cris: fix a build error in kernel/fork.c
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2 of the License, or
16  * (at your option) any later version.
17  *
18  * This program is distributed in the hope that it will be useful,
19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
21  * GNU General Public License for more details.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/shmem_fs.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
45 #include <linux/fs.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
52 #include "internal.h"
53
54 #include <asm/uaccess.h>
55
56 #include <trace/events/vmscan.h>
57
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES      5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
61
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
65
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
69 #else
70 static int really_do_swap_account __initdata = 0;
71 #endif
72
73 #else
74 #define do_swap_account         (0)
75 #endif
76
77
78 /*
79  * Statistics for memory cgroup.
80  */
81 enum mem_cgroup_stat_index {
82         /*
83          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
84          */
85         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
86         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
87         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
88         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
91         MEM_CGROUP_STAT_NSTATS,
92 };
93
94 enum mem_cgroup_events_index {
95         MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
96         MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
97         MEM_CGROUP_EVENTS_COUNT,        /* # of pages paged in/out */
98         MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
99         MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
100         MEM_CGROUP_EVENTS_NSTATS,
101 };
102 /*
103  * Per memcg event counter is incremented at every pagein/pageout. With THP,
104  * it will be incremated by the number of pages. This counter is used for
105  * for trigger some periodic events. This is straightforward and better
106  * than using jiffies etc. to handle periodic memcg event.
107  */
108 enum mem_cgroup_events_target {
109         MEM_CGROUP_TARGET_THRESH,
110         MEM_CGROUP_TARGET_SOFTLIMIT,
111         MEM_CGROUP_TARGET_NUMAINFO,
112         MEM_CGROUP_NTARGETS,
113 };
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET  (1024)
117
118 struct mem_cgroup_stat_cpu {
119         long count[MEM_CGROUP_STAT_NSTATS];
120         unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121         unsigned long targets[MEM_CGROUP_NTARGETS];
122 };
123
124 /*
125  * per-zone information in memory controller.
126  */
127 struct mem_cgroup_per_zone {
128         /*
129          * spin_lock to protect the per cgroup LRU
130          */
131         struct list_head        lists[NR_LRU_LISTS];
132         unsigned long           count[NR_LRU_LISTS];
133
134         struct zone_reclaim_stat reclaim_stat;
135         struct rb_node          tree_node;      /* RB tree node */
136         unsigned long long      usage_in_excess;/* Set to the value by which */
137                                                 /* the soft limit is exceeded*/
138         bool                    on_tree;
139         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
140                                                 /* use container_of        */
141 };
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
144
145 struct mem_cgroup_per_node {
146         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
147 };
148
149 struct mem_cgroup_lru_info {
150         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
151 };
152
153 /*
154  * Cgroups above their limits are maintained in a RB-Tree, independent of
155  * their hierarchy representation
156  */
157
158 struct mem_cgroup_tree_per_zone {
159         struct rb_root rb_root;
160         spinlock_t lock;
161 };
162
163 struct mem_cgroup_tree_per_node {
164         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
165 };
166
167 struct mem_cgroup_tree {
168         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
169 };
170
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
172
173 struct mem_cgroup_threshold {
174         struct eventfd_ctx *eventfd;
175         u64 threshold;
176 };
177
178 /* For threshold */
179 struct mem_cgroup_threshold_ary {
180         /* An array index points to threshold just below usage. */
181         int current_threshold;
182         /* Size of entries[] */
183         unsigned int size;
184         /* Array of thresholds */
185         struct mem_cgroup_threshold entries[0];
186 };
187
188 struct mem_cgroup_thresholds {
189         /* Primary thresholds array */
190         struct mem_cgroup_threshold_ary *primary;
191         /*
192          * Spare threshold array.
193          * This is needed to make mem_cgroup_unregister_event() "never fail".
194          * It must be able to store at least primary->size - 1 entries.
195          */
196         struct mem_cgroup_threshold_ary *spare;
197 };
198
199 /* for OOM */
200 struct mem_cgroup_eventfd_list {
201         struct list_head list;
202         struct eventfd_ctx *eventfd;
203 };
204
205 static void mem_cgroup_threshold(struct mem_cgroup *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
207
208 enum {
209         SCAN_BY_LIMIT,
210         SCAN_BY_SYSTEM,
211         NR_SCAN_CONTEXT,
212         SCAN_BY_SHRINK, /* not recorded now */
213 };
214
215 enum {
216         SCAN,
217         SCAN_ANON,
218         SCAN_FILE,
219         ROTATE,
220         ROTATE_ANON,
221         ROTATE_FILE,
222         FREED,
223         FREED_ANON,
224         FREED_FILE,
225         ELAPSED,
226         NR_SCANSTATS,
227 };
228
229 struct scanstat {
230         spinlock_t      lock;
231         unsigned long   stats[NR_SCAN_CONTEXT][NR_SCANSTATS];
232         unsigned long   rootstats[NR_SCAN_CONTEXT][NR_SCANSTATS];
233 };
234
235 const char *scanstat_string[NR_SCANSTATS] = {
236         "scanned_pages",
237         "scanned_anon_pages",
238         "scanned_file_pages",
239         "rotated_pages",
240         "rotated_anon_pages",
241         "rotated_file_pages",
242         "freed_pages",
243         "freed_anon_pages",
244         "freed_file_pages",
245         "elapsed_ns",
246 };
247 #define SCANSTAT_WORD_LIMIT     "_by_limit"
248 #define SCANSTAT_WORD_SYSTEM    "_by_system"
249 #define SCANSTAT_WORD_HIERARCHY "_under_hierarchy"
250
251
252 /*
253  * The memory controller data structure. The memory controller controls both
254  * page cache and RSS per cgroup. We would eventually like to provide
255  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
256  * to help the administrator determine what knobs to tune.
257  *
258  * TODO: Add a water mark for the memory controller. Reclaim will begin when
259  * we hit the water mark. May be even add a low water mark, such that
260  * no reclaim occurs from a cgroup at it's low water mark, this is
261  * a feature that will be implemented much later in the future.
262  */
263 struct mem_cgroup {
264         struct cgroup_subsys_state css;
265         /*
266          * the counter to account for memory usage
267          */
268         struct res_counter res;
269         /*
270          * the counter to account for mem+swap usage.
271          */
272         struct res_counter memsw;
273         /*
274          * Per cgroup active and inactive list, similar to the
275          * per zone LRU lists.
276          */
277         struct mem_cgroup_lru_info info;
278         /*
279          * While reclaiming in a hierarchy, we cache the last child we
280          * reclaimed from.
281          */
282         int last_scanned_child;
283         int last_scanned_node;
284 #if MAX_NUMNODES > 1
285         nodemask_t      scan_nodes;
286         atomic_t        numainfo_events;
287         atomic_t        numainfo_updating;
288 #endif
289         /*
290          * Should the accounting and control be hierarchical, per subtree?
291          */
292         bool use_hierarchy;
293
294         bool            oom_lock;
295         atomic_t        under_oom;
296
297         atomic_t        refcnt;
298
299         int     swappiness;
300         /* OOM-Killer disable */
301         int             oom_kill_disable;
302
303         /* set when res.limit == memsw.limit */
304         bool            memsw_is_minimum;
305
306         /* protect arrays of thresholds */
307         struct mutex thresholds_lock;
308
309         /* thresholds for memory usage. RCU-protected */
310         struct mem_cgroup_thresholds thresholds;
311
312         /* thresholds for mem+swap usage. RCU-protected */
313         struct mem_cgroup_thresholds memsw_thresholds;
314
315         /* For oom notifier event fd */
316         struct list_head oom_notify;
317         /* For recording LRU-scan statistics */
318         struct scanstat scanstat;
319         /*
320          * Should we move charges of a task when a task is moved into this
321          * mem_cgroup ? And what type of charges should we move ?
322          */
323         unsigned long   move_charge_at_immigrate;
324         /*
325          * percpu counter.
326          */
327         struct mem_cgroup_stat_cpu *stat;
328         /*
329          * used when a cpu is offlined or other synchronizations
330          * See mem_cgroup_read_stat().
331          */
332         struct mem_cgroup_stat_cpu nocpu_base;
333         spinlock_t pcp_counter_lock;
334 };
335
336 /* Stuffs for move charges at task migration. */
337 /*
338  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
339  * left-shifted bitmap of these types.
340  */
341 enum move_type {
342         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
343         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
344         NR_MOVE_TYPE,
345 };
346
347 /* "mc" and its members are protected by cgroup_mutex */
348 static struct move_charge_struct {
349         spinlock_t        lock; /* for from, to */
350         struct mem_cgroup *from;
351         struct mem_cgroup *to;
352         unsigned long precharge;
353         unsigned long moved_charge;
354         unsigned long moved_swap;
355         struct task_struct *moving_task;        /* a task moving charges */
356         wait_queue_head_t waitq;                /* a waitq for other context */
357 } mc = {
358         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
359         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
360 };
361
362 static bool move_anon(void)
363 {
364         return test_bit(MOVE_CHARGE_TYPE_ANON,
365                                         &mc.to->move_charge_at_immigrate);
366 }
367
368 static bool move_file(void)
369 {
370         return test_bit(MOVE_CHARGE_TYPE_FILE,
371                                         &mc.to->move_charge_at_immigrate);
372 }
373
374 /*
375  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
376  * limit reclaim to prevent infinite loops, if they ever occur.
