072                 goto direct_uncharge;
3073
3074         /*
3075          * In typical case, batch->memcg == mem. This means we can
3076          * merge a series of uncharges to an uncharge of res_counter.
3077          * If not, we uncharge res_counter ony by one.
3078          */
3079         if (batch->memcg != mem)
3080                 goto direct_uncharge;
3081         /* remember freed charge and uncharge it later */
3082         batch->nr_pages++;
3083         if (uncharge_memsw)
3084                 batch->memsw_nr_pages++;
3085         return;
3086 direct_uncharge:
3087         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3088         if (uncharge_memsw)
3089                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3090         if (unlikely(batch->memcg != mem))
3091                 memcg_oom_recover(mem);
3092         return;
3093 }
3094
3095 /*
3096  * uncharge if !page_mapped(page)
3097  */
3098 static struct mem_cgroup *
3099 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3100 {
3101         struct mem_cgroup *mem = NULL;
3102         unsigned int nr_pages = 1;
3103         struct page_cgroup *pc;
3104
3105         if (mem_cgroup_disabled())
3106                 return NULL;
3107
3108         if (PageSwapCache(page))
3109                 return NULL;
3110
3111         if (PageTransHuge(page)) {
3112                 nr_pages <<= compound_order(page);
3113                 VM_BUG_ON(!PageTransHuge(page));
3114         }
3115         /*
3116          * Check if our page_cgroup is valid
3117          */
3118         pc = lookup_page_cgroup(page);
3119         if (unlikely(!pc || !PageCgroupUsed(pc)))
3120                 return NULL;
3121
3122         lock_page_cgroup(pc);
3123
3124         mem = pc->mem_cgroup;
3125
3126         if (!PageCgroupUsed(pc))
3127                 goto unlock_out;
3128
3129         switch (ctype) {
3130         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3131         case MEM_CGROUP_CHARGE_TYPE_DROP:
3132                 /* See mem_cgroup_prepare_migration() */
3133                 if (page_mapped(page) || PageCgroupMigration(pc))
3134                         goto unlock_out;
3135                 break;
3136         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3137                 if (!PageAnon(page)) {  /* Shared memory */
3138                         if (page->mapping && !page_is_file_cache(page))
3139                                 goto unlock_out;
3140                 } else if (page_mapped(page)) /* Anon */
3141                                 goto unlock_out;
3142                 break;
3143         default:
3144                 break;
3145         }
3146
3147         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3148
3149         ClearPageCgroupUsed(pc);
3150         /*
3151          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3152          * freed from LRU. This is safe because uncharged page is expected not
3153          * to be reused (freed soon). Exception is SwapCache, it's handled by
3154          * special functions.
3155          */
3156
3157         unlock_page_cgroup(pc);
3158         /*
3159          * even after unlock, we have mem->res.usage here and this memcg
3160          * will never be freed.
3161          */
3162         memcg_check_events(mem, page);
3163         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3164                 mem_cgroup_swap_statistics(mem, true);
3165                 mem_cgroup_get(mem);
3166         }
3167         if (!mem_cgroup_is_root(mem))
3168                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3169
3170         return mem;
3171
3172 unlock_out:
3173         unlock_page_cgroup(pc);
3174         return NULL;
3175 }
3176
3177 void mem_cgroup_uncharge_page(struct page *page)
3178 {
3179         /* early check. */
3180         if (page_mapped(page))
3181                 return;
3182         if (page->mapping && !PageAnon(page))
3183                 return;
3184         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3185 }
3186
3187 void mem_cgroup_uncharge_cache_page(struct page *page)
3188 {
3189         VM_BUG_ON(page_mapped(page));
3190         VM_BUG_ON(page->mapping);
3191         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3192 }
3193
3194 /*
3195  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3196  * In that cases, pages are freed continuously and we can expect pages
3197  * are in the same memcg. All these calls itself limits the number of
3198  * pages freed at once, then uncharge_start/end() is called properly.
3199  * This may be called prural(2) times in a context,
3200  */
3201
3202 void mem_cgroup_uncharge_start(void)
3203 {
3204         current->memcg_batch.do_batch++;
3205         /* We can do nest. */
3206         if (current->memcg_batch.do_batch == 1) {
3207                 current->memcg_batch.memcg = NULL;
3208                 current->memcg_batch.nr_pages = 0;
3209                 current->memcg_batch.memsw_nr_pages = 0;
3210         }
3211 }
3212
3213 void mem_cgroup_uncharge_end(void)
3214 {
3215         struct memcg_batch_info *batch = &current->memcg_batch;
3216
3217         if (!batch->do_batch)
3218                 return;
3219
3220         batch->do_batch--;
3221         if (batch->do_batch) /* If stacked, do nothing. */
3222                 return;
3223
3224         if (!batch->memcg)
3225                 return;
3226         /*
3227          * This "batch->memcg" is valid without any css_get/put etc...
3228          * bacause we hide charges behind us.
3229          */
3230         if (batch->nr_pages)
3231                 res_counter_uncharge(&batch->memcg->res,
3232                                      batch->nr_pages * PAGE_SIZE);
3233         if (batch->memsw_nr_pages)
3234                 res_counter_uncharge(&batch->memcg->memsw,
3235                                      batch->memsw_nr_pages * PAGE_SIZE);
3236         memcg_oom_recover(batch->memcg);
3237         /* forget this pointer (for sanity check) */
3238         batch->memcg = NULL;
3239 }
3240
3241 #ifdef CONFIG_SWAP
3242 /*
3243  * called after __delete_from_swap_cache() and drop "page" account.
3244  * memcg information is recorded to swap_cgroup of "ent"
3245  */
3246 void
3247 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3248 {
3249         struct mem_cgroup *memcg;
3250         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3251
3252         if (!swapout) /* this was a swap cache but the swap is unused ! */
3253                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3254
3255         memcg = __mem_cgroup_uncharge_common(page, ctype);
3256
3257         /*
3258          * record memcg information,  if swapout && memcg != NULL,
3259          * mem_cgroup_get() was called in uncharge().
3260          */
3261         if (do_swap_account && swapout && memcg)
3262                 swap_cgroup_record(ent, css_id(&memcg->css));
3263 }
3264 #endif
3265
3266 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3267 /*
3268  * called from swap_entry_free(). remove record in swap_cgroup and
3269  * uncharge "memsw" account.
3270  */
3271 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3272 {
3273         struct mem_cgroup *memcg;
3274         unsigned short id;
3275
3276         if (!do_swap_account)
3277                 return;
3278
3279         id = swap_cgroup_record(ent, 0);
3280         rcu_read_lock();
3281         memcg = mem_cgroup_lookup(id);
3282         if (memcg) {
3283                 /*
3284                  * We uncharge this because swap is freed.
3285                  * This memcg can be obsolete one. We avoid calling css_tryget
3286                  */
3287                 if (!mem_cgroup_is_root(memcg))
3288                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3289                 mem_cgroup_swap_statistics(memcg, false);
3290                 mem_cgroup_put(memcg);
3291         }
3292         rcu_read_unlock();
3293 }
3294
3295 /**
3296  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3297  * @entry: swap entry to be moved
3298  * @from:  mem_cgroup which the entry is moved from
3299  * @to:  mem_cgroup which the entry is moved to
3300  * @need_fixup: whether we should fixup res_counters and refcounts.
3301  *
3302  * It succeeds only when the swap_cgroup's record for this entry is the same
3303  * as the mem_cgroup's id of @from.
3304  *
3305  * Returns 0 on success, -EINVAL on failure.
3306  *
3307  * The caller must have charged to @to, IOW, called res_counter_charge() about
3308  * both res and memsw, and called css_get().
3309  */
3310 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3311                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3312 {
3313         unsigned short old_id, new_id;
3314
3315         old_id = css_id(&from->css);
3316         new_id = css_id(&to->css);
3317
3318         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3319                 mem_cgroup_swap_statistics(from, false);
3320                 mem_cgroup_swap_statistics(to, true);
3321                 /*
3322                  * This function is only called from task migration context now.
3323                  * It postpones res_counter and refcount handling till the end
3324                  * of task migration(mem_cgroup_clear_mc()) for performance
3325                  * improvement. But we cannot postpone mem_cgroup_get(to)
3326                  * because if the process that has been moved to @to does
3327                  * swap-in, the refcount of @to might be decreased to 0.
3328                  */
3329                 mem_cgroup_get(to);
3330                 if (need_fixup) {
3331                         if (!mem_cgroup_is_root(from))
3332                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3333                         mem_cgroup_put(from);
3334                         /*
3335                          * we charged both to->res and to->memsw, so we should
3336                          * uncharge to->res.
3337                          */
3338                         if (!mem_cgroup_is_root(to))
3339                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3340                 }
3341                 return 0;
3342         }
3343         return -EINVAL;
3344 }
3345 #else
3346 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3347                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3348 {
3349         return -EINVAL;
3350 }
3351 #endif
3352
3353 /*
3354  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3355  * page belongs to.
3356  */
3357 int mem_cgroup_prepare_migration(struct page *page,
3358         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3359 {
3360         struct mem_cgroup *mem = NULL;
3361         struct page_cgroup *pc;
3362         enum charge_type ctype;
3363         int ret = 0;
3364
3365         *ptr = NULL;
3366
3367         VM_BUG_ON(PageTransHuge(page));
3368         if (mem_cgroup_disabled())
3369                 return 0;
3370
3371         pc = lookup_page_cgroup(page);
3372         lock_page_cgroup(pc);
3373         if (PageCgroupUsed(pc)) {
3374                 mem = pc->mem_cgroup;
3375                 css_get(&mem->css);
3376                 /*
3377                  * At migrating an anonymous page, its mapcount goes down
3378                  * to 0 and uncharge() will be called. But, even if it's fully
3379                  * unmapped, migration may fail and this page has to be
3380                  * charged again. We set MIGRATION flag here and delay uncharge
3381                  * until end_migration() is called
3382                  *
3383                  * Corner Case Thinking
3384                  * A)
3385                  * When the old page was mapped as Anon and it's unmap-and-freed
3386                  * while migration was ongoing.
