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