377  */
378 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
379 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
380
381 enum charge_type {
382         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
383         MEM_CGROUP_CHARGE_TYPE_MAPPED,
384         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
385         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
386         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
387         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
388         NR_CHARGE_TYPE,
389 };
390
391 /* for encoding cft->private value on file */
392 #define _MEM                    (0)
393 #define _MEMSWAP                (1)
394 #define _OOM_TYPE               (2)
395 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
396 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
397 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
398 /* Used for OOM nofiier */
399 #define OOM_CONTROL             (0)
400
401 /*
402  * Reclaim flags for mem_cgroup_hierarchical_reclaim
403  */
404 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
405 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
406 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
407 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
408 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
409 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
410
411 static void mem_cgroup_get(struct mem_cgroup *mem);
412 static void mem_cgroup_put(struct mem_cgroup *mem);
413 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
414 static void drain_all_stock_async(struct mem_cgroup *mem);
415
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
418 {
419         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
420 }
421
422 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
423 {
424         return &mem->css;
425 }
426
427 static struct mem_cgroup_per_zone *
428 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
429 {
430         int nid = page_to_nid(page);
431         int zid = page_zonenum(page);
432
433         return mem_cgroup_zoneinfo(mem, nid, zid);
434 }
435
436 static struct mem_cgroup_tree_per_zone *
437 soft_limit_tree_node_zone(int nid, int zid)
438 {
439         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
440 }
441
442 static struct mem_cgroup_tree_per_zone *
443 soft_limit_tree_from_page(struct page *page)
444 {
445         int nid = page_to_nid(page);
446         int zid = page_zonenum(page);
447
448         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
449 }
450
451 static void
452 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
453                                 struct mem_cgroup_per_zone *mz,
454                                 struct mem_cgroup_tree_per_zone *mctz,
455                                 unsigned long long new_usage_in_excess)
456 {
457         struct rb_node **p = &mctz->rb_root.rb_node;
458         struct rb_node *parent = NULL;
459         struct mem_cgroup_per_zone *mz_node;
460
461         if (mz->on_tree)
462                 return;
463
464         mz->usage_in_excess = new_usage_in_excess;
465         if (!mz->usage_in_excess)
466                 return;
467         while (*p) {
468                 parent = *p;
469                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
470                                         tree_node);
471                 if (mz->usage_in_excess < mz_node->usage_in_excess)
472                         p = &(*p)->rb_left;
473                 /*
474                  * We can't avoid mem cgroups that are over their soft
475                  * limit by the same amount
476                  */
477                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
478                         p = &(*p)->rb_right;
479         }
480         rb_link_node(&mz->tree_node, parent, p);
481         rb_insert_color(&mz->tree_node, &mctz->rb_root);
482         mz->on_tree = true;
483 }
484
485 static void
486 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
487                                 struct mem_cgroup_per_zone *mz,
488                                 struct mem_cgroup_tree_per_zone *mctz)
489 {
490         if (!mz->on_tree)
491                 return;
492         rb_erase(&mz->tree_node, &mctz->rb_root);
493         mz->on_tree = false;
494 }
495
496 static void
497 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
498                                 struct mem_cgroup_per_zone *mz,
499                                 struct mem_cgroup_tree_per_zone *mctz)
500 {
501         spin_lock(&mctz->lock);
502         __mem_cgroup_remove_exceeded(mem, mz, mctz);
503         spin_unlock(&mctz->lock);
504 }
505
506
507 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
508 {
509         unsigned long long excess;
510         struct mem_cgroup_per_zone *mz;
511         struct mem_cgroup_tree_per_zone *mctz;
512         int nid = page_to_nid(page);
513         int zid = page_zonenum(page);
514         mctz = soft_limit_tree_from_page(page);
515
516         /*
517          * Necessary to update all ancestors when hierarchy is used.
518          * because their event counter is not touched.
519          */
520         for (; mem; mem = parent_mem_cgroup(mem)) {
521                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
522                 excess = res_counter_soft_limit_excess(&mem->res);
523                 /*
524                  * We have to update the tree if mz is on RB-tree or
525                  * mem is over its softlimit.
526                  */
527                 if (excess || mz->on_tree) {
528                         spin_lock(&mctz->lock);
529                         /* if on-tree, remove it */
530                         if (mz->on_tree)
531                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
532                         /*
533                          * Insert again. mz->usage_in_excess will be updated.
534                          * If excess is 0, no tree ops.
535                          */
536                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
537                         spin_unlock(&mctz->lock);
538                 }
539         }
540 }
541
542 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
543 {
544         int node, zone;
545         struct mem_cgroup_per_zone *mz;
546         struct mem_cgroup_tree_per_zone *mctz;
547
548         for_each_node_state(node, N_POSSIBLE) {
549                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
550                         mz = mem_cgroup_zoneinfo(mem, node, zone);
551                         mctz = soft_limit_tree_node_zone(node, zone);
552                         mem_cgroup_remove_exceeded(mem, mz, mctz);
553                 }
554         }
555 }
556
557 static struct mem_cgroup_per_zone *
558 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
559 {
560         struct rb_node *rightmost = NULL;
561         struct mem_cgroup_per_zone *mz;
562
563 retry:
564         mz = NULL;
565         rightmost = rb_last(&mctz->rb_root);
566         if (!rightmost)
567                 goto done;              /* Nothing to reclaim from */
568
569         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
570         /*
571          * Remove the node now but someone else can add it back,
572          * we will to add it back at the end of reclaim to its correct
573          * position in the tree.
574          */
575         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
576         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
577                 !css_tryget(&mz->mem->css))
578                 goto retry;
579 done:
580         return mz;
581 }
582
583 static struct mem_cgroup_per_zone *
584 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
585 {
586         struct mem_cgroup_per_zone *mz;
587
588         spin_lock(&mctz->lock);
589         mz = __mem_cgroup_largest_soft_limit_node(mctz);
590         spin_unlock(&mctz->lock);
591         return mz;
592 }
593
594 /*
595  * Implementation Note: reading percpu statistics for memcg.
596  *
597  * Both of vmstat[] and percpu_counter has threshold and do periodic
598  * synchronization to implement "quick" read. There are trade-off between
599  * reading cost and precision of value. Then, we may have a chance to implement
600  * a periodic synchronizion of counter in memcg's counter.
601  *
602  * But this _read() function is used for user interface now. The user accounts
603  * memory usage by memory cgroup and he _always_ requires exact value because
604  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
605  * have to visit all online cpus and make sum. So, for now, unnecessary
606  * synchronization is not implemented. (just implemented for cpu hotplug)
607  *
608  * If there are kernel internal actions which can make use of some not-exact
609  * value, and reading all cpu value can be performance bottleneck in some
610  * common workload, threashold and synchonization as vmstat[] should be
611  * implemented.
612  */
613 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
614                                  enum mem_cgroup_stat_index idx)
615 {
616         long val = 0;
617         int cpu;
618
619         get_online_cpus();
620         for_each_online_cpu(cpu)
621                 val += per_cpu(mem->stat->count[idx], cpu);
622 #ifdef CONFIG_HOTPLUG_CPU
623         spin_lock(&mem->pcp_counter_lock);
624         val += mem->nocpu_base.count[idx];
625         spin_unlock(&mem->pcp_counter_lock);
626 #endif
627         put_online_cpus();
628         return val;
629 }
630
631 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
632                                          bool charge)
633 {
634         int val = (charge) ? 1 : -1;
635         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
636 }
637
638 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
639 {
640         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
641 }
642
643 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
644 {
645         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
646 }
647
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
649                                             enum mem_cgroup_events_index idx)
650 {
651         unsigned long val = 0;
652         int cpu;
653
654         for_each_online_cpu(cpu)
655                 val += per_cpu(mem->stat->events[idx], cpu);
656 #ifdef CONFIG_HOTPLUG_CPU
657         spin_lock(&mem->pcp_counter_lock);
658         val += mem->nocpu_base.events[idx];
659         spin_unlock(&mem->pcp_counter_lock);
660 #endif
661         return val;
662 }
663
664 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
665                                          bool file, int nr_pages)
666 {
667         preempt_disable();
668
669         if (file)
670                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
671         else
672                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
673
674         /* pagein of a big page is an event. So, ignore page size */
675         if (nr_pages > 0)
676                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
677         else {
678                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
679                 nr_pages = -nr_pages; /* for event */
680         }
681
682         __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
683
684         preempt_enable();
685 }
686
687 unsigned long
688 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
689                         unsigned int lru_mask)
690 {
691         struct mem_cgroup_per_zone *mz;
692         enum lru_list l;
693         unsigned long ret = 0;
694
695         mz = mem_cgroup_zoneinfo(mem, nid, zid);
696
697         for_each_lru(l) {
698                 if (BIT(l) & lru_mask)
699                         ret += MEM_CGROUP_ZSTAT(mz, l);
700         }
701         return ret;
702 }
703
704 static unsigned long
705 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
706                         int nid, unsigned int lru_mask)
707 {
708         u64 total = 0;
709         int zid;
710
711         for (zid = 0; zid < MAX_NR_ZONES; zid++)
712                 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
713
714         return total;
715 }
716
717 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
718                         unsigned int lru_mask)
719 {
720         int nid;
721         u64 total = 0;
722
723         for_each_node_state(nid, N_HIGH_MEMORY)
724                 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
725         return total;
726 }
727
728 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
729 {
730         unsigned long val, next;
731
732         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
733         next = this_cpu_read(mem->stat->targets[target]);
734         /* from time_after() in jiffies.h */
735         return ((long)next - (long)val < 0);
736 }
737
738 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
739 {
740         unsigned long val, next;
741
742         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
743
744         switch (target) {
745         case MEM_CGROUP_TARGET_THRESH:
746                 next = val + THRESHOLDS_EVENTS_TARGET;
747                 break;
748         case MEM_CGROUP_TARGET_SOFTLIMIT:
749                 next = val + SOFTLIMIT_EVENTS_TARGET;
750                 break;
751         case MEM_CGROUP_TARGET_NUMAINFO:
752                 next = val + NUMAINFO_EVENTS_TARGET;
753                 break;
754         default:
755                 return;
756         }
757
758         this_cpu_write(mem->stat->targets[target], next);
759 }
760
761 /*
762  * Check events in order.
763  *
764  */
765 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
766 {
767         /* threshold event is triggered in finer grain than soft limit */
768         if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
769                 mem_cgroup_threshold(mem);
770                 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
771                 if (unlikely(__memcg_event_check(mem,
772                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
773                         mem_cgroup_update_tree(mem, page);
774                         __mem_cgroup_target_update(mem,
775                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
776                 }
777 #if MAX_NUMNODES > 1
778                 if (unlikely(__memcg_event_check(mem,
779                         MEM_CGROUP_TARGET_NUMAINFO))) {
780                         atomic_inc(&mem->numainfo_events);
781                         __mem_cgroup_target_update(mem,
782                                 MEM_CGROUP_TARGET_NUMAINFO);
783                 }
784 #endif
785         }
786 }
787
788 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
789 {
790         return container_of(cgroup_subsys_state(cont,
791                                 mem_cgroup_subsys_id), struct mem_cgroup,
792                                 css);
793 }
794
795 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
796 {
797         /*
798          * mm_update_next_owner() may clear mm->owner to NULL
799          * if it races with swapoff, page migration, etc.