3387                  * If unmap finds the old page, uncharge() of it will be delayed
3388                  * until end_migration(). If unmap finds a new page, it's
3389                  * uncharged when it make mapcount to be 1->0. If unmap code
3390                  * finds swap_migration_entry, the new page will not be mapped
3391                  * and end_migration() will find it(mapcount==0).
3392                  *
3393                  * B)
3394                  * When the old page was mapped but migraion fails, the kernel
3395                  * remaps it. A charge for it is kept by MIGRATION flag even
3396                  * if mapcount goes down to 0. We can do remap successfully
3397                  * without charging it again.
3398                  *
3399                  * C)
3400                  * The "old" page is under lock_page() until the end of
3401                  * migration, so, the old page itself will not be swapped-out.
3402                  * If the new page is swapped out before end_migraton, our
3403                  * hook to usual swap-out path will catch the event.
3404                  */
3405                 if (PageAnon(page))
3406                         SetPageCgroupMigration(pc);
3407         }
3408         unlock_page_cgroup(pc);
3409         /*
3410          * If the page is not charged at this point,
3411          * we return here.
3412          */
3413         if (!mem)
3414                 return 0;
3415
3416         *ptr = mem;
3417         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3418         css_put(&mem->css);/* drop extra refcnt */
3419         if (ret || *ptr == NULL) {
3420                 if (PageAnon(page)) {
3421                         lock_page_cgroup(pc);
3422                         ClearPageCgroupMigration(pc);
3423                         unlock_page_cgroup(pc);
3424                         /*
3425                          * The old page may be fully unmapped while we kept it.
3426                          */
3427                         mem_cgroup_uncharge_page(page);
3428                 }
3429                 return -ENOMEM;
3430         }
3431         /*
3432          * We charge new page before it's used/mapped. So, even if unlock_page()
3433          * is called before end_migration, we can catch all events on this new
3434          * page. In the case new page is migrated but not remapped, new page's
3435          * mapcount will be finally 0 and we call uncharge in end_migration().
3436          */
3437         pc = lookup_page_cgroup(newpage);
3438         if (PageAnon(page))
3439                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3440         else if (page_is_file_cache(page))
3441                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3442         else
3443                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3444         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3445         return ret;
3446 }
3447
3448 /* remove redundant charge if migration failed*/
3449 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3450         struct page *oldpage, struct page *newpage, bool migration_ok)
3451 {
3452         struct page *used, *unused;
3453         struct page_cgroup *pc;
3454
3455         if (!mem)
3456                 return;
3457         /* blocks rmdir() */
3458         cgroup_exclude_rmdir(&mem->css);
3459         if (!migration_ok) {
3460                 used = oldpage;
3461                 unused = newpage;
3462         } else {
3463                 used = newpage;
3464                 unused = oldpage;
3465         }
3466         /*
3467          * We disallowed uncharge of pages under migration because mapcount
3468          * of the page goes down to zero, temporarly.
3469          * Clear the flag and check the page should be charged.
3470          */
3471         pc = lookup_page_cgroup(oldpage);
3472         lock_page_cgroup(pc);
3473         ClearPageCgroupMigration(pc);
3474         unlock_page_cgroup(pc);
3475
3476         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3477
3478         /*
3479          * If a page is a file cache, radix-tree replacement is very atomic
3480          * and we can skip this check. When it was an Anon page, its mapcount
3481          * goes down to 0. But because we added MIGRATION flage, it's not
3482          * uncharged yet. There are several case but page->mapcount check
3483          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3484          * check. (see prepare_charge() also)
3485          */
3486         if (PageAnon(used))
3487                 mem_cgroup_uncharge_page(used);
3488         /*
3489          * At migration, we may charge account against cgroup which has no
3490          * tasks.
3491          * So, rmdir()->pre_destroy() can be called while we do this charge.
3492          * In that case, we need to call pre_destroy() again. check it here.
3493          */
3494         cgroup_release_and_wakeup_rmdir(&mem->css);
3495 }
3496
3497 /*
3498  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3499  * Calling hierarchical_reclaim is not enough because we should update
3500  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3501  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3502  * not from the memcg which this page would be charged to.
3503  * try_charge_swapin does all of these works properly.
3504  */
3505 int mem_cgroup_shmem_charge_fallback(struct page *page,
3506                             struct mm_struct *mm,
3507                             gfp_t gfp_mask)
3508 {
3509         struct mem_cgroup *mem;
3510         int ret;
3511
3512         if (mem_cgroup_disabled())
3513                 return 0;
3514
3515         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3516         if (!ret)
3517                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3518
3519         return ret;
3520 }
3521
3522 #ifdef CONFIG_DEBUG_VM
3523 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3524 {
3525         struct page_cgroup *pc;
3526
3527         pc = lookup_page_cgroup(page);
3528         if (likely(pc) && PageCgroupUsed(pc))
3529                 return pc;
3530         return NULL;
3531 }
3532
3533 bool mem_cgroup_bad_page_check(struct page *page)
3534 {
3535         if (mem_cgroup_disabled())
3536                 return false;
3537
3538         return lookup_page_cgroup_used(page) != NULL;
3539 }
3540
3541 void mem_cgroup_print_bad_page(struct page *page)
3542 {
3543         struct page_cgroup *pc;
3544
3545         pc = lookup_page_cgroup_used(page);
3546         if (pc) {
3547                 int ret = -1;
3548                 char *path;
3549
3550                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3551                        pc, pc->flags, pc->mem_cgroup);
3552
3553                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3554                 if (path) {
3555                         rcu_read_lock();
3556                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3557                                                         path, PATH_MAX);
3558                         rcu_read_unlock();
3559                 }
3560
3561                 printk(KERN_CONT "(%s)\n",
3562                                 (ret < 0) ? "cannot get the path" : path);
3563                 kfree(path);
3564         }
3565 }
3566 #endif
3567
3568 static DEFINE_MUTEX(set_limit_mutex);
3569
3570 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3571                                 unsigned long long val)
3572 {
3573         int retry_count;
3574         u64 memswlimit, memlimit;
3575         int ret = 0;
3576         int children = mem_cgroup_count_children(memcg);
3577         u64 curusage, oldusage;
3578         int enlarge;
3579
3580         /*
3581          * For keeping hierarchical_reclaim simple, how long we should retry
3582          * is depends on callers. We set our retry-count to be function
3583          * of # of children which we should visit in this loop.
3584          */
3585         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3586
3587         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3588
3589         enlarge = 0;
3590         while (retry_count) {
3591                 if (signal_pending(current)) {
3592                         ret = -EINTR;
3593                         break;
3594                 }
3595                 /*
3596                  * Rather than hide all in some function, I do this in
3597                  * open coded manner. You see what this really does.
3598                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3599                  */
3600                 mutex_lock(&set_limit_mutex);
3601                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3602                 if (memswlimit < val) {
3603                         ret = -EINVAL;
3604                         mutex_unlock(&set_limit_mutex);
3605                         break;
3606                 }
3607
3608                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3609                 if (memlimit < val)
3610                         enlarge = 1;
3611
3612                 ret = res_counter_set_limit(&memcg->res, val);
3613                 if (!ret) {
3614                         if (memswlimit == val)
3615                                 memcg->memsw_is_minimum = true;
3616                         else
3617                                 memcg->memsw_is_minimum = false;
3618                 }
3619                 mutex_unlock(&set_limit_mutex);
3620
3621                 if (!ret)
3622                         break;
3623
3624                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3625                                                 MEM_CGROUP_RECLAIM_SHRINK,
3626                                                 NULL);
3627                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3628                 /* Usage is reduced ? */
3629                 if (curusage >= oldusage)
3630                         retry_count--;
3631                 else
3632                         oldusage = curusage;
3633         }
3634         if (!ret && enlarge)
3635                 memcg_oom_recover(memcg);
3636
3637         return ret;
3638 }
3639
3640 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3641                                         unsigned long long val)
3642 {
3643         int retry_count;
3644         u64 memlimit, memswlimit, oldusage, curusage;
3645         int children = mem_cgroup_count_children(memcg);
3646         int ret = -EBUSY;
3647         int enlarge = 0;
3648
3649         /* see mem_cgroup_resize_res_limit */
3650         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3651         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3652         while (retry_count) {
3653                 if (signal_pending(current)) {
3654                         ret = -EINTR;
3655                         break;
3656                 }
3657                 /*
3658                  * Rather than hide all in some function, I do this in
3659                  * open coded manner. You see what this really does.
3660                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3661                  */
3662                 mutex_lock(&set_limit_mutex);
3663                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3664                 if (memlimit > val) {
3665                         ret = -EINVAL;
3666                         mutex_unlock(&set_limit_mutex);
3667                         break;
3668                 }
3669                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3670                 if (memswlimit < val)
3671                         enlarge = 1;
3672                 ret = res_counter_set_limit(&memcg->memsw, val);
3673                 if (!ret) {
3674                         if (memlimit == val)
3675                                 memcg->memsw_is_minimum = true;
3676                         else
3677                                 memcg->memsw_is_minimum = false;
3678                 }
3679                 mutex_unlock(&set_limit_mutex);
3680
3681                 if (!ret)
3682                         break;
3683
3684                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3685                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3686                                                 MEM_CGROUP_RECLAIM_SHRINK,
3687                                                 NULL);
3688                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3689                 /* Usage is reduced ? */
3690                 if (curusage >= oldusage)
3691                         retry_count--;
3692                 else
3693                         oldusage = curusage;
3694         }
3695         if (!ret && enlarge)
3696                 memcg_oom_recover(memcg);
3697         return ret;
3698 }
3699
3700 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3701                                             gfp_t gfp_mask,
3702                                             unsigned long *total_scanned)
3703 {
3704         unsigned long nr_reclaimed = 0;
3705         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3706         unsigned long reclaimed;
3707         int loop = 0;
3708         struct mem_cgroup_tree_per_zone *mctz;
3709         unsigned long long excess;
3710         unsigned long nr_scanned;
3711
3712         if (order > 0)
3713                 return 0;
3714
3715         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3716         /*
3717          * This loop can run a while, specially if mem_cgroup's continuously
3718          * keep exceeding their soft limit and putting the system under
3719          * pressure
3720          */
3721         do {
3722                 if (next_mz)
3723                         mz = next_mz;
3724                 else
3725                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3726                 if (!mz)
3727                         break;
3728
3729                 nr_scanned = 0;
3730                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3731                                                 gfp_mask,
3732                                                 MEM_CGROUP_RECLAIM_SOFT,
3733                                                 &nr_scanned);
3734                 nr_reclaimed += reclaimed;
3735                 *total_scanned += nr_scanned;
3736                 spin_lock(&mctz->lock);
3737
3738                 /*
3739                  * If we failed to reclaim anything from this memory cgroup
3740                  * it is time to move on to the next cgroup
3741                  */
3742                 next_mz = NULL;
3743                 if (!reclaimed) {
3744                         do {
3745                                 /*
3746                                  * Loop until we find yet another one.
3747                                  *
3748                                  * By the time we get the soft_limit lock
3749                                  * again, someone might have aded the
3750                                  * group back on the RB tree. Iterate to
3751                                  * make sure we get a different mem.
3752                                  * mem_cgroup_largest_soft_limit_node returns
3753                                  * NULL if no other cgroup is present on
3754                                  * the tree
3755                                  */
3756                                 next_mz =
3757                                 __mem_cgroup_largest_soft_limit_node(mctz);
3758                                 if (next_mz == mz)
3759                                         css_put(&next_mz->mem->css);
3760                                 else /* next_mz == NULL or other memcg */
3761                                         break;
3762                         } while (1);
3763                 }
3764                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3765                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3766                 /*
3767                  * One school of thought says that we should not add
3768                  * back the node to the tree if reclaim returns 0.
3769                  * But our reclaim could return 0, simply because due
3770                  * to priority we are exposing a smaller subset of
3771                  * memory to reclaim from. Consider this as a longer
3772                  * term TODO.
3773                  */
3774                 /* If excess == 0, no tree ops */
3775                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3776                 spin_unlock(&mctz->lock);
3777                 css_put(&mz->mem->css);
3778                 loop++;
3779                 /*
3780                  * Could not reclaim anything and there are no more
3781                  * mem cgroups to try or we seem to be looping without
3782                  * reclaiming anything.
3783                  */
3784                 if (!nr_reclaimed &&
3785                         (next_mz == NULL ||
3786                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3787                         break;
3788         } while (!nr_reclaimed);
3789         if (next_mz)
3790                 css_put(&next_mz->mem->css);
3791         return nr_reclaimed;
3792 }
3793
3794 /*
3795  * This routine traverse page_cgroup in given list and drop them all.
3796  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3797  */
3798 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3799                                 int node, int zid, enum lru_list lru)
3800 {
3801         struct zone *zone;
3802         struct mem_cgroup_per_zone *mz;
3803         struct page_cgroup *pc, *busy;
3804         unsigned long flags, loop;
3805         struct list_head *list;
3806         int ret = 0;
3807
3808         zone = &NODE_DATA(node)->node_zones[zid];
3809         mz = mem_cgroup_zoneinfo(mem, node, zid);
3810         list = &mz->lists[lru];
3811
3812         loop = MEM_CGROUP_ZSTAT(mz, lru);
3813         /* give some margin against EBUSY etc...*/
3814         loop += 256;
3815         busy = NULL;
3816         while (loop--) {
3817                 struct page *page;
3818
3819                 ret = 0;
3820                 spin_lock_irqsave(&zone->lru_lock, flags);
3821                 if (list_empty(list)) {
3822                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3823                         break;
3824                 }
3825                 pc = list_entry(list->prev, struct page_cgroup, lru);
3826                 if (busy == pc) {
3827                         list_move(&pc->lru, list);
3828                         busy = NULL;
3829                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3830                         continue;
3831                 }
3832                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3833
3834                 page = lookup_cgroup_page(pc);
3835
3836                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3837                 if (ret == -ENOMEM)
3838                         break;
3839
3840                 if (ret == -EBUSY || ret == -EINVAL) {
3841                         /* found lock contention or "pc" is obsolete. */
3842                         busy = pc;
3843                         cond_resched();
3844                 } else
3845                         busy = NULL;
3846         }
3847
3848         if (!ret && !list_empty(list))
3849                 return -EBUSY;
3850         return ret;
3851 }
3852
3853 /*
3854  * make mem_cgroup's charge to be 0 if there is no task.
3855  * This enables deleting this mem_cgroup.
3856  */
3857 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3858 {
3859         int ret;
3860         int node, zid, shrink;
3861         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3862         struct cgroup *cgrp = mem->css.cgroup;
3863
3864         css_get(&mem->css);
3865
3866         shrink = 0;
3867         /* should free all ? */
3868         if (free_all)
3869                 goto try_to_free;
3870 move_account:
3871         do {
3872                 ret = -EBUSY;
3873                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3874                         goto out;
3875                 ret = -EINTR;
3876                 if (signal_pending(current))
3877                         goto out;
3878                 /* This is for making all *used* pages to be on LRU. */
3879                 lru_add_drain_all();
3880                 drain_all_stock_sync(mem);
3881                 ret = 0;
3882                 mem_cgroup_start_move(mem);
3883                 for_each_node_state(node, N_HIGH_MEMORY) {
3884                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3885                                 enum lru_list l;
3886                                 for_each_lru(l) {
3887                                         ret = mem_cgroup_force_empty_list(mem,
3888                                                         node, zid, l);
3889                                         if (ret)
3890                                                 break;
3891                                 }
3892                         }
3893                         if (ret)
3894                                 break;
3895                 }
3896                 mem_cgroup_end_move(mem);
3897                 memcg_oom_recover(mem);
3898                 /* it seems parent cgroup doesn't have enough mem */
3899                 if (ret == -ENOMEM)
3900                         goto try_to_free;
3901                 cond_resched();
3902         /* "ret" should also be checked to ensure all lists are empty. */
3903         } while (mem->res.usage > 0 || ret);
3904 out:
3905         css_put(&mem->css);
3906         return ret;
3907
3908 try_to_free:
3909         /* returns EBUSY if there is a task or if we come here twice. */
3910         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3911                 ret = -EBUSY;
3912                 goto out;
3913         }
3914         /* we call try-to-free pages for make this cgroup empty */
3915         lru_add_drain_all();
3916         /* try to free all pages in this cgroup */
3917         shrink = 1;
3918         while (nr_retries && mem->res.usage > 0) {
3919                 struct memcg_scanrecord rec;
3920                 int progress;
3921
3922                 if (signal_pending(current)) {
3923                         ret = -EINTR;
3924                         goto out;
3925                 }
3926                 rec.context = SCAN_BY_SHRINK;
3927                 rec.mem = mem;
3928                 rec.root = mem;
3929                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3930                                                 false, &rec);
3931                 if (!progress) {
3932                         nr_retries--;
3933                         /* maybe some writeback is necessary */
3934                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3935                 }
3936
3937         }
3938         lru_add_drain();
3939         /* try move_account...there may be some *locked* pages. */
3940         goto move_account;
3941 }
3942
3943 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3944 {
3945         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3946 }
3947
3948
3949 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3950 {
3951         return mem_cgroup_from_cont(cont)->use_hierarchy;
3952 }
3953
3954 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3955                                         u64 val)
3956 {
3957         int retval = 0;
3958         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3959         struct cgroup *parent = cont->parent;
3960         struct mem_cgroup *parent_mem = NULL;
3961
3962         if (parent)
3963                 parent_mem = mem_cgroup_from_cont(parent);
3964
3965         cgroup_lock();
3966         /*
3967          * If parent's use_hierarchy is set, we can't make any modifications
3968          * in the child subtrees. If it is unset, then the change can
3969          * occur, provided the current cgroup has no children.
3970          *
3971          * For the root cgroup, parent_mem is NULL, we allow value to be
3972          * set if there are no children.
3973          */
3974         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3975                                 (val == 1 || val == 0)) {
3976                 if (list_empty(&cont->children))
3977                         mem->use_hierarchy = val;
3978                 else
3979                         retval = -EBUSY;
3980         } else
3981                 retval = -EINVAL;
3982         cgroup_unlock();
3983
3984         return retval;
3985 }
3986
3987
3988 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3989                                                enum mem_cgroup_stat_index idx)
3990 {
3991         struct mem_cgroup *iter;
3992         long val = 0;
3993
3994         /* Per-cpu values can be negative, use a signed accumulator */
3995         for_each_mem_cgroup_tree(iter, mem)
3996                 val += mem_cgroup_read_stat(iter, idx);
3997
3998         if (val < 0) /* race ? */
3999                 val = 0;
4000         return val;
4001 }
4002
4003 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
4004 {
4005         u64 val;
4006
4007         if (!mem_cgroup_is_root(mem)) {
4008                 if (!swap)
4009                         return res_counter_read_u64(&mem->res, RES_USAGE);
4010                 else
4011                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
4012         }
4013
4014         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
4015         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
4016
4017         if (swap)
4018                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4019
4020         return val << PAGE_SHIFT;
4021 }
4022
4023 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
4024 {
4025         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4026         u64 val;
4027         int type, name;
4028
4029         type = MEMFILE_TYPE(cft->private);
4030         name = MEMFILE_ATTR(cft->private);
4031         switch (type) {
4032         case _MEM:
4033                 if (name == RES_USAGE)
4034                         val = mem_cgroup_usage(mem, false);
4035                 else
4036                         val = res_counter_read_u64(&mem->res, name);
4037                 break;
4038         case _MEMSWAP:
4039                 if (name == RES_USAGE)
4040                         val = mem_cgroup_usage(mem, true);
4041                 else
4042                         val = res_counter_read_u64(&mem->memsw, name);
4043                 break;
4044         default:
4045                 BUG();
4046                 break;
4047         }
4048         return val;
4049 }
4050 /*
4051  * The user of this function is...
4052  * RES_LIMIT.