800          * So this can be called with p == NULL.
801          */
802         if (unlikely(!p))
803                 return NULL;
804
805         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
806                                 struct mem_cgroup, css);
807 }
808
809 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
810 {
811         struct mem_cgroup *mem = NULL;
812
813         if (!mm)
814                 return NULL;
815         /*
816          * Because we have no locks, mm->owner's may be being moved to other
817          * cgroup. We use css_tryget() here even if this looks
818          * pessimistic (rather than adding locks here).
819          */
820         rcu_read_lock();
821         do {
822                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
823                 if (unlikely(!mem))
824                         break;
825         } while (!css_tryget(&mem->css));
826         rcu_read_unlock();
827         return mem;
828 }
829
830 /* The caller has to guarantee "mem" exists before calling this */
831 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
832 {
833         struct cgroup_subsys_state *css;
834         int found;
835
836         if (!mem) /* ROOT cgroup has the smallest ID */
837                 return root_mem_cgroup; /*css_put/get against root is ignored*/
838         if (!mem->use_hierarchy) {
839                 if (css_tryget(&mem->css))
840                         return mem;
841                 return NULL;
842         }
843         rcu_read_lock();
844         /*
845          * searching a memory cgroup which has the smallest ID under given
846          * ROOT cgroup. (ID >= 1)
847          */
848         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
849         if (css && css_tryget(css))
850                 mem = container_of(css, struct mem_cgroup, css);
851         else
852                 mem = NULL;
853         rcu_read_unlock();
854         return mem;
855 }
856
857 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
858                                         struct mem_cgroup *root,
859                                         bool cond)
860 {
861         int nextid = css_id(&iter->css) + 1;
862         int found;
863         int hierarchy_used;
864         struct cgroup_subsys_state *css;
865
866         hierarchy_used = iter->use_hierarchy;
867
868         css_put(&iter->css);
869         /* If no ROOT, walk all, ignore hierarchy */
870         if (!cond || (root && !hierarchy_used))
871                 return NULL;
872
873         if (!root)
874                 root = root_mem_cgroup;
875
876         do {
877                 iter = NULL;
878                 rcu_read_lock();
879
880                 css = css_get_next(&mem_cgroup_subsys, nextid,
881                                 &root->css, &found);
882                 if (css && css_tryget(css))
883                         iter = container_of(css, struct mem_cgroup, css);
884                 rcu_read_unlock();
885                 /* If css is NULL, no more cgroups will be found */
886                 nextid = found + 1;
887         } while (css && !iter);
888
889         return iter;
890 }
891 /*
892  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
893  * be careful that "break" loop is not allowed. We have reference count.
894  * Instead of that modify "cond" to be false and "continue" to exit the loop.
895  */
896 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
897         for (iter = mem_cgroup_start_loop(root);\
898              iter != NULL;\
899              iter = mem_cgroup_get_next(iter, root, cond))
900
901 #define for_each_mem_cgroup_tree(iter, root) \
902         for_each_mem_cgroup_tree_cond(iter, root, true)
903
904 #define for_each_mem_cgroup_all(iter) \
905         for_each_mem_cgroup_tree_cond(iter, NULL, true)
906
907
908 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
909 {
910         return (mem == root_mem_cgroup);
911 }
912
913 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
914 {
915         struct mem_cgroup *mem;
916
917         if (!mm)
918                 return;
919
920         rcu_read_lock();
921         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
922         if (unlikely(!mem))
923                 goto out;
924
925         switch (idx) {
926         case PGMAJFAULT:
927                 mem_cgroup_pgmajfault(mem, 1);
928                 break;
929         case PGFAULT:
930                 mem_cgroup_pgfault(mem, 1);
931                 break;
932         default:
933                 BUG();
934         }
935 out:
936         rcu_read_unlock();
937 }
938 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
939
940 /*
941  * Following LRU functions are allowed to be used without PCG_LOCK.
942  * Operations are called by routine of global LRU independently from memcg.
943  * What we have to take care of here is validness of pc->mem_cgroup.
944  *
945  * Changes to pc->mem_cgroup happens when
946  * 1. charge
947  * 2. moving account
948  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
949  * It is added to LRU before charge.
950  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
951  * When moving account, the page is not on LRU. It's isolated.
952  */
953
954 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
955 {
956         struct page_cgroup *pc;
957         struct mem_cgroup_per_zone *mz;
958
959         if (mem_cgroup_disabled())
960                 return;
961         pc = lookup_page_cgroup(page);
962         /* can happen while we handle swapcache. */
963         if (!TestClearPageCgroupAcctLRU(pc))
964                 return;
965         VM_BUG_ON(!pc->mem_cgroup);
966         /*
967          * We don't check PCG_USED bit. It's cleared when the "page" is finally
968          * removed from global LRU.
969          */
970         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
971         /* huge page split is done under lru_lock. so, we have no races. */
972         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
973         if (mem_cgroup_is_root(pc->mem_cgroup))
974                 return;
975         VM_BUG_ON(list_empty(&pc->lru));
976         list_del_init(&pc->lru);
977 }
978
979 void mem_cgroup_del_lru(struct page *page)
980 {
981         mem_cgroup_del_lru_list(page, page_lru(page));
982 }
983
984 /*
985  * Writeback is about to end against a page which has been marked for immediate
986  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
987  * inactive list.
988  */
989 void mem_cgroup_rotate_reclaimable_page(struct page *page)
990 {
991         struct mem_cgroup_per_zone *mz;
992         struct page_cgroup *pc;
993         enum lru_list lru = page_lru(page);
994
995         if (mem_cgroup_disabled())
996                 return;
997
998         pc = lookup_page_cgroup(page);
999         /* unused or root page is not rotated. */
1000         if (!PageCgroupUsed(pc))
1001                 return;
1002         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1003         smp_rmb();
1004         if (mem_cgroup_is_root(pc->mem_cgroup))
1005                 return;
1006         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1007         list_move_tail(&pc->lru, &mz->lists[lru]);
1008 }
1009
1010 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1011 {
1012         struct mem_cgroup_per_zone *mz;
1013         struct page_cgroup *pc;
1014
1015         if (mem_cgroup_disabled())
1016                 return;
1017
1018         pc = lookup_page_cgroup(page);
1019         /* unused or root page is not rotated. */
1020         if (!PageCgroupUsed(pc))
1021                 return;
1022         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1023         smp_rmb();
1024         if (mem_cgroup_is_root(pc->mem_cgroup))
1025                 return;
1026         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1027         list_move(&pc->lru, &mz->lists[lru]);
1028 }
1029
1030 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1031 {
1032         struct page_cgroup *pc;
1033         struct mem_cgroup_per_zone *mz;
1034
1035         if (mem_cgroup_disabled())
1036                 return;
1037         pc = lookup_page_cgroup(page);
1038         VM_BUG_ON(PageCgroupAcctLRU(pc));
1039         if (!PageCgroupUsed(pc))
1040                 return;
1041         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1042         smp_rmb();
1043         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1044         /* huge page split is done under lru_lock. so, we have no races. */
1045         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1046         SetPageCgroupAcctLRU(pc);
1047         if (mem_cgroup_is_root(pc->mem_cgroup))
1048                 return;
1049         list_add(&pc->lru, &mz->lists[lru]);
1050 }
1051
1052 /*
1053  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1054  * while it's linked to lru because the page may be reused after it's fully
1055  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1056  * It's done under lock_page and expected that zone->lru_lock isnever held.
1057  */
1058 static void mem_cgroup_lru_del_before_commit(struct page *page)
1059 {
1060         unsigned long flags;
1061         struct zone *zone = page_zone(page);
1062         struct page_cgroup *pc = lookup_page_cgroup(page);
1063
1064         /*
1065          * Doing this check without taking ->lru_lock seems wrong but this
1066          * is safe. Because if page_cgroup's USED bit is unset, the page
1067          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1068          * set, the commit after this will fail, anyway.
1069          * This all charge/uncharge is done under some mutual execustion.
1070          * So, we don't need to taking care of changes in USED bit.
1071          */
1072         if (likely(!PageLRU(page)))
1073                 return;
1074
1075         spin_lock_irqsave(&zone->lru_lock, flags);
1076         /*
1077          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1078          * is guarded by lock_page() because the page is SwapCache.
1079          */
1080         if (!PageCgroupUsed(pc))
1081                 mem_cgroup_del_lru_list(page, page_lru(page));
1082         spin_unlock_irqrestore(&zone->lru_lock, flags);
1083 }
1084
1085 static void mem_cgroup_lru_add_after_commit(struct page *page)
1086 {
1087         unsigned long flags;
1088         struct zone *zone = page_zone(page);
1089         struct page_cgroup *pc = lookup_page_cgroup(page);
1090
1091         /* taking care of that the page is added to LRU while we commit it */
1092         if (likely(!PageLRU(page)))
1093                 return;
1094         spin_lock_irqsave(&zone->lru_lock, flags);
1095         /* link when the page is linked to LRU but page_cgroup isn't */
1096         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1097                 mem_cgroup_add_lru_list(page, page_lru(page));
1098         spin_unlock_irqrestore(&zone->lru_lock, flags);
1099 }
1100
1101
1102 void mem_cgroup_move_lists(struct page *page,
1103                            enum lru_list from, enum lru_list to)
1104 {
1105         if (mem_cgroup_disabled())
1106                 return;
1107         mem_cgroup_del_lru_list(page, from);
1108         mem_cgroup_add_lru_list(page, to);
1109 }
1110
1111 /*
1112  * Checks whether given mem is same or in the root_mem's
1113  * hierarchy subtree
1114  */
1115 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_mem,
1116                 struct mem_cgroup *mem)
1117 {
1118         if (root_mem != mem) {
1119                 return (root_mem->use_hierarchy &&
1120                         css_is_ancestor(&mem->css, &root_mem->css));
1121         }
1122
1123         return true;
1124 }
1125
1126 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1127 {
1128         int ret;
1129         struct mem_cgroup *curr = NULL;
1130         struct task_struct *p;
1131
1132         p = find_lock_task_mm(task);
1133         if (!p)
1134                 return 0;
1135         curr = try_get_mem_cgroup_from_mm(p->mm);
1136         task_unlock(p);
1137         if (!curr)
1138                 return 0;
1139         /*
1140          * We should check use_hierarchy of "mem" not "curr". Because checking
1141          * use_hierarchy of "curr" here make this function true if hierarchy is
1142          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1143          * hierarchy(even if use_hierarchy is disabled in "mem").