4053  */
4054 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4055                             const char *buffer)
4056 {
4057         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4058         int type, name;
4059         unsigned long long val;
4060         int ret;
4061
4062         type = MEMFILE_TYPE(cft->private);
4063         name = MEMFILE_ATTR(cft->private);
4064         switch (name) {
4065         case RES_LIMIT:
4066                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4067                         ret = -EINVAL;
4068                         break;
4069                 }
4070                 /* This function does all necessary parse...reuse it */
4071                 ret = res_counter_memparse_write_strategy(buffer, &val);
4072                 if (ret)
4073                         break;
4074                 if (type == _MEM)
4075                         ret = mem_cgroup_resize_limit(memcg, val);
4076                 else
4077                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
4078                 break;
4079         case RES_SOFT_LIMIT:
4080                 ret = res_counter_memparse_write_strategy(buffer, &val);
4081                 if (ret)
4082                         break;
4083                 /*
4084                  * For memsw, soft limits are hard to implement in terms
4085                  * of semantics, for now, we support soft limits for
4086                  * control without swap
4087                  */
4088                 if (type == _MEM)
4089                         ret = res_counter_set_soft_limit(&memcg->res, val);
4090                 else
4091                         ret = -EINVAL;
4092                 break;
4093         default:
4094                 ret = -EINVAL; /* should be BUG() ? */
4095                 break;
4096         }
4097         return ret;
4098 }
4099
4100 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4101                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4102 {
4103         struct cgroup *cgroup;
4104         unsigned long long min_limit, min_memsw_limit, tmp;
4105
4106         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4107         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4108         cgroup = memcg->css.cgroup;
4109         if (!memcg->use_hierarchy)
4110                 goto out;
4111
4112         while (cgroup->parent) {
4113                 cgroup = cgroup->parent;
4114                 memcg = mem_cgroup_from_cont(cgroup);
4115                 if (!memcg->use_hierarchy)
4116                         break;
4117                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4118                 min_limit = min(min_limit, tmp);
4119                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4120                 min_memsw_limit = min(min_memsw_limit, tmp);
4121         }
4122 out:
4123         *mem_limit = min_limit;
4124         *memsw_limit = min_memsw_limit;
4125         return;
4126 }
4127
4128 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4129 {
4130         struct mem_cgroup *mem;
4131         int type, name;
4132
4133         mem = mem_cgroup_from_cont(cont);
4134         type = MEMFILE_TYPE(event);
4135         name = MEMFILE_ATTR(event);
4136         switch (name) {
4137         case RES_MAX_USAGE:
4138                 if (type == _MEM)
4139                         res_counter_reset_max(&mem->res);
4140                 else
4141                         res_counter_reset_max(&mem->memsw);
4142                 break;
4143         case RES_FAILCNT:
4144                 if (type == _MEM)
4145                         res_counter_reset_failcnt(&mem->res);
4146                 else
4147                         res_counter_reset_failcnt(&mem->memsw);
4148                 break;
4149         }
4150
4151         return 0;
4152 }
4153
4154 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4155                                         struct cftype *cft)
4156 {
4157         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4158 }
4159
4160 #ifdef CONFIG_MMU
4161 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4162                                         struct cftype *cft, u64 val)
4163 {
4164         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4165
4166         if (val >= (1 << NR_MOVE_TYPE))
4167                 return -EINVAL;
4168         /*
4169          * We check this value several times in both in can_attach() and
4170          * attach(), so we need cgroup lock to prevent this value from being
4171          * inconsistent.
4172          */
4173         cgroup_lock();
4174         mem->move_charge_at_immigrate = val;
4175         cgroup_unlock();
4176
4177         return 0;
4178 }
4179 #else
4180 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4181                                         struct cftype *cft, u64 val)
4182 {
4183         return -ENOSYS;
4184 }
4185 #endif
4186
4187
4188 /* For read statistics */
4189 enum {
4190         MCS_CACHE,
4191         MCS_RSS,
4192         MCS_FILE_MAPPED,
4193         MCS_PGPGIN,
4194         MCS_PGPGOUT,
4195         MCS_SWAP,
4196         MCS_PGFAULT,
4197         MCS_PGMAJFAULT,
4198         MCS_INACTIVE_ANON,
4199         MCS_ACTIVE_ANON,
4200         MCS_INACTIVE_FILE,
4201         MCS_ACTIVE_FILE,
4202         MCS_UNEVICTABLE,
4203         NR_MCS_STAT,
4204 };
4205
4206 struct mcs_total_stat {
4207         s64 stat[NR_MCS_STAT];
4208 };
4209
4210 struct {
4211         char *local_name;
4212         char *total_name;
4213 } memcg_stat_strings[NR_MCS_STAT] = {
4214         {"cache", "total_cache"},
4215         {"rss", "total_rss"},
4216         {"mapped_file", "total_mapped_file"},
4217         {"pgpgin", "total_pgpgin"},
4218         {"pgpgout", "total_pgpgout"},
4219         {"swap", "total_swap"},
4220         {"pgfault", "total_pgfault"},
4221         {"pgmajfault", "total_pgmajfault"},
4222         {"inactive_anon", "total_inactive_anon"},
4223         {"active_anon", "total_active_anon"},
4224         {"inactive_file", "total_inactive_file"},
4225         {"active_file", "total_active_file"},
4226         {"unevictable", "total_unevictable"}
4227 };
4228
4229
4230 static void
4231 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4232 {
4233         s64 val;
4234
4235         /* per cpu stat */
4236         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4237         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4238         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4239         s->stat[MCS_RSS] += val * PAGE_SIZE;
4240         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4241         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4242         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4243         s->stat[MCS_PGPGIN] += val;
4244         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4245         s->stat[MCS_PGPGOUT] += val;
4246         if (do_swap_account) {
4247                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4248                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4249         }
4250         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4251         s->stat[MCS_PGFAULT] += val;
4252         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4253         s->stat[MCS_PGMAJFAULT] += val;
4254
4255         /* per zone stat */
4256         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4257         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4258         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4259         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4260         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4261         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4262         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4263         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4264         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4265         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4266 }
4267
4268 static void
4269 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4270 {
4271         struct mem_cgroup *iter;
4272
4273         for_each_mem_cgroup_tree(iter, mem)
4274                 mem_cgroup_get_local_stat(iter, s);
4275 }
4276
4277 #ifdef CONFIG_NUMA
4278 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4279 {
4280         int nid;
4281         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4282         unsigned long node_nr;
4283         struct cgroup *cont = m->private;
4284         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4285
4286         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4287         seq_printf(m, "total=%lu", total_nr);
4288         for_each_node_state(nid, N_HIGH_MEMORY) {
4289                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4290                 seq_printf(m, " N%d=%lu", nid, node_nr);
4291         }
4292         seq_putc(m, '\n');
4293
4294         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4295         seq_printf(m, "file=%lu", file_nr);
4296         for_each_node_state(nid, N_HIGH_MEMORY) {
4297                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4298                                 LRU_ALL_FILE);
4299                 seq_printf(m, " N%d=%lu", nid, node_nr);
4300         }
4301         seq_putc(m, '\n');
4302
4303         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4304         seq_printf(m, "anon=%lu", anon_nr);
4305         for_each_node_state(nid, N_HIGH_MEMORY) {
4306                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4307                                 LRU_ALL_ANON);
4308                 seq_printf(m, " N%d=%lu", nid, node_nr);
4309         }
4310         seq_putc(m, '\n');
4311
4312         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4313         seq_printf(m, "unevictable=%lu", unevictable_nr);
4314         for_each_node_state(nid, N_HIGH_MEMORY) {
4315                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4316                                 BIT(LRU_UNEVICTABLE));
4317                 seq_printf(m, " N%d=%lu", nid, node_nr);
4318         }
4319         seq_putc(m, '\n');
4320         return 0;
4321 }
4322 #endif /* CONFIG_NUMA */
4323
4324 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4325                                  struct cgroup_map_cb *cb)
4326 {
4327         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4328         struct mcs_total_stat mystat;
4329         int i;
4330
4331         memset(&mystat, 0, sizeof(mystat));
4332         mem_cgroup_get_local_stat(mem_cont, &mystat);
4333
4334
4335         for (i = 0; i < NR_MCS_STAT; i++) {
4336                 if (i == MCS_SWAP && !do_swap_account)
4337                         continue;
4338                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4339         }
4340
4341         /* Hierarchical information */
4342         {
4343                 unsigned long long limit, memsw_limit;
4344                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4345                 cb->fill(cb, "hierarchical_memory_limit", limit);
4346                 if (do_swap_account)
4347                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4348         }
4349
4350         memset(&mystat, 0, sizeof(mystat));
4351         mem_cgroup_get_total_stat(mem_cont, &mystat);
4352         for (i = 0; i < NR_MCS_STAT; i++) {
4353                 if (i == MCS_SWAP && !do_swap_account)
4354                         continue;
4355                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4356         }
4357
4358 #ifdef CONFIG_DEBUG_VM
4359         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4360
4361         {
4362                 int nid, zid;
4363                 struct mem_cgroup_per_zone *mz;
4364                 unsigned long recent_rotated[2] = {0, 0};
4365                 unsigned long recent_scanned[2] = {0, 0};
4366
4367                 for_each_online_node(nid)
4368                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4369                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4370
4371                                 recent_rotated[0] +=
4372                                         mz->reclaim_stat.recent_rotated[0];
4373                                 recent_rotated[1] +=
4374                                         mz->reclaim_stat.recent_rotated[1];
4375                                 recent_scanned[0] +=
4376                                         mz->reclaim_stat.recent_scanned[0];
4377                                 recent_scanned[1] +=
4378                                         mz->reclaim_stat.recent_scanned[1];
4379                         }
4380                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4381                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4382                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4383                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4384         }
4385 #endif
4386
4387         return 0;
4388 }
4389
4390 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4391 {
4392         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4393
4394         return mem_cgroup_swappiness(memcg);
4395 }
4396
4397 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4398                                        u64 val)
4399 {
4400         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4401         struct mem_cgroup *parent;
4402
4403         if (val > 100)
4404                 return -EINVAL;
4405
4406         if (cgrp->parent == NULL)
4407                 return -EINVAL;
4408
4409         parent = mem_cgroup_from_cont(cgrp->parent);
4410
4411         cgroup_lock();
4412
4413         /* If under hierarchy, only empty-root can set this value */
4414         if ((parent->use_hierarchy) ||
4415             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4416                 cgroup_unlock();
4417                 return -EINVAL;
4418         }
4419
4420         memcg->swappiness = val;
4421
4422         cgroup_unlock();
4423
4424         return 0;
4425 }
4426
4427 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4428 {
4429         struct mem_cgroup_threshold_ary *t;
4430         u64 usage;
4431         int i;
4432
4433         rcu_read_lock();
4434         if (!swap)
4435                 t = rcu_dereference(memcg->thresholds.primary);
4436         else
4437                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4438
4439         if (!t)
4440                 goto unlock;
4441
4442         usage = mem_cgroup_usage(memcg, swap);
4443
4444         /*
4445          * current_threshold points to threshold just below usage.
4446          * If it's not true, a threshold was crossed after last
4447          * call of __mem_cgroup_threshold().
4448          */
4449         i = t->current_threshold;
4450
4451         /*
4452          * Iterate backward over array of thresholds starting from
4453          * current_threshold and check if a threshold is crossed.
4454          * If none of thresholds below usage is crossed, we read
4455          * only one element of the array here.
4456          */
4457         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4458                 eventfd_signal(t->entries[i].eventfd, 1);
4459
4460         /* i = current_threshold + 1 */
4461         i++;
4462
4463         /*
4464          * Iterate forward over array of thresholds starting from
4465          * current_threshold+1 and check if a threshold is crossed.
4466          * If none of thresholds above usage is crossed, we read
4467          * only one element of the array here.
4468          */
4469         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4470                 eventfd_signal(t->entries[i].eventfd, 1);
4471
4472         /* Update current_threshold */
4473         t->current_threshold = i - 1;
4474 unlock:
4475         rcu_read_unlock();
4476 }
4477
4478 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4479 {
4480         while (memcg) {
4481                 __mem_cgroup_threshold(memcg, false);
4482                 if (do_swap_account)
4483                         __mem_cgroup_threshold(memcg, true);
4484
4485                 memcg = parent_mem_cgroup(memcg);
4486         }
4487 }
4488
4489 static int compare_thresholds(const void *a, const void *b)
4490 {
4491         const struct mem_cgroup_threshold *_a = a;
4492         const struct mem_cgroup_threshold *_b = b;
4493
4494         return _a->threshold - _b->threshold;
4495 }
4496
4497 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4498 {
4499         struct mem_cgroup_eventfd_list *ev;
4500
4501         list_for_each_entry(ev, &mem->oom_notify, list)
4502                 eventfd_signal(ev->eventfd, 1);
4503         return 0;
4504 }
4505
4506 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4507 {
4508         struct mem_cgroup *iter;
4509
4510         for_each_mem_cgroup_tree(iter, mem)
4511                 mem_cgroup_oom_notify_cb(iter);
4512 }
4513
4514 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4515         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4516 {
4517         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4518         struct mem_cgroup_thresholds *thresholds;
4519         struct mem_cgroup_threshold_ary *new;
4520         int type = MEMFILE_TYPE(cft->private);
4521         u64 threshold, usage;
4522         int i, size, ret;
4523
4524         ret = res_counter_memparse_write_strategy(args, &threshold);
4525         if (ret)
4526                 return ret;
4527
4528         mutex_lock(&memcg->thresholds_lock);
4529
4530         if (type == _MEM)
4531                 thresholds = &memcg->thresholds;
4532         else if (type == _MEMSWAP)
4533                 thresholds = &memcg->memsw_thresholds;
4534         else
4535                 BUG();
4536
4537         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4538
4539         /* Check if a threshold crossed before adding a new one */
4540         if (thresholds->primary)
4541                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4542
4543         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4544
4545         /* Allocate memory for new array of thresholds */
4546         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4547                         GFP_KERNEL);
4548         if (!new) {
4549                 ret = -ENOMEM;
4550                 goto unlock;
4551         }
4552         new->size = size;
4553
4554         /* Copy thresholds (if any) to new array */
4555         if (thresholds->primary) {
4556                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4557                                 sizeof(struct mem_cgroup_threshold));
4558         }
4559
4560         /* Add new threshold */
4561         new->entries[size - 1].eventfd = eventfd;
4562         new->entries[size - 1].threshold = threshold;
4563
4564         /* Sort thresholds. Registering of new threshold isn't time-critical */
4565         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4566                         compare_thresholds, NULL);
4567
4568         /* Find current threshold */
4569         new->current_threshold = -1;
4570         for (i = 0; i < size; i++) {
4571                 if (new->entries[i].threshold < usage) {
4572                         /*
4573                          * new->current_threshold will not be used until
4574                          * rcu_assign_pointer(), so it's safe to increment
4575                          * it here.
4576                          */
4577                         ++new->current_threshold;
4578                 }
4579         }
4580
4581         /* Free old spare buffer and save old primary buffer as spare */
4582         kfree(thresholds->spare);
4583         thresholds->spare = thresholds->primary;
4584
4585         rcu_assign_pointer(thresholds->primary, new);
4586
4587         /* To be sure that nobody uses thresholds */
4588         synchronize_rcu();
4589
4590 unlock:
4591         mutex_unlock(&memcg->thresholds_lock);
4592
4593         return ret;
4594 }
4595
4596 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4597         struct cftype *cft, struct eventfd_ctx *eventfd)
4598 {
4599         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4600         struct mem_cgroup_thresholds *thresholds;
4601         struct mem_cgroup_threshold_ary *new;
4602         int type = MEMFILE_TYPE(cft->private);
4603         u64 usage;
4604         int i, j, size;
4605
4606         mutex_lock(&memcg->thresholds_lock);
4607         if (type == _MEM)
4608                 thresholds = &memcg->thresholds;
4609         else if (type == _MEMSWAP)
4610                 thresholds = &memcg->memsw_thresholds;
4611         else
4612                 BUG();
4613
4614         /*
4615          * Something went wrong if we trying to unregister a threshold
4616          * if we don't have thresholds
4617          */
4618         BUG_ON(!thresholds);
4619
4620         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4621
4622         /* Check if a threshold crossed before removing */
4623         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4624
4625         /* Calculate new number of threshold */
4626         size = 0;
4627         for (i = 0; i < thresholds->primary->size; i++) {
4628                 if (thresholds->primary->entries[i].eventfd != eventfd)
4629                         size++;
4630         }
4631
4632         new = thresholds->spare;
4633
4634         /* Set thresholds array to NULL if we don't have thresholds */
4635         if (!size) {
4636                 kfree(new);
4637                 new = NULL;
4638                 goto swap_buffers;
4639         }
4640
4641         new->size = size;
4642
4643         /* Copy thresholds and find current threshold */
4644         new->current_threshold = -1;
4645         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4646                 if (thresholds->primary->entries[i].eventfd == eventfd)
4647                         continue;
4648
4649                 new->entries[j] = thresholds->primary->entries[i];
4650                 if (new->entries[j].threshold < usage) {
4651                         /*
4652                          * new->current_threshold will not be used
4653                          * until rcu_assign_pointer(), so it's safe to increment
4654                          * it here.