1144          */
1145         ret = mem_cgroup_same_or_subtree(mem, curr);
1146         css_put(&curr->css);
1147         return ret;
1148 }
1149
1150 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1151 {
1152         unsigned long active;
1153         unsigned long inactive;
1154         unsigned long gb;
1155         unsigned long inactive_ratio;
1156
1157         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1158         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1159
1160         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1161         if (gb)
1162                 inactive_ratio = int_sqrt(10 * gb);
1163         else
1164                 inactive_ratio = 1;
1165
1166         if (present_pages) {
1167                 present_pages[0] = inactive;
1168                 present_pages[1] = active;
1169         }
1170
1171         return inactive_ratio;
1172 }
1173
1174 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1175 {
1176         unsigned long active;
1177         unsigned long inactive;
1178         unsigned long present_pages[2];
1179         unsigned long inactive_ratio;
1180
1181         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1182
1183         inactive = present_pages[0];
1184         active = present_pages[1];
1185
1186         if (inactive * inactive_ratio < active)
1187                 return 1;
1188
1189         return 0;
1190 }
1191
1192 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1193 {
1194         unsigned long active;
1195         unsigned long inactive;
1196
1197         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1198         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1199
1200         return (active > inactive);
1201 }
1202
1203 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1204                                                       struct zone *zone)
1205 {
1206         int nid = zone_to_nid(zone);
1207         int zid = zone_idx(zone);
1208         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1209
1210         return &mz->reclaim_stat;
1211 }
1212
1213 struct zone_reclaim_stat *
1214 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1215 {
1216         struct page_cgroup *pc;
1217         struct mem_cgroup_per_zone *mz;
1218
1219         if (mem_cgroup_disabled())
1220                 return NULL;
1221
1222         pc = lookup_page_cgroup(page);
1223         if (!PageCgroupUsed(pc))
1224                 return NULL;
1225         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1226         smp_rmb();
1227         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1228         return &mz->reclaim_stat;
1229 }
1230
1231 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1232                                         struct list_head *dst,
1233                                         unsigned long *scanned, int order,
1234                                         int mode, struct zone *z,
1235                                         struct mem_cgroup *mem_cont,
1236                                         int active, int file)
1237 {
1238         unsigned long nr_taken = 0;
1239         struct page *page;
1240         unsigned long scan;
1241         LIST_HEAD(pc_list);
1242         struct list_head *src;
1243         struct page_cgroup *pc, *tmp;
1244         int nid = zone_to_nid(z);
1245         int zid = zone_idx(z);
1246         struct mem_cgroup_per_zone *mz;
1247         int lru = LRU_FILE * file + active;
1248         int ret;
1249
1250         BUG_ON(!mem_cont);
1251         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1252         src = &mz->lists[lru];
1253
1254         scan = 0;
1255         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1256                 if (scan >= nr_to_scan)
1257                         break;
1258
1259                 if (unlikely(!PageCgroupUsed(pc)))
1260                         continue;
1261
1262                 page = lookup_cgroup_page(pc);
1263
1264                 if (unlikely(!PageLRU(page)))
1265                         continue;
1266
1267                 scan++;
1268                 ret = __isolate_lru_page(page, mode, file);
1269                 switch (ret) {
1270                 case 0:
1271                         list_move(&page->lru, dst);
1272                         mem_cgroup_del_lru(page);
1273                         nr_taken += hpage_nr_pages(page);
1274                         break;
1275                 case -EBUSY:
1276                         /* we don't affect global LRU but rotate in our LRU */
1277                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1278                         break;
1279                 default:
1280                         break;
1281                 }
1282         }
1283
1284         *scanned = scan;
1285
1286         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1287                                       0, 0, 0, mode);
1288
1289         return nr_taken;
1290 }
1291
1292 #define mem_cgroup_from_res_counter(counter, member)    \
1293         container_of(counter, struct mem_cgroup, member)
1294
1295 /**
1296  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1297  * @mem: the memory cgroup
1298  *
1299  * Returns the maximum amount of memory @mem can be charged with, in
1300  * pages.
1301  */
1302 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1303 {
1304         unsigned long long margin;
1305
1306         margin = res_counter_margin(&mem->res);
1307         if (do_swap_account)
1308                 margin = min(margin, res_counter_margin(&mem->memsw));
1309         return margin >> PAGE_SHIFT;
1310 }
1311
1312 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1313 {
1314         struct cgroup *cgrp = memcg->css.cgroup;
1315
1316         /* root ? */
1317         if (cgrp->parent == NULL)
1318                 return vm_swappiness;
1319
1320         return memcg->swappiness;
1321 }
1322
1323 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1324 {
1325         int cpu;
1326
1327         get_online_cpus();
1328         spin_lock(&mem->pcp_counter_lock);
1329         for_each_online_cpu(cpu)
1330                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1331         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1332         spin_unlock(&mem->pcp_counter_lock);
1333         put_online_cpus();
1334
1335         synchronize_rcu();
1336 }
1337
1338 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1339 {
1340         int cpu;
1341
1342         if (!mem)
1343                 return;
1344         get_online_cpus();
1345         spin_lock(&mem->pcp_counter_lock);
1346         for_each_online_cpu(cpu)
1347                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1348         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1349         spin_unlock(&mem->pcp_counter_lock);
1350         put_online_cpus();
1351 }
1352 /*
1353  * 2 routines for checking "mem" is under move_account() or not.
1354  *
1355  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1356  *                        for avoiding race in accounting. If true,
1357  *                        pc->mem_cgroup may be overwritten.
1358  *
1359  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1360  *                        under hierarchy of moving cgroups. This is for
1361  *                        waiting at hith-memory prressure caused by "move".
1362  */
1363
1364 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1365 {
1366         VM_BUG_ON(!rcu_read_lock_held());
1367         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1368 }
1369
1370 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1371 {
1372         struct mem_cgroup *from;
1373         struct mem_cgroup *to;
1374         bool ret = false;
1375         /*
1376          * Unlike task_move routines, we access mc.to, mc.from not under
1377          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1378          */
1379         spin_lock(&mc.lock);
1380         from = mc.from;
1381         to = mc.to;
1382         if (!from)
1383                 goto unlock;
1384
1385         ret = mem_cgroup_same_or_subtree(mem, from)
1386                 || mem_cgroup_same_or_subtree(mem, to);
1387 unlock:
1388         spin_unlock(&mc.lock);
1389         return ret;
1390 }
1391
1392 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1393 {
1394         if (mc.moving_task && current != mc.moving_task) {
1395                 if (mem_cgroup_under_move(mem)) {
1396                         DEFINE_WAIT(wait);
1397                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1398                         /* moving charge context might have finished. */
1399                         if (mc.moving_task)
1400                                 schedule();
1401                         finish_wait(&mc.waitq, &wait);
1402                         return true;
1403                 }
1404         }
1405         return false;
1406 }
1407
1408 /**
1409  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1410  * @memcg: The memory cgroup that went over limit
1411  * @p: Task that is going to be killed
1412  *
1413  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1414  * enabled
1415  */
1416 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1417 {
1418         struct cgroup *task_cgrp;
1419         struct cgroup *mem_cgrp;
1420         /*
1421          * Need a buffer in BSS, can't rely on allocations. The code relies
1422          * on the assumption that OOM is serialized for memory controller.
1423          * If this assumption is broken, revisit this code.
1424          */
1425         static char memcg_name[PATH_MAX];
1426         int ret;
1427
1428         if (!memcg || !p)
1429                 return;
1430
1431
1432         rcu_read_lock();
1433
1434         mem_cgrp = memcg->css.cgroup;
1435         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1436
1437         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1438         if (ret < 0) {
1439                 /*
1440                  * Unfortunately, we are unable to convert to a useful name
1441                  * But we'll still print out the usage information
1442                  */
1443                 rcu_read_unlock();
1444                 goto done;
1445         }
1446         rcu_read_unlock();
1447
1448         printk(KERN_INFO "Task in %s killed", memcg_name);
1449
1450         rcu_read_lock();
1451         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1452         if (ret < 0) {
1453                 rcu_read_unlock();
1454                 goto done;
1455         }
1456         rcu_read_unlock();
1457
1458         /*
1459          * Continues from above, so we don't need an KERN_ level
1460          */
1461         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1462 done:
1463
1464         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1465                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1466                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1467                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1468         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1469                 "failcnt %llu\n",
1470                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1471                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1472                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1473 }
1474
1475 /*
1476  * This function returns the number of memcg under hierarchy tree. Returns
1477  * 1(self count) if no children.
1478  */
1479 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1480 {
1481         int num = 0;
1482         struct mem_cgroup *iter;
1483
1484         for_each_mem_cgroup_tree(iter, mem)
1485                 num++;
1486         return num;
1487 }
1488
1489 /*
1490  * Return the memory (and swap, if configured) limit for a memcg.
1491  */
1492 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1493 {
1494         u64 limit;
1495         u64 memsw;
1496
1497         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1498         limit += total_swap_pages << PAGE_SHIFT;
1499
1500         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1501         /*
1502          * If memsw is finite and limits the amount of swap space available
1503          * to this memcg, return that limit.
1504          */
1505         return min(limit, memsw);
1506 }
1507
1508 /*
1509  * Visit the first child (need not be the first child as per the ordering
1510  * of the cgroup list, since we track last_scanned_child) of @mem and use
1511  * that to reclaim free pages from.