4655                          */
4656                         ++new->current_threshold;
4657                 }
4658                 j++;
4659         }
4660
4661 swap_buffers:
4662         /* Swap primary and spare array */
4663         thresholds->spare = thresholds->primary;
4664         rcu_assign_pointer(thresholds->primary, new);
4665
4666         /* To be sure that nobody uses thresholds */
4667         synchronize_rcu();
4668
4669         mutex_unlock(&memcg->thresholds_lock);
4670 }
4671
4672 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4673         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4674 {
4675         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4676         struct mem_cgroup_eventfd_list *event;
4677         int type = MEMFILE_TYPE(cft->private);
4678
4679         BUG_ON(type != _OOM_TYPE);
4680         event = kmalloc(sizeof(*event), GFP_KERNEL);
4681         if (!event)
4682                 return -ENOMEM;
4683
4684         spin_lock(&memcg_oom_lock);
4685
4686         event->eventfd = eventfd;
4687         list_add(&event->list, &memcg->oom_notify);
4688
4689         /* already in OOM ? */
4690         if (atomic_read(&memcg->under_oom))
4691                 eventfd_signal(eventfd, 1);
4692         spin_unlock(&memcg_oom_lock);
4693
4694         return 0;
4695 }
4696
4697 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4698         struct cftype *cft, struct eventfd_ctx *eventfd)
4699 {
4700         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4701         struct mem_cgroup_eventfd_list *ev, *tmp;
4702         int type = MEMFILE_TYPE(cft->private);
4703
4704         BUG_ON(type != _OOM_TYPE);
4705
4706         spin_lock(&memcg_oom_lock);
4707
4708         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4709                 if (ev->eventfd == eventfd) {
4710                         list_del(&ev->list);
4711                         kfree(ev);
4712                 }
4713         }
4714
4715         spin_unlock(&memcg_oom_lock);
4716 }
4717
4718 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4719         struct cftype *cft,  struct cgroup_map_cb *cb)
4720 {
4721         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4722
4723         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4724
4725         if (atomic_read(&mem->under_oom))
4726                 cb->fill(cb, "under_oom", 1);
4727         else
4728                 cb->fill(cb, "under_oom", 0);
4729         return 0;
4730 }
4731
4732 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4733         struct cftype *cft, u64 val)
4734 {
4735         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4736         struct mem_cgroup *parent;
4737
4738         /* cannot set to root cgroup and only 0 and 1 are allowed */
4739         if (!cgrp->parent || !((val == 0) || (val == 1)))
4740                 return -EINVAL;
4741
4742         parent = mem_cgroup_from_cont(cgrp->parent);
4743
4744         cgroup_lock();
4745         /* oom-kill-disable is a flag for subhierarchy. */
4746         if ((parent->use_hierarchy) ||
4747             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4748                 cgroup_unlock();
4749                 return -EINVAL;
4750         }
4751         mem->oom_kill_disable = val;
4752         if (!val)
4753                 memcg_oom_recover(mem);
4754         cgroup_unlock();
4755         return 0;
4756 }
4757
4758 #ifdef CONFIG_NUMA
4759 static const struct file_operations mem_control_numa_stat_file_operations = {
4760         .read = seq_read,
4761         .llseek = seq_lseek,
4762         .release = single_release,
4763 };
4764
4765 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4766 {
4767         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4768
4769         file->f_op = &mem_control_numa_stat_file_operations;
4770         return single_open(file, mem_control_numa_stat_show, cont);
4771 }
4772 #endif /* CONFIG_NUMA */
4773
4774 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4775                                 struct cftype *cft,
4776                                 struct cgroup_map_cb *cb)
4777 {
4778         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4779         char string[64];
4780         int i;
4781
4782         for (i = 0; i < NR_SCANSTATS; i++) {
4783                 strcpy(string, scanstat_string[i]);
4784                 strcat(string, SCANSTAT_WORD_LIMIT);
4785                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4786         }
4787
4788         for (i = 0; i < NR_SCANSTATS; i++) {
4789                 strcpy(string, scanstat_string[i]);
4790                 strcat(string, SCANSTAT_WORD_SYSTEM);
4791                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4792         }
4793
4794         for (i = 0; i < NR_SCANSTATS; i++) {
4795                 strcpy(string, scanstat_string[i]);
4796                 strcat(string, SCANSTAT_WORD_LIMIT);
4797                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4798                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4799         }
4800         for (i = 0; i < NR_SCANSTATS; i++) {
4801                 strcpy(string, scanstat_string[i]);
4802                 strcat(string, SCANSTAT_WORD_SYSTEM);
4803                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4804                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4805         }
4806         return 0;
4807 }
4808
4809 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4810                                 unsigned int event)
4811 {
4812         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4813
4814         spin_lock(&mem->scanstat.lock);
4815         memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4816         memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4817         spin_unlock(&mem->scanstat.lock);
4818         return 0;
4819 }
4820
4821
4822 static struct cftype mem_cgroup_files[] = {
4823         {
4824                 .name = "usage_in_bytes",
4825                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4826                 .read_u64 = mem_cgroup_read,
4827                 .register_event = mem_cgroup_usage_register_event,
4828                 .unregister_event = mem_cgroup_usage_unregister_event,
4829         },
4830         {
4831                 .name = "max_usage_in_bytes",
4832                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4833                 .trigger = mem_cgroup_reset,
4834                 .read_u64 = mem_cgroup_read,
4835         },
4836         {
4837                 .name = "limit_in_bytes",
4838                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4839                 .write_string = mem_cgroup_write,
4840                 .read_u64 = mem_cgroup_read,
4841         },
4842         {
4843                 .name = "soft_limit_in_bytes",
4844                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4845                 .write_string = mem_cgroup_write,
4846                 .read_u64 = mem_cgroup_read,
4847         },
4848         {
4849                 .name = "failcnt",
4850                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4851                 .trigger = mem_cgroup_reset,
4852                 .read_u64 = mem_cgroup_read,
4853         },
4854         {
4855                 .name = "stat",
4856                 .read_map = mem_control_stat_show,
4857         },
4858         {
4859                 .name = "force_empty",
4860                 .trigger = mem_cgroup_force_empty_write,
4861         },
4862         {
4863                 .name = "use_hierarchy",
4864                 .write_u64 = mem_cgroup_hierarchy_write,
4865                 .read_u64 = mem_cgroup_hierarchy_read,
4866         },
4867         {
4868                 .name = "swappiness",
4869                 .read_u64 = mem_cgroup_swappiness_read,
4870                 .write_u64 = mem_cgroup_swappiness_write,
4871         },
4872         {
4873                 .name = "move_charge_at_immigrate",
4874                 .read_u64 = mem_cgroup_move_charge_read,
4875                 .write_u64 = mem_cgroup_move_charge_write,
4876         },
4877         {
4878                 .name = "oom_control",
4879                 .read_map = mem_cgroup_oom_control_read,
4880                 .write_u64 = mem_cgroup_oom_control_write,
4881                 .register_event = mem_cgroup_oom_register_event,
4882                 .unregister_event = mem_cgroup_oom_unregister_event,
4883                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4884         },
4885 #ifdef CONFIG_NUMA
4886         {
4887                 .name = "numa_stat",
4888                 .open = mem_control_numa_stat_open,
4889                 .mode = S_IRUGO,
4890         },
4891 #endif
4892         {
4893                 .name = "vmscan_stat",
4894                 .read_map = mem_cgroup_vmscan_stat_read,
4895                 .trigger = mem_cgroup_reset_vmscan_stat,
4896         },
4897 };
4898
4899 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4900 static struct cftype memsw_cgroup_files[] = {
4901         {
4902                 .name = "memsw.usage_in_bytes",
4903                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4904                 .read_u64 = mem_cgroup_read,
4905                 .register_event = mem_cgroup_usage_register_event,
4906                 .unregister_event = mem_cgroup_usage_unregister_event,
4907         },
4908         {
4909                 .name = "memsw.max_usage_in_bytes",
4910                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4911                 .trigger = mem_cgroup_reset,
4912                 .read_u64 = mem_cgroup_read,
4913         },
4914         {
4915                 .name = "memsw.limit_in_bytes",
4916                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4917                 .write_string = mem_cgroup_write,
4918                 .read_u64 = mem_cgroup_read,
4919         },
4920         {
4921                 .name = "memsw.failcnt",
4922                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4923                 .trigger = mem_cgroup_reset,
4924                 .read_u64 = mem_cgroup_read,
4925         },
4926 };
4927
4928 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4929 {
4930         if (!do_swap_account)
4931                 return 0;
4932         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4933                                 ARRAY_SIZE(memsw_cgroup_files));
4934 };
4935 #else
4936 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4937 {
4938         return 0;
4939 }
4940 #endif
4941
4942 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4943 {
4944         struct mem_cgroup_per_node *pn;
4945         struct mem_cgroup_per_zone *mz;
4946         enum lru_list l;
4947         int zone, tmp = node;
4948         /*
4949          * This routine is called against possible nodes.
4950          * But it's BUG to call kmalloc() against offline node.
4951          *
4952          * TODO: this routine can waste much memory for nodes which will
4953          *       never be onlined. It's better to use memory hotplug callback
4954          *       function.
4955          */
4956         if (!node_state(node, N_NORMAL_MEMORY))
4957                 tmp = -1;
4958         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4959         if (!pn)
4960                 return 1;
4961
4962         mem->info.nodeinfo[node] = pn;
4963         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4964                 mz = &pn->zoneinfo[zone];
4965                 for_each_lru(l)
4966                         INIT_LIST_HEAD(&mz->lists[l]);
4967                 mz->usage_in_excess = 0;
4968                 mz->on_tree = false;
4969                 mz->mem = mem;
4970         }
4971         return 0;
4972 }
4973
4974 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4975 {
4976         kfree(mem->info.nodeinfo[node]);
4977 }
4978
4979 static struct mem_cgroup *mem_cgroup_alloc(void)
4980 {
4981         struct mem_cgroup *mem;
4982         int size = sizeof(struct mem_cgroup);
4983
4984         /* Can be very big if MAX_NUMNODES is very big */
4985         if (size < PAGE_SIZE)
4986                 mem = kzalloc(size, GFP_KERNEL);
4987         else
4988                 mem = vzalloc(size);
4989
4990         if (!mem)
4991                 return NULL;
4992
4993         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4994         if (!mem->stat)
4995                 goto out_free;
4996         spin_lock_init(&mem->pcp_counter_lock);
4997         return mem;
4998
4999 out_free:
5000         if (size < PAGE_SIZE)
5001                 kfree(mem);
5002         else
5003                 vfree(mem);
5004         return NULL;
5005 }
5006
5007 /*
5008  * At destroying mem_cgroup, references from swap_cgroup can remain.
5009  * (scanning all at force_empty is too costly...)
5010  *
5011  * Instead of clearing all references at force_empty, we remember
5012  * the number of reference from swap_cgroup and free mem_cgroup when
5013  * it goes down to 0.
5014  *
5015  * Removal of cgroup itself succeeds regardless of refs from swap.
5016  */
5017
5018 static void __mem_cgroup_free(struct mem_cgroup *mem)
5019 {
5020         int node;
5021
5022         mem_cgroup_remove_from_trees(mem);
5023         free_css_id(&mem_cgroup_subsys, &mem->css);
5024
5025         for_each_node_state(node, N_POSSIBLE)
5026                 free_mem_cgroup_per_zone_info(mem, node);
5027
5028         free_percpu(mem->stat);
5029         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
5030                 kfree(mem);
5031         else
5032                 vfree(mem);
5033 }
5034
5035 static void mem_cgroup_get(struct mem_cgroup *mem)
5036 {
5037         atomic_inc(&mem->refcnt);
5038 }
5039
5040 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
5041 {
5042         if (atomic_sub_and_test(count, &mem->refcnt)) {
5043                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
5044                 __mem_cgroup_free(mem);
5045                 if (parent)
5046                         mem_cgroup_put(parent);
5047         }
5048 }
5049
5050 static void mem_cgroup_put(struct mem_cgroup *mem)
5051 {
5052         __mem_cgroup_put(mem, 1);
5053 }
5054
5055 /*
5056  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5057  */
5058 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5059 {
5060         if (!mem->res.parent)
5061                 return NULL;
5062         return mem_cgroup_from_res_counter(mem->res.parent, res);
5063 }
5064
5065 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5066 static void __init enable_swap_cgroup(void)
5067 {
5068         if (!mem_cgroup_disabled() && really_do_swap_account)
5069                 do_swap_account = 1;
5070 }
5071 #else
5072 static void __init enable_swap_cgroup(void)
5073 {
5074 }
5075 #endif
5076
5077 static int mem_cgroup_soft_limit_tree_init(void)
5078 {
5079         struct mem_cgroup_tree_per_node *rtpn;
5080         struct mem_cgroup_tree_per_zone *rtpz;
5081         int tmp, node, zone;
5082
5083         for_each_node_state(node, N_POSSIBLE) {
5084                 tmp = node;
5085                 if (!node_state(node, N_NORMAL_MEMORY))
5086                         tmp = -1;
5087                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5088                 if (!rtpn)
5089                         return 1;
5090
5091                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5092
5093                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5094                         rtpz = &rtpn->rb_tree_per_zone[zone];
5095                         rtpz->rb_root = RB_ROOT;
5096                         spin_lock_init(&rtpz->lock);
5097                 }
5098         }
5099         return 0;
5100 }
5101
5102 static struct cgroup_subsys_state * __ref
5103 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5104 {
5105         struct mem_cgroup *mem, *parent;
5106         long error = -ENOMEM;
5107         int node;
5108
5109         mem = mem_cgroup_alloc();
5110         if (!mem)
5111                 return ERR_PTR(error);
5112
5113         for_each_node_state(node, N_POSSIBLE)
5114                 if (alloc_mem_cgroup_per_zone_info(mem, node))
5115                         goto free_out;
5116
5117         /* root ? */
5118         if (cont->parent == NULL) {
5119                 int cpu;
5120                 enable_swap_cgroup();
5121                 parent = NULL;
5122                 root_mem_cgroup = mem;
5123                 if (mem_cgroup_soft_limit_tree_init())
5124                         goto free_out;
5125                 for_each_possible_cpu(cpu) {
5126                         struct memcg_stock_pcp *stock =
5127                                                 &per_cpu(memcg_stock, cpu);
5128                         INIT_WORK(&stock->work, drain_local_stock);
5129                 }
5130                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5131         } else {
5132                 parent = mem_cgroup_from_cont(cont->parent);
5133                 mem->use_hierarchy = parent->use_hierarchy;
5134                 mem->oom_kill_disable = parent->oom_kill_disable;
5135         }
5136
5137         if (parent && parent->use_hierarchy) {
5138                 res_counter_init(&mem->res, &parent->res);
5139                 res_counter_init(&mem->memsw, &parent->memsw);
5140                 /*
5141                  * We increment refcnt of the parent to ensure that we can
5142                  * safely access it on res_counter_charge/uncharge.