1512  */
1513 static struct mem_cgroup *
1514 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1515 {
1516         struct mem_cgroup *ret = NULL;
1517         struct cgroup_subsys_state *css;
1518         int nextid, found;
1519
1520         if (!root_mem->use_hierarchy) {
1521                 css_get(&root_mem->css);
1522                 ret = root_mem;
1523         }
1524
1525         while (!ret) {
1526                 rcu_read_lock();
1527                 nextid = root_mem->last_scanned_child + 1;
1528                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1529                                    &found);
1530                 if (css && css_tryget(css))
1531                         ret = container_of(css, struct mem_cgroup, css);
1532
1533                 rcu_read_unlock();
1534                 /* Updates scanning parameter */
1535                 if (!css) {
1536                         /* this means start scan from ID:1 */
1537                         root_mem->last_scanned_child = 0;
1538                 } else
1539                         root_mem->last_scanned_child = found;
1540         }
1541
1542         return ret;
1543 }
1544
1545 /**
1546  * test_mem_cgroup_node_reclaimable
1547  * @mem: the target memcg
1548  * @nid: the node ID to be checked.
1549  * @noswap : specify true here if the user wants flle only information.
1550  *
1551  * This function returns whether the specified memcg contains any
1552  * reclaimable pages on a node. Returns true if there are any reclaimable
1553  * pages in the node.
1554  */
1555 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1556                 int nid, bool noswap)
1557 {
1558         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1559                 return true;
1560         if (noswap || !total_swap_pages)
1561                 return false;
1562         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1563                 return true;
1564         return false;
1565
1566 }
1567 #if MAX_NUMNODES > 1
1568
1569 /*
1570  * Always updating the nodemask is not very good - even if we have an empty
1571  * list or the wrong list here, we can start from some node and traverse all
1572  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1573  *
1574  */
1575 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1576 {
1577         int nid;
1578         /*
1579          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1580          * pagein/pageout changes since the last update.
1581          */
1582         if (!atomic_read(&mem->numainfo_events))
1583                 return;
1584         if (atomic_inc_return(&mem->numainfo_updating) > 1)
1585                 return;
1586
1587         /* make a nodemask where this memcg uses memory from */
1588         mem->scan_nodes = node_states[N_HIGH_MEMORY];
1589
1590         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1591
1592                 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1593                         node_clear(nid, mem->scan_nodes);
1594         }
1595
1596         atomic_set(&mem->numainfo_events, 0);
1597         atomic_set(&mem->numainfo_updating, 0);
1598 }
1599
1600 /*
1601  * Selecting a node where we start reclaim from. Because what we need is just
1602  * reducing usage counter, start from anywhere is O,K. Considering
1603  * memory reclaim from current node, there are pros. and cons.
1604  *
1605  * Freeing memory from current node means freeing memory from a node which
1606  * we'll use or we've used. So, it may make LRU bad. And if several threads
1607  * hit limits, it will see a contention on a node. But freeing from remote
1608  * node means more costs for memory reclaim because of memory latency.
1609  *
1610  * Now, we use round-robin. Better algorithm is welcomed.
1611  */
1612 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1613 {
1614         int node;
1615
1616         mem_cgroup_may_update_nodemask(mem);
1617         node = mem->last_scanned_node;
1618
1619         node = next_node(node, mem->scan_nodes);
1620         if (node == MAX_NUMNODES)
1621                 node = first_node(mem->scan_nodes);
1622         /*
1623          * We call this when we hit limit, not when pages are added to LRU.
1624          * No LRU may hold pages because all pages are UNEVICTABLE or
1625          * memcg is too small and all pages are not on LRU. In that case,
1626          * we use curret node.
1627          */
1628         if (unlikely(node == MAX_NUMNODES))
1629                 node = numa_node_id();
1630
1631         mem->last_scanned_node = node;
1632         return node;
1633 }
1634
1635 /*
1636  * Check all nodes whether it contains reclaimable pages or not.
1637  * For quick scan, we make use of scan_nodes. This will allow us to skip
1638  * unused nodes. But scan_nodes is lazily updated and may not cotain
1639  * enough new information. We need to do double check.
1640  */
1641 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1642 {
1643         int nid;
1644
1645         /*
1646          * quick check...making use of scan_node.
1647          * We can skip unused nodes.
1648          */
1649         if (!nodes_empty(mem->scan_nodes)) {
1650                 for (nid = first_node(mem->scan_nodes);
1651                      nid < MAX_NUMNODES;
1652                      nid = next_node(nid, mem->scan_nodes)) {
1653
1654                         if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1655                                 return true;
1656                 }
1657         }
1658         /*
1659          * Check rest of nodes.
1660          */
1661         for_each_node_state(nid, N_HIGH_MEMORY) {
1662                 if (node_isset(nid, mem->scan_nodes))
1663                         continue;
1664                 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1665                         return true;
1666         }
1667         return false;
1668 }
1669
1670 #else
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1672 {
1673         return 0;
1674 }
1675
1676 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1677 {
1678         return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1679 }
1680 #endif
1681
1682 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1683                            struct memcg_scanrecord *rec)
1684 {
1685
1686         stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1687         stats[SCAN_ANON] += rec->nr_scanned[0];
1688         stats[SCAN_FILE] += rec->nr_scanned[1];
1689
1690         stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1691         stats[ROTATE_ANON] += rec->nr_rotated[0];
1692         stats[ROTATE_FILE] += rec->nr_rotated[1];
1693
1694         stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1695         stats[FREED_ANON] += rec->nr_freed[0];
1696         stats[FREED_FILE] += rec->nr_freed[1];
1697
1698         stats[ELAPSED] += rec->elapsed;
1699 }
1700
1701 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1702 {
1703         struct mem_cgroup *mem;
1704         int context = rec->context;
1705
1706         if (context >= NR_SCAN_CONTEXT)
1707                 return;
1708
1709         mem = rec->mem;
1710         spin_lock(&mem->scanstat.lock);
1711         __mem_cgroup_record_scanstat(mem->scanstat.stats[context], rec);
1712         spin_unlock(&mem->scanstat.lock);
1713
1714         mem = rec->root;
1715         spin_lock(&mem->scanstat.lock);
1716         __mem_cgroup_record_scanstat(mem->scanstat.rootstats[context], rec);
1717         spin_unlock(&mem->scanstat.lock);
1718 }
1719
1720 /*
1721  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1722  * we reclaimed from, so that we don't end up penalizing one child extensively
1723  * based on its position in the children list.
1724  *
1725  * root_mem is the original ancestor that we've been reclaim from.
1726  *
1727  * We give up and return to the caller when we visit root_mem twice.
1728  * (other groups can be removed while we're walking....)
1729  *
1730  * If shrink==true, for avoiding to free too much, this returns immedieately.
1731  */
1732 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1733                                                 struct zone *zone,
1734                                                 gfp_t gfp_mask,
1735                                                 unsigned long reclaim_options,
1736                                                 unsigned long *total_scanned)
1737 {
1738         struct mem_cgroup *victim;
1739         int ret, total = 0;
1740         int loop = 0;
1741         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1742         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1743         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1744         struct memcg_scanrecord rec;
1745         unsigned long excess;
1746         unsigned long scanned;
1747
1748         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1749
1750         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1751         if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1752                 noswap = true;
1753
1754         if (shrink)
1755                 rec.context = SCAN_BY_SHRINK;
1756         else if (check_soft)
1757                 rec.context = SCAN_BY_SYSTEM;
1758         else
1759                 rec.context = SCAN_BY_LIMIT;
1760
1761         rec.root = root_mem;
1762
1763         while (1) {
1764                 victim = mem_cgroup_select_victim(root_mem);
1765                 if (victim == root_mem) {
1766                         loop++;
1767                         /*
1768                          * We are not draining per cpu cached charges during
1769                          * soft limit reclaim  because global reclaim doesn't
1770                          * care about charges. It tries to free some memory and
1771                          * charges will not give any.
1772                          */
1773                         if (!check_soft && loop >= 1)
1774                                 drain_all_stock_async(root_mem);
1775                         if (loop >= 2) {
1776                                 /*
1777                                  * If we have not been able to reclaim
1778                                  * anything, it might because there are
1779                                  * no reclaimable pages under this hierarchy
1780                                  */
1781                                 if (!check_soft || !total) {
1782                                         css_put(&victim->css);
1783                                         break;
1784                                 }
1785                                 /*
1786                                  * We want to do more targeted reclaim.
1787                                  * excess >> 2 is not to excessive so as to
1788                                  * reclaim too much, nor too less that we keep
1789                                  * coming back to reclaim from this cgroup
1790                                  */
1791                                 if (total >= (excess >> 2) ||
1792                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1793                                         css_put(&victim->css);
1794                                         break;
1795                                 }
1796                         }
1797                 }
1798                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1799                         /* this cgroup's local usage == 0 */
1800                         css_put(&victim->css);
1801                         continue;
1802                 }
1803                 rec.mem = victim;
1804                 rec.nr_scanned[0] = 0;
1805                 rec.nr_scanned[1] = 0;
1806                 rec.nr_rotated[0] = 0;
1807                 rec.nr_rotated[1] = 0;
1808                 rec.nr_freed[0] = 0;
1809                 rec.nr_freed[1] = 0;
1810                 rec.elapsed = 0;
1811                 /* we use swappiness of local cgroup */
1812                 if (check_soft) {
1813                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1814                                 noswap, zone, &rec, &scanned);
1815                         *total_scanned += scanned;
1816                 } else
1817                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1818                                                 noswap, &rec);
1819                 mem_cgroup_record_scanstat(&rec);
1820                 css_put(&victim->css);
1821                 /*
1822                  * At shrinking usage, we can't check we should stop here or
1823                  * reclaim more. It's depends on callers. last_scanned_child
1824                  * will work enough for keeping fairness under tree.
1825                  */
1826                 if (shrink)
1827                         return ret;
1828                 total += ret;
1829                 if (check_soft) {
1830                         if (!res_counter_soft_limit_excess(&root_mem->res))
1831                                 return total;
1832                 } else if (mem_cgroup_margin(root_mem))
1833                         return total;
1834         }
1835         return total;
1836 }
1837
1838 /*
1839  * Check OOM-Killer is already running under our hierarchy.
1840  * If someone is running, return false.
1841  * Has to be called with memcg_oom_lock
1842  */
1843 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1844 {
1845         int lock_count = -1;
1846         struct mem_cgroup *iter, *failed = NULL;
1847         bool cond = true;
1848
1849         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1850                 bool locked = iter->oom_lock;
1851
1852                 iter->oom_lock = true;
1853                 if (lock_count == -1)
1854                         lock_count = iter->oom_lock;
1855                 else if (lock_count != locked) {
1856                         /*
1857                          * this subtree of our hierarchy is already locked
1858                          * so we cannot give a lock.
1859                          */
1860                         lock_count = 0;
1861                         failed = iter;
1862                         cond = false;
1863                 }
1864         }
1865
1866         if (!failed)
1867                 goto done;
1868
1869         /*
1870          * OK, we failed to lock the whole subtree so we have to clean up
1871          * what we set up to the failing subtree
1872          */
1873         cond = true;
1874         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1875                 if (iter == failed) {
1876                         cond = false;
1877                         continue;
1878                 }
1879                 iter->oom_lock = false;
1880         }
1881 done:
1882         return lock_count;
1883 }
1884
1885 /*
1886  * Has to be called with memcg_oom_lock
1887  */
1888 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1889 {
1890         struct mem_cgroup *iter;
1891
1892         for_each_mem_cgroup_tree(iter, mem)
1893                 iter->oom_lock = false;
1894         return 0;
1895 }
1896
1897 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1898 {
1899         struct mem_cgroup *iter;
1900
1901         for_each_mem_cgroup_tree(iter, mem)
1902                 atomic_inc(&iter->under_oom);
1903 }
1904
1905 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1906 {
1907         struct mem_cgroup *iter;
1908
1909         /*
1910          * When a new child is created while the hierarchy is under oom,
1911          * mem_cgroup_oom_lock() may not be called. We have to use
1912          * atomic_add_unless() here.
1913          */
1914         for_each_mem_cgroup_tree(iter, mem)
1915                 atomic_add_unless(&iter->under_oom, -1, 0);
1916 }
1917
1918 static DEFINE_SPINLOCK(memcg_oom_lock);
1919 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1920
1921 struct oom_wait_info {
1922         struct mem_cgroup *mem;
1923         wait_queue_t    wait;
1924 };
1925
1926 static int memcg_oom_wake_function(wait_queue_t *wait,
1927         unsigned mode, int sync, void *arg)
1928 {
1929         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg,
1930                           *oom_wait_mem;
1931         struct oom_wait_info *oom_wait_info;
1932
1933         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1934         oom_wait_mem = oom_wait_info->mem;
1935
1936         /*
1937          * Both of oom_wait_info->mem and wake_mem are stable under us.
1938          * Then we can use css_is_ancestor without taking care of RCU.
1939          */
1940         if (!mem_cgroup_same_or_subtree(oom_wait_mem, wake_mem)
1941                         && !mem_cgroup_same_or_subtree(wake_mem, oom_wait_mem))
1942                 return 0;
1943         return autoremove_wake_function(wait, mode, sync, arg);
1944 }
1945
1946 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1947 {
1948         /* for filtering, pass "mem" as argument. */
1949         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1950 }
1951
1952 static void memcg_oom_recover(struct mem_cgroup *mem)
1953 {
1954         if (mem && atomic_read(&mem->under_oom))
1955                 memcg_wakeup_oom(mem);
1956 }
1957
1958 /*
1959  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1960  */
1961 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1962 {
1963         struct oom_wait_info owait;
1964         bool locked, need_to_kill;
1965
1966         owait.mem = mem;
1967         owait.wait.flags = 0;
1968         owait.wait.func = memcg_oom_wake_function;
1969         owait.wait.private = current;
1970         INIT_LIST_HEAD(&owait.wait.task_list);
1971         need_to_kill = true;
1972         mem_cgroup_mark_under_oom(mem);
1973
1974         /* At first, try to OOM lock hierarchy under mem.*/
1975         spin_lock(&memcg_oom_lock);
1976         locked = mem_cgroup_oom_lock(mem);
1977         /*
1978          * Even if signal_pending(), we can't quit charge() loop without
1979          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1980          * under OOM is always welcomed, use TASK_KILLABLE here.
1981          */
1982         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1983         if (!locked || mem->oom_kill_disable)
1984                 need_to_kill = false;
1985         if (locked)
1986                 mem_cgroup_oom_notify(mem);
1987         spin_unlock(&memcg_oom_lock);
1988
1989         if (need_to_kill) {
1990                 finish_wait(&memcg_oom_waitq, &owait.wait);
1991                 mem_cgroup_out_of_memory(mem, mask);
1992         } else {
1993                 schedule();
1994                 finish_wait(&memcg_oom_waitq, &owait.wait);
1995         }
1996         spin_lock(&memcg_oom_lock);
1997         if (locked)
1998                 mem_cgroup_oom_unlock(mem);
1999         memcg_wakeup_oom(mem);
2000         spin_unlock(&memcg_oom_lock);
2001
2002         mem_cgroup_unmark_under_oom(mem);
2003
2004         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2005                 return false;
2006         /* Give chance to dying process */
2007         schedule_timeout(1);
2008         return true;
2009 }
2010
2011 /*
2012  * Currently used to update mapped file statistics, but the routine can be
2013  * generalized to update other statistics as well.
2014  *
2015  * Notes: Race condition
2016  *
2017  * We usually use page_cgroup_lock() for accessing page_cgroup member but
2018  * it tends to be costly. But considering some conditions, we doesn't need
2019  * to do so _always_.
2020  *
2021  * Considering "charge", lock_page_cgroup() is not required because all
2022  * file-stat operations happen after a page is attached to radix-tree. There
2023  * are no race with "charge".
2024  *
2025  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2026  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2027  * if there are race with "uncharge". Statistics itself is properly handled
2028  * by flags.
2029  *
2030  * Considering "move", this is an only case we see a race. To make the race
2031  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2032  * possibility of race condition. If there is, we take a lock.
2033  */
2034
2035 void mem_cgroup_update_page_stat(struct page *page,
2036                                  enum mem_cgroup_page_stat_item idx, int val)
2037 {
2038         struct mem_cgroup *mem;
2039         struct page_cgroup *pc = lookup_page_cgroup(page);
2040         bool need_unlock = false;
2041         unsigned long uninitialized_var(flags);
2042
2043         if (unlikely(!pc))
2044                 return;
2045
2046         rcu_read_lock();
2047         mem = pc->mem_cgroup;
2048         if (unlikely(!mem || !PageCgroupUsed(pc)))
2049                 goto out;
2050         /* pc->mem_cgroup is unstable ? */
2051         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
2052                 /* take a lock against to access pc->mem_cgroup */
2053                 move_lock_page_cgroup(pc, &flags);
2054                 need_unlock = true;
2055                 mem = pc->mem_cgroup;
2056                 if (!mem || !PageCgroupUsed(pc))
2057                         goto out;
2058         }
2059
2060         switch (idx) {
2061         case MEMCG_NR_FILE_MAPPED:
2062                 if (val > 0)
2063                         SetPageCgroupFileMapped(pc);
2064                 else if (!page_mapped(page))
2065                         ClearPageCgroupFileMapped(pc);
2066                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2067                 break;
2068         default:
2069                 BUG();
2070         }
2071
2072         this_cpu_add(mem->stat->count[idx], val);
2073
2074 out:
2075         if (unlikely(need_unlock))
2076                 move_unlock_page_cgroup(pc, &flags);
2077         rcu_read_unlock();
2078         return;
2079 }
2080 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2081
2082 /*
2083  * size of first charge trial. "32" comes from vmscan.c's magic value.
2084  * TODO: maybe necessary to use big numbers in big irons.
2085  */
2086 #define CHARGE_BATCH    32U
2087 struct memcg_stock_pcp {
2088         struct mem_cgroup *cached; /* this never be root cgroup */
2089         unsigned int nr_pages;
2090         struct work_struct work;
2091         unsigned long flags;
2092 #define FLUSHING_CACHED_CHARGE  (0)
2093 };
2094 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2095
2096 /*
2097  * Try to consume stocked charge on this cpu. If success, one page is consumed
2098  * from local stock and true is returned. If the stock is 0 or charges from a
2099  * cgroup which is not current target, returns false. This stock will be
2100  * refilled.
2101  */
2102 static bool consume_stock(struct mem_cgroup *mem)
2103 {
2104         struct memcg_stock_pcp *stock;
2105         bool ret = true;
2106
2107         stock = &get_cpu_var(memcg_stock);
2108         if (mem == stock->cached && stock->nr_pages)
2109                 stock->nr_pages--;
2110         else /* need to call res_counter_charge */
2111                 ret = false;
2112         put_cpu_var(memcg_stock);
2113         return ret;
2114 }
2115
2116 /*
2117  * Returns stocks cached in percpu to res_counter and reset cached information.
2118  */
2119 static void drain_stock(struct memcg_stock_pcp *stock)
2120 {
2121         struct mem_cgroup *old = stock->cached;
2122
2123         if (stock->nr_pages) {
2124                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2125
2126                 res_counter_uncharge(&old->res, bytes);
2127                 if (do_swap_account)
2128                         res_counter_uncharge(&old->memsw, bytes);
2129                 stock->nr_pages = 0;
2130         }
2131         stock->cached = NULL;
2132 }
2133
2134 /*
2135  * This must be called under preempt disabled or must be called by
2136  * a thread which is pinned to local cpu.
2137  */
2138 static void drain_local_stock(struct work_struct *dummy)
2139 {
2140         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2141         drain_stock(stock);
2142         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2143 }
2144
2145 /*
2146  * Cache charges(val) which is from res_counter, to local per_cpu area.
2147  * This will be consumed by consume_stock() function, later.