5143                  * This refcnt will be decremented when freeing this
5144                  * mem_cgroup(see mem_cgroup_put).
5145                  */
5146                 mem_cgroup_get(parent);
5147         } else {
5148                 res_counter_init(&mem->res, NULL);
5149                 res_counter_init(&mem->memsw, NULL);
5150         }
5151         mem->last_scanned_child = 0;
5152         mem->last_scanned_node = MAX_NUMNODES;
5153         INIT_LIST_HEAD(&mem->oom_notify);
5154
5155         if (parent)
5156                 mem->swappiness = mem_cgroup_swappiness(parent);
5157         atomic_set(&mem->refcnt, 1);
5158         mem->move_charge_at_immigrate = 0;
5159         mutex_init(&mem->thresholds_lock);
5160         spin_lock_init(&mem->scanstat.lock);
5161         return &mem->css;
5162 free_out:
5163         __mem_cgroup_free(mem);
5164         root_mem_cgroup = NULL;
5165         return ERR_PTR(error);
5166 }
5167
5168 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5169                                         struct cgroup *cont)
5170 {
5171         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5172
5173         return mem_cgroup_force_empty(mem, false);
5174 }
5175
5176 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5177                                 struct cgroup *cont)
5178 {
5179         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5180
5181         mem_cgroup_put(mem);
5182 }
5183
5184 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5185                                 struct cgroup *cont)
5186 {
5187         int ret;
5188
5189         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5190                                 ARRAY_SIZE(mem_cgroup_files));
5191
5192         if (!ret)
5193                 ret = register_memsw_files(cont, ss);
5194         return ret;
5195 }
5196
5197 #ifdef CONFIG_MMU
5198 /* Handlers for move charge at task migration. */
5199 #define PRECHARGE_COUNT_AT_ONCE 256
5200 static int mem_cgroup_do_precharge(unsigned long count)
5201 {
5202         int ret = 0;
5203         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5204         struct mem_cgroup *mem = mc.to;
5205
5206         if (mem_cgroup_is_root(mem)) {
5207                 mc.precharge += count;
5208                 /* we don't need css_get for root */
5209                 return ret;
5210         }
5211         /* try to charge at once */
5212         if (count > 1) {
5213                 struct res_counter *dummy;
5214                 /*
5215                  * "mem" cannot be under rmdir() because we've already checked
5216                  * by cgroup_lock_live_cgroup() that it is not removed and we
5217                  * are still under the same cgroup_mutex. So we can postpone
5218                  * css_get().
5219                  */
5220                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5221                         goto one_by_one;
5222                 if (do_swap_account && res_counter_charge(&mem->memsw,
5223                                                 PAGE_SIZE * count, &dummy)) {
5224                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5225                         goto one_by_one;
5226                 }
5227                 mc.precharge += count;
5228                 return ret;
5229         }
5230 one_by_one:
5231         /* fall back to one by one charge */
5232         while (count--) {
5233                 if (signal_pending(current)) {
5234                         ret = -EINTR;
5235                         break;
5236                 }
5237                 if (!batch_count--) {
5238                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5239                         cond_resched();
5240                 }
5241                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5242                 if (ret || !mem)
5243                         /* mem_cgroup_clear_mc() will do uncharge later */
5244                         return -ENOMEM;
5245                 mc.precharge++;
5246         }
5247         return ret;
5248 }
5249
5250 /**
5251  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5252  * @vma: the vma the pte to be checked belongs
5253  * @addr: the address corresponding to the pte to be checked
5254  * @ptent: the pte to be checked
5255  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5256  *
5257  * Returns
5258  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5259  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5260  *     move charge. if @target is not NULL, the page is stored in target->page
5261  *     with extra refcnt got(Callers should handle it).
5262  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5263  *     target for charge migration. if @target is not NULL, the entry is stored
5264  *     in target->ent.
5265  *
5266  * Called with pte lock held.
5267  */
5268 union mc_target {
5269         struct page     *page;
5270         swp_entry_t     ent;
5271 };
5272
5273 enum mc_target_type {
5274         MC_TARGET_NONE, /* not used */
5275         MC_TARGET_PAGE,
5276         MC_TARGET_SWAP,
5277 };
5278
5279 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5280                                                 unsigned long addr, pte_t ptent)
5281 {
5282         struct page *page = vm_normal_page(vma, addr, ptent);
5283
5284         if (!page || !page_mapped(page))
5285                 return NULL;
5286         if (PageAnon(page)) {
5287                 /* we don't move shared anon */
5288                 if (!move_anon() || page_mapcount(page) > 2)
5289                         return NULL;
5290         } else if (!move_file())
5291                 /* we ignore mapcount for file pages */
5292                 return NULL;
5293         if (!get_page_unless_zero(page))
5294                 return NULL;
5295
5296         return page;
5297 }
5298
5299 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5300                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5301 {
5302         int usage_count;
5303         struct page *page = NULL;
5304         swp_entry_t ent = pte_to_swp_entry(ptent);
5305
5306         if (!move_anon() || non_swap_entry(ent))
5307                 return NULL;
5308         usage_count = mem_cgroup_count_swap_user(ent, &page);
5309         if (usage_count > 1) { /* we don't move shared anon */
5310                 if (page)
5311                         put_page(page);
5312                 return NULL;
5313         }
5314         if (do_swap_account)
5315                 entry->val = ent.val;
5316
5317         return page;
5318 }
5319
5320 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5321                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5322 {
5323         struct page *page = NULL;
5324         struct inode *inode;
5325         struct address_space *mapping;
5326         pgoff_t pgoff;
5327
5328         if (!vma->vm_file) /* anonymous vma */
5329                 return NULL;
5330         if (!move_file())
5331                 return NULL;
5332
5333         inode = vma->vm_file->f_path.dentry->d_inode;
5334         mapping = vma->vm_file->f_mapping;
5335         if (pte_none(ptent))
5336                 pgoff = linear_page_index(vma, addr);
5337         else /* pte_file(ptent) is true */
5338                 pgoff = pte_to_pgoff(ptent);
5339
5340         /* page is moved even if it's not RSS of this task(page-faulted). */
5341         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5342                 page = find_get_page(mapping, pgoff);
5343         } else { /* shmem/tmpfs file. we should take account of swap too. */
5344                 swp_entry_t ent;
5345                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5346                 if (do_swap_account)
5347                         entry->val = ent.val;
5348         }
5349
5350         return page;
5351 }
5352
5353 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5354                 unsigned long addr, pte_t ptent, union mc_target *target)
5355 {
5356         struct page *page = NULL;
5357         struct page_cgroup *pc;
5358         int ret = 0;
5359         swp_entry_t ent = { .val = 0 };
5360
5361         if (pte_present(ptent))
5362                 page = mc_handle_present_pte(vma, addr, ptent);
5363         else if (is_swap_pte(ptent))
5364                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5365         else if (pte_none(ptent) || pte_file(ptent))
5366                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5367
5368         if (!page && !ent.val)
5369                 return 0;
5370         if (page) {
5371                 pc = lookup_page_cgroup(page);
5372                 /*
5373                  * Do only loose check w/o page_cgroup lock.
5374                  * mem_cgroup_move_account() checks the pc is valid or not under
5375                  * the lock.