2148  */
2149 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2150 {
2151         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2152
2153         if (stock->cached != mem) { /* reset if necessary */
2154                 drain_stock(stock);
2155                 stock->cached = mem;
2156         }
2157         stock->nr_pages += nr_pages;
2158         put_cpu_var(memcg_stock);
2159 }
2160
2161 /*
2162  * Drains all per-CPU charge caches for given root_mem resp. subtree
2163  * of the hierarchy under it. sync flag says whether we should block
2164  * until the work is done.
2165  */
2166 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2167 {
2168         int cpu, curcpu;
2169
2170         /* Notify other cpus that system-wide "drain" is running */
2171         get_online_cpus();
2172         /*
2173          * Get a hint for avoiding draining charges on the current cpu,
2174          * which must be exhausted by our charging.  It is not required that
2175          * this be a precise check, so we use raw_smp_processor_id() instead of
2176          * getcpu()/putcpu().
2177          */
2178         curcpu = raw_smp_processor_id();
2179         for_each_online_cpu(cpu) {
2180                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2181                 struct mem_cgroup *mem;
2182
2183                 mem = stock->cached;
2184                 if (!mem || !stock->nr_pages)
2185                         continue;
2186                 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2187                         continue;
2188                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2189                         if (cpu == curcpu)
2190                                 drain_local_stock(&stock->work);
2191                         else
2192                                 schedule_work_on(cpu, &stock->work);
2193                 }
2194         }
2195
2196         if (!sync)
2197                 goto out;
2198
2199         for_each_online_cpu(cpu) {
2200                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2201                 if (mem_cgroup_same_or_subtree(root_mem, stock->cached) &&
2202                                 test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2203                         flush_work(&stock->work);
2204         }
2205 out:
2206         put_online_cpus();
2207 }
2208
2209 /*
2210  * Tries to drain stocked charges in other cpus. This function is asynchronous
2211  * and just put a work per cpu for draining localy on each cpu. Caller can
2212  * expects some charges will be back to res_counter later but cannot wait for
2213  * it.
2214  */
2215 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2216 {
2217         drain_all_stock(root_mem, false);
2218 }
2219
2220 /* This is a synchronous drain interface. */
2221 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2222 {
2223         /* called when force_empty is called */
2224         drain_all_stock(root_mem, true);
2225 }
2226
2227 /*
2228  * This function drains percpu counter value from DEAD cpu and
2229  * move it to local cpu. Note that this function can be preempted.
2230  */
2231 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2232 {
2233         int i;
2234
2235         spin_lock(&mem->pcp_counter_lock);
2236         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2237                 long x = per_cpu(mem->stat->count[i], cpu);
2238
2239                 per_cpu(mem->stat->count[i], cpu) = 0;
2240                 mem->nocpu_base.count[i] += x;
2241         }
2242         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2243                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2244
2245                 per_cpu(mem->stat->events[i], cpu) = 0;
2246                 mem->nocpu_base.events[i] += x;
2247         }
2248         /* need to clear ON_MOVE value, works as a kind of lock. */
2249         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2250         spin_unlock(&mem->pcp_counter_lock);
2251 }
2252
2253 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2254 {
2255         int idx = MEM_CGROUP_ON_MOVE;
2256
2257         spin_lock(&mem->pcp_counter_lock);
2258         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2259         spin_unlock(&mem->pcp_counter_lock);
2260 }
2261
2262 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2263                                         unsigned long action,
2264                                         void *hcpu)
2265 {
2266         int cpu = (unsigned long)hcpu;
2267         struct memcg_stock_pcp *stock;
2268         struct mem_cgroup *iter;
2269
2270         if ((action == CPU_ONLINE)) {
2271                 for_each_mem_cgroup_all(iter)
2272                         synchronize_mem_cgroup_on_move(iter, cpu);
2273                 return NOTIFY_OK;
2274         }
2275
2276         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2277                 return NOTIFY_OK;
2278
2279         for_each_mem_cgroup_all(iter)
2280                 mem_cgroup_drain_pcp_counter(iter, cpu);
2281
2282         stock = &per_cpu(memcg_stock, cpu);
2283         drain_stock(stock);
2284         return NOTIFY_OK;
2285 }
2286
2287
2288 /* See __mem_cgroup_try_charge() for details */
2289 enum {
2290         CHARGE_OK,              /* success */
2291         CHARGE_RETRY,           /* need to retry but retry is not bad */
2292         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2293         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2294         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2295 };
2296
2297 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2298                                 unsigned int nr_pages, bool oom_check)
2299 {
2300         unsigned long csize = nr_pages * PAGE_SIZE;
2301         struct mem_cgroup *mem_over_limit;
2302         struct res_counter *fail_res;
2303         unsigned long flags = 0;
2304         int ret;
2305
2306         ret = res_counter_charge(&mem->res, csize, &fail_res);
2307
2308         if (likely(!ret)) {
2309                 if (!do_swap_account)
2310                         return CHARGE_OK;
2311                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2312                 if (likely(!ret))
2313                         return CHARGE_OK;
2314
2315                 res_counter_uncharge(&mem->res, csize);
2316                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2317                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2318         } else
2319                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2320         /*
2321          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2322          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2323          *
2324          * Never reclaim on behalf of optional batching, retry with a
2325          * single page instead.
2326          */
2327         if (nr_pages == CHARGE_BATCH)
2328                 return CHARGE_RETRY;
2329
2330         if (!(gfp_mask & __GFP_WAIT))
2331                 return CHARGE_WOULDBLOCK;
2332
2333         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2334                                               gfp_mask, flags, NULL);
2335         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2336                 return CHARGE_RETRY;
2337         /*
2338          * Even though the limit is exceeded at this point, reclaim
2339          * may have been able to free some pages.  Retry the charge
2340          * before killing the task.
2341          *
2342          * Only for regular pages, though: huge pages are rather
2343          * unlikely to succeed so close to the limit, and we fall back
2344          * to regular pages anyway in case of failure.
2345          */
2346         if (nr_pages == 1 && ret)
2347                 return CHARGE_RETRY;
2348
2349         /*
2350          * At task move, charge accounts can be doubly counted. So, it's
2351          * better to wait until the end of task_move if something is going on.
2352          */
2353         if (mem_cgroup_wait_acct_move(mem_over_limit))
2354                 return CHARGE_RETRY;
2355
2356         /* If we don't need to call oom-killer at el, return immediately */
2357         if (!oom_check)
2358                 return CHARGE_NOMEM;
2359         /* check OOM */
2360         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2361                 return CHARGE_OOM_DIE;
2362
2363         return CHARGE_RETRY;
2364 }
2365
2366 /*
2367  * Unlike exported interface, "oom" parameter is added. if oom==true,
2368  * oom-killer can be invoked.
2369  */
2370 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2371                                    gfp_t gfp_mask,
2372                                    unsigned int nr_pages,
2373                                    struct mem_cgroup **memcg,
2374                                    bool oom)
2375 {
2376         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2377         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2378         struct mem_cgroup *mem = NULL;
2379         int ret;
2380
2381         /*
2382          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2383          * in system level. So, allow to go ahead dying process in addition to
2384          * MEMDIE process.
2385          */
2386         if (unlikely(test_thread_flag(TIF_MEMDIE)
2387                      || fatal_signal_pending(current)))
2388                 goto bypass;
2389
2390         /*
2391          * We always charge the cgroup the mm_struct belongs to.
2392          * The mm_struct's mem_cgroup changes on task migration if the
2393          * thread group leader migrates. It's possible that mm is not
2394          * set, if so charge the init_mm (happens for pagecache usage).
2395          */
2396         if (!*memcg && !mm)
2397                 goto bypass;
2398 again:
2399         if (*memcg) { /* css should be a valid one */
2400                 mem = *memcg;
2401                 VM_BUG_ON(css_is_removed(&mem->css));
2402                 if (mem_cgroup_is_root(mem))
2403                         goto done;
2404                 if (nr_pages == 1 && consume_stock(mem))
2405                         goto done;
2406                 css_get(&mem->css);
2407         } else {
2408                 struct task_struct *p;
2409
2410                 rcu_read_lock();
2411                 p = rcu_dereference(mm->owner);
2412                 /*
2413                  * Because we don't have task_lock(), "p" can exit.
2414                  * In that case, "mem" can point to root or p can be NULL with
2415                  * race with swapoff. Then, we have small risk of mis-accouning.
2416                  * But such kind of mis-account by race always happens because
2417                  * we don't have cgroup_mutex(). It's overkill and we allo that
2418                  * small race, here.
2419                  * (*) swapoff at el will charge against mm-struct not against
2420                  * task-struct. So, mm->owner can be NULL.
2421                  */
2422                 mem = mem_cgroup_from_task(p);
2423                 if (!mem || mem_cgroup_is_root(mem)) {
2424                         rcu_read_unlock();
2425                         goto done;
2426                 }
2427                 if (nr_pages == 1 && consume_stock(mem)) {
2428                         /*
2429                          * It seems dagerous to access memcg without css_get().
2430                          * But considering how consume_stok works, it's not
2431                          * necessary. If consume_stock success, some charges
2432                          * from this memcg are cached on this cpu. So, we
2433                          * don't need to call css_get()/css_tryget() before
2434                          * calling consume_stock().
2435                          */
2436                         rcu_read_unlock();
2437                         goto done;
2438                 }
2439                 /* after here, we may be blocked. we need to get refcnt */
2440                 if (!css_tryget(&mem->css)) {
2441                         rcu_read_unlock();
2442                         goto again;
2443                 }
2444                 rcu_read_unlock();
2445         }
2446
2447         do {
2448                 bool oom_check;
2449
2450                 /* If killed, bypass charge */
2451                 if (fatal_signal_pending(current)) {
2452                         css_put(&mem->css);
2453                         goto bypass;
2454                 }
2455
2456                 oom_check = false;
2457                 if (oom && !nr_oom_retries) {
2458                         oom_check = true;
2459                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2460                 }
2461
2462                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2463                 switch (ret) {
2464                 case CHARGE_OK:
2465                         break;
2466                 case CHARGE_RETRY: /* not in OOM situation but retry */
2467                         batch = nr_pages;
2468                         css_put(&mem->css);
2469                         mem = NULL;
2470                         goto again;
2471                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2472                         css_put(&mem->css);
2473                         goto nomem;
2474                 case CHARGE_NOMEM: /* OOM routine works */
2475                         if (!oom) {
2476                                 css_put(&mem->css);
2477                                 goto nomem;
2478                         }
2479                         /* If oom, we never return -ENOMEM */
2480                         nr_oom_retries--;
2481                         break;
2482                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2483                         css_put(&mem->css);
2484                         goto bypass;
2485                 }
2486         } while (ret != CHARGE_OK);
2487
2488         if (batch > nr_pages)
2489                 refill_stock(mem, batch - nr_pages);
2490         css_put(&mem->css);
2491 done:
2492         *memcg = mem;
2493         return 0;
2494 nomem:
2495         *memcg = NULL;
2496         return -ENOMEM;
2497 bypass:
2498         *memcg = NULL;
2499         return 0;
2500 }
2501
2502 /*
2503  * Somemtimes we have to undo a charge we got by try_charge().