5376                  */
5377                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5378                         ret = MC_TARGET_PAGE;
5379                         if (target)
5380                                 target->page = page;
5381                 }
5382                 if (!ret || !target)
5383                         put_page(page);
5384         }
5385         /* There is a swap entry and a page doesn't exist or isn't charged */
5386         if (ent.val && !ret &&
5387                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5388                 ret = MC_TARGET_SWAP;
5389                 if (target)
5390                         target->ent = ent;
5391         }
5392         return ret;
5393 }
5394
5395 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5396                                         unsigned long addr, unsigned long end,
5397                                         struct mm_walk *walk)
5398 {
5399         struct vm_area_struct *vma = walk->private;
5400         pte_t *pte;
5401         spinlock_t *ptl;
5402
5403         split_huge_page_pmd(walk->mm, pmd);
5404
5405         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5406         for (; addr != end; pte++, addr += PAGE_SIZE)
5407                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5408                         mc.precharge++; /* increment precharge temporarily */
5409         pte_unmap_unlock(pte - 1, ptl);
5410         cond_resched();
5411
5412         return 0;
5413 }
5414
5415 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5416 {
5417         unsigned long precharge;
5418         struct vm_area_struct *vma;
5419
5420         down_read(&mm->mmap_sem);
5421         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5422                 struct mm_walk mem_cgroup_count_precharge_walk = {
5423                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5424                         .mm = mm,
5425                         .private = vma,
5426                 };
5427                 if (is_vm_hugetlb_page(vma))
5428                         continue;
5429                 walk_page_range(vma->vm_start, vma->vm_end,
5430                                         &mem_cgroup_count_precharge_walk);
5431         }
5432         up_read(&mm->mmap_sem);
5433
5434         precharge = mc.precharge;
5435         mc.precharge = 0;
5436
5437         return precharge;
5438 }
5439
5440 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5441 {
5442         unsigned long precharge = mem_cgroup_count_precharge(mm);
5443
5444         VM_BUG_ON(mc.moving_task);
5445         mc.moving_task = current;
5446         return mem_cgroup_do_precharge(precharge);
5447 }
5448
5449 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5450 static void __mem_cgroup_clear_mc(void)
5451 {
5452         struct mem_cgroup *from = mc.from;
5453         struct mem_cgroup *to = mc.to;
5454
5455         /* we must uncharge all the leftover precharges from mc.to */
5456         if (mc.precharge) {
5457                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5458                 mc.precharge = 0;
5459         }
5460         /*
5461          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5462          * we must uncharge here.
5463          */
5464         if (mc.moved_charge) {
5465                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5466                 mc.moved_charge = 0;
5467         }
5468         /* we must fixup refcnts and charges */
5469         if (mc.moved_swap) {
5470                 /* uncharge swap account from the old cgroup */
5471                 if (!mem_cgroup_is_root(mc.from))
5472                         res_counter_uncharge(&mc.from->memsw,
5473                                                 PAGE_SIZE * mc.moved_swap);
5474                 __mem_cgroup_put(mc.from, mc.moved_swap);
5475
5476                 if (!mem_cgroup_is_root(mc.to)) {
5477                         /*
5478                          * we charged both to->res and to->memsw, so we should
5479                          * uncharge to->res.
5480                          */
5481                         res_counter_uncharge(&mc.to->res,
5482                                                 PAGE_SIZE * mc.moved_swap);
5483                 }
5484                 /* we've already done mem_cgroup_get(mc.to) */
5485                 mc.moved_swap = 0;
5486         }
5487         memcg_oom_recover(from);
5488         memcg_oom_recover(to);
5489         wake_up_all(&mc.waitq);
5490 }
5491
5492 static void mem_cgroup_clear_mc(void)
5493 {
5494         struct mem_cgroup *from = mc.from;
5495
5496         /*
5497          * we must clear moving_task before waking up waiters at the end of
5498          * task migration.
5499          */
5500         mc.moving_task = NULL;
5501         __mem_cgroup_clear_mc();
5502         spin_lock(&mc.lock);
5503         mc.from = NULL;
5504         mc.to = NULL;
5505         spin_unlock(&mc.lock);
5506         mem_cgroup_end_move(from);
5507 }
5508
5509 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5510                                 struct cgroup *cgroup,
5511                                 struct task_struct *p)
5512 {
5513         int ret = 0;
5514         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5515
5516         if (mem->move_charge_at_immigrate) {
5517                 struct mm_struct *mm;
5518                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5519
5520                 VM_BUG_ON(from == mem);
5521
5522                 mm = get_task_mm(p);
5523                 if (!mm)
5524                         return 0;
5525                 /* We move charges only when we move a owner of the mm */
5526                 if (mm->owner == p) {
5527                         VM_BUG_ON(mc.from);
5528                         VM_BUG_ON(mc.to);
5529                         VM_BUG_ON(mc.precharge);
5530                         VM_BUG_ON(mc.moved_charge);
5531                         VM_BUG_ON(mc.moved_swap);
5532                         mem_cgroup_start_move(from);
5533                         spin_lock(&mc.lock);
5534                         mc.from = from;
5535                         mc.to = mem;
5536                         spin_unlock(&mc.lock);
5537                         /* We set mc.moving_task later */
5538
5539                         ret = mem_cgroup_precharge_mc(mm);
5540                         if (ret)
5541                                 mem_cgroup_clear_mc();
5542                 }
5543                 mmput(mm);
5544         }
5545         return ret;
5546 }
5547
5548 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5549                                 struct cgroup *cgroup,
5550                                 struct task_struct *p)
5551 {
5552         mem_cgroup_clear_mc();
5553 }
5554
5555 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5556                                 unsigned long addr, unsigned long end,
5557                                 struct mm_walk *walk)
5558 {
5559         int ret = 0;
5560         struct vm_area_struct *vma = walk->private;
5561         pte_t *pte;
5562         spinlock_t *ptl;
5563
5564         split_huge_page_pmd(walk->mm, pmd);
5565 retry:
5566         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5567         for (; addr != end; addr += PAGE_SIZE) {
5568                 pte_t ptent = *(pte++);
5569                 union mc_target target;
5570                 int type;
5571                 struct page *page;
5572                 struct page_cgroup *pc;
5573                 swp_entry_t ent;
5574
5575                 if (!mc.precharge)
5576                         break;
5577
5578                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5579                 switch (type) {
5580                 case MC_TARGET_PAGE:
5581                         page = target.page;
5582                         if (isolate_lru_page(page))
5583                                 goto put;
5584                         pc = lookup_page_cgroup(page);
5585                         if (!mem_cgroup_move_account(page, 1, pc,
5586                                                      mc.from, mc.to, false)) {
5587                                 mc.precharge--;
5588                                 /* we uncharge from mc.from later. */
5589                                 mc.moved_charge++;
5590                         }
5591                         putback_lru_page(page);
5592 put:                    /* is_target_pte_for_mc() gets the page */
5593                         put_page(page);
5594                         break;
5595                 case MC_TARGET_SWAP:
5596                         ent = target.ent;
5597                         if (!mem_cgroup_move_swap_account(ent,
5598                                                 mc.from, mc.to, false)) {
5599                                 mc.precharge--;
5600                                 /* we fixup refcnts and charges later. */
5601                                 mc.moved_swap++;
5602                         }
5603                         break;
5604                 default:
5605                         break;
5606                 }
5607         }
5608         pte_unmap_unlock(pte - 1, ptl);
5609         cond_resched();
5610
5611         if (addr != end) {
5612                 /*
5613                  * We have consumed all precharges we got in can_attach().
5614                  * We try charge one by one, but don't do any additional
5615                  * charges to mc.to if we have failed in charge once in attach()
5616                  * phase.
5617                  */
5618                 ret = mem_cgroup_do_precharge(1);
5619                 if (!ret)
5620                         goto retry;
5621         }
5622
5623         return ret;
5624 }
5625
5626 static void mem_cgroup_move_charge(struct mm_struct *mm)
5627 {
5628         struct vm_area_struct *vma;
5629
5630         lru_add_drain_all();
5631 retry:
5632         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5633                 /*
5634                  * Someone who are holding the mmap_sem might be waiting in
5635                  * waitq. So we cancel all extra charges, wake up all waiters,
5636                  * and retry. Because we cancel precharges, we might not be able
5637                  * to move enough charges, but moving charge is a best-effort
5638                  * feature anyway, so it wouldn't be a big problem.
5639                  */
5640                 __mem_cgroup_clear_mc();
5641                 cond_resched();
5642                 goto retry;
5643         }
5644         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5645                 int ret;
5646                 struct mm_walk mem_cgroup_move_charge_walk = {
5647                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5648                         .mm = mm,
5649                         .private = vma,
5650                 };
5651                 if (is_vm_hugetlb_page(vma))
5652                         continue;
5653                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5654                                                 &mem_cgroup_move_charge_walk);
5655                 if (ret)
5656                         /*
5657                          * means we have consumed all precharges and failed in
5658                          * doing additional charge. Just abandon here.
5659                          */
5660                         break;
5661         }
5662         up_read(&mm->mmap_sem);
5663 }
5664
5665 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5666                                 struct cgroup *cont,
5667                                 struct cgroup *old_cont,
5668                                 struct task_struct *p)
5669 {
5670         struct mm_struct *mm = get_task_mm(p);
5671
5672         if (mm) {
5673                 if (mc.to)
5674                         mem_cgroup_move_charge(mm);
5675                 put_swap_token(mm);
5676                 mmput(mm);
5677         }
5678         if (mc.to)
5679                 mem_cgroup_clear_mc();
5680 }
5681 #else   /* !CONFIG_MMU */
5682 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5683                                 struct cgroup *cgroup,
5684                                 struct task_struct *p)
5685 {
5686         return 0;
5687 }
5688 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5689                                 struct cgroup *cgroup,
5690                                 struct task_struct *p)
5691 {
5692 }
5693 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5694                                 struct cgroup *cont,
5695                                 struct cgroup *old_cont,
5696                                 struct task_struct *p)
5697 {
5698 }
5699 #endif
5700
5701 struct cgroup_subsys mem_cgroup_subsys = {
5702         .name = "memory",
5703         .subsys_id = mem_cgroup_subsys_id,
5704         .create = mem_cgroup_create,
5705         .pre_destroy = mem_cgroup_pre_destroy,
5706         .destroy = mem_cgroup_destroy,
5707         .populate = mem_cgroup_populate,
5708         .can_attach = mem_cgroup_can_attach,
5709         .cancel_attach = mem_cgroup_cancel_attach,
5710         .attach = mem_cgroup_move_task,
5711         .early_init = 0,
5712         .use_id = 1,
5713 };
5714
5715 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5716 static int __init enable_swap_account(char *s)
5717 {
5718         /* consider enabled if no parameter or 1 is given */
5719         if (!strcmp(s, "1"))
5720                 really_do_swap_account = 1;
5721         else if (!strcmp(s, "0"))
5722                 really_do_swap_account = 0;
5723         return 1;
5724 }
5725 __setup("swapaccount=", enable_swap_account);
5726
5727 #endif