2504  * This function is for that and do uncharge, put css's refcnt.
2505  * gotten by try_charge().
2506  */
2507 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2508                                        unsigned int nr_pages)
2509 {
2510         if (!mem_cgroup_is_root(mem)) {
2511                 unsigned long bytes = nr_pages * PAGE_SIZE;
2512
2513                 res_counter_uncharge(&mem->res, bytes);
2514                 if (do_swap_account)
2515                         res_counter_uncharge(&mem->memsw, bytes);
2516         }
2517 }
2518
2519 /*
2520  * A helper function to get mem_cgroup from ID. must be called under
2521  * rcu_read_lock(). The caller must check css_is_removed() or some if
2522  * it's concern. (dropping refcnt from swap can be called against removed
2523  * memcg.)
2524  */
2525 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2526 {
2527         struct cgroup_subsys_state *css;
2528
2529         /* ID 0 is unused ID */
2530         if (!id)
2531                 return NULL;
2532         css = css_lookup(&mem_cgroup_subsys, id);
2533         if (!css)
2534                 return NULL;
2535         return container_of(css, struct mem_cgroup, css);
2536 }
2537
2538 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2539 {
2540         struct mem_cgroup *mem = NULL;
2541         struct page_cgroup *pc;
2542         unsigned short id;
2543         swp_entry_t ent;
2544
2545         VM_BUG_ON(!PageLocked(page));
2546
2547         pc = lookup_page_cgroup(page);
2548         lock_page_cgroup(pc);
2549         if (PageCgroupUsed(pc)) {
2550                 mem = pc->mem_cgroup;
2551                 if (mem && !css_tryget(&mem->css))
2552                         mem = NULL;
2553         } else if (PageSwapCache(page)) {
2554                 ent.val = page_private(page);
2555                 id = lookup_swap_cgroup(ent);
2556                 rcu_read_lock();
2557                 mem = mem_cgroup_lookup(id);
2558                 if (mem && !css_tryget(&mem->css))
2559                         mem = NULL;
2560                 rcu_read_unlock();
2561         }
2562         unlock_page_cgroup(pc);
2563         return mem;
2564 }
2565
2566 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2567                                        struct page *page,
2568                                        unsigned int nr_pages,
2569                                        struct page_cgroup *pc,
2570                                        enum charge_type ctype)
2571 {
2572         lock_page_cgroup(pc);
2573         if (unlikely(PageCgroupUsed(pc))) {
2574                 unlock_page_cgroup(pc);
2575                 __mem_cgroup_cancel_charge(mem, nr_pages);
2576                 return;
2577         }
2578         /*
2579          * we don't need page_cgroup_lock about tail pages, becase they are not
2580          * accessed by any other context at this point.
2581          */
2582         pc->mem_cgroup = mem;
2583         /*
2584          * We access a page_cgroup asynchronously without lock_page_cgroup().
2585          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2586          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2587          * before USED bit, we need memory barrier here.
2588          * See mem_cgroup_add_lru_list(), etc.
2589          */
2590         smp_wmb();
2591         switch (ctype) {
2592         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2593         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2594                 SetPageCgroupCache(pc);
2595                 SetPageCgroupUsed(pc);
2596                 break;
2597         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2598                 ClearPageCgroupCache(pc);
2599                 SetPageCgroupUsed(pc);
2600                 break;
2601         default:
2602                 break;
2603         }
2604
2605         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2606         unlock_page_cgroup(pc);
2607         /*
2608          * "charge_statistics" updated event counter. Then, check it.
2609          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2610          * if they exceeds softlimit.
2611          */
2612         memcg_check_events(mem, page);
2613 }
2614
2615 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2616
2617 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2618                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2619 /*
2620  * Because tail pages are not marked as "used", set it. We're under
2621  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2622  */
2623 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2624 {
2625         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2626         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2627         unsigned long flags;
2628
2629         if (mem_cgroup_disabled())
2630                 return;
2631         /*
2632          * We have no races with charge/uncharge but will have races with
2633          * page state accounting.
2634          */
2635         move_lock_page_cgroup(head_pc, &flags);
2636
2637         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2638         smp_wmb(); /* see __commit_charge() */
2639         if (PageCgroupAcctLRU(head_pc)) {
2640                 enum lru_list lru;
2641                 struct mem_cgroup_per_zone *mz;
2642
2643                 /*
2644                  * LRU flags cannot be copied because we need to add tail
2645                  *.page to LRU by generic call and our hook will be called.
2646                  * We hold lru_lock, then, reduce counter directly.
2647                  */
2648                 lru = page_lru(head);
2649                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2650                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2651         }
2652         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2653         move_unlock_page_cgroup(head_pc, &flags);
2654 }
2655 #endif
2656
2657 /**
2658  * mem_cgroup_move_account - move account of the page
2659  * @page: the page
2660  * @nr_pages: number of regular pages (>1 for huge pages)
2661  * @pc: page_cgroup of the page.
2662  * @from: mem_cgroup which the page is moved from.
2663  * @to: mem_cgroup which the page is moved to. @from != @to.
2664  * @uncharge: whether we should call uncharge and css_put against @from.
2665  *
2666  * The caller must confirm following.
2667  * - page is not on LRU (isolate_page() is useful.)
2668  * - compound_lock is held when nr_pages > 1
2669  *
2670  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2671  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2672  * true, this function does "uncharge" from old cgroup, but it doesn't if
2673  * @uncharge is false, so a caller should do "uncharge".
2674  */
2675 static int mem_cgroup_move_account(struct page *page,
2676                                    unsigned int nr_pages,
2677                                    struct page_cgroup *pc,
2678                                    struct mem_cgroup *from,
2679                                    struct mem_cgroup *to,
2680                                    bool uncharge)
2681 {
2682         unsigned long flags;
2683         int ret;
2684
2685         VM_BUG_ON(from == to);
2686         VM_BUG_ON(PageLRU(page));
2687         /*
2688          * The page is isolated from LRU. So, collapse function
2689          * will not handle this page. But page splitting can happen.
2690          * Do this check under compound_page_lock(). The caller should
2691          * hold it.
2692          */
2693         ret = -EBUSY;
2694         if (nr_pages > 1 && !PageTransHuge(page))
2695                 goto out;
2696
2697         lock_page_cgroup(pc);
2698
2699         ret = -EINVAL;
2700         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2701                 goto unlock;
2702
2703         move_lock_page_cgroup(pc, &flags);
2704
2705         if (PageCgroupFileMapped(pc)) {
2706                 /* Update mapped_file data for mem_cgroup */
2707                 preempt_disable();
2708                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2709                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2710                 preempt_enable();
2711         }
2712         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2713         if (uncharge)
2714                 /* This is not "cancel", but cancel_charge does all we need. */
2715                 __mem_cgroup_cancel_charge(from, nr_pages);
2716
2717         /* caller should have done css_get */
2718         pc->mem_cgroup = to;
2719         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2720         /*
2721          * We charges against "to" which may not have any tasks. Then, "to"
2722          * can be under rmdir(). But in current implementation, caller of
2723          * this function is just force_empty() and move charge, so it's
2724          * guaranteed that "to" is never removed. So, we don't check rmdir
2725          * status here.
2726          */
2727         move_unlock_page_cgroup(pc, &flags);
2728         ret = 0;
2729 unlock:
2730         unlock_page_cgroup(pc);
2731         /*
2732          * check events
2733          */
2734         memcg_check_events(to, page);
2735         memcg_check_events(from, page);
2736 out:
2737         return ret;
2738 }
2739
2740 /*
2741  * move charges to its parent.
2742  */
2743
2744 static int mem_cgroup_move_parent(struct page *page,
2745                                   struct page_cgroup *pc,
2746                                   struct mem_cgroup *child,
2747                                   gfp_t gfp_mask)
2748 {
2749         struct cgroup *cg = child->css.cgroup;
2750         struct cgroup *pcg = cg->parent;
2751         struct mem_cgroup *parent;
2752         unsigned int nr_pages;
2753         unsigned long uninitialized_var(flags);
2754         int ret;
2755
2756         /* Is ROOT ? */
2757         if (!pcg)
2758                 return -EINVAL;
2759
2760         ret = -EBUSY;
2761         if (!get_page_unless_zero(page))
2762                 goto out;
2763         if (isolate_lru_page(page))
2764                 goto put;
2765
2766         nr_pages = hpage_nr_pages(page);
2767
2768         parent = mem_cgroup_from_cont(pcg);
2769         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2770         if (ret || !parent)
2771                 goto put_back;
2772
2773         if (nr_pages > 1)
2774                 flags = compound_lock_irqsave(page);
2775
2776         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2777         if (ret)
2778                 __mem_cgroup_cancel_charge(parent, nr_pages);
2779
2780         if (nr_pages > 1)
2781                 compound_unlock_irqrestore(page, flags);
2782 put_back:
2783         putback_lru_page(page);
2784 put:
2785         put_page(page);
2786 out:
2787         return ret;
2788 }
2789
2790 /*
2791  * Charge the memory controller for page usage.
2792  * Return
2793  * 0 if the charge was successful
2794  * < 0 if the cgroup is over its limit
2795  */
2796 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2797                                 gfp_t gfp_mask, enum charge_type ctype)
2798 {
2799         struct mem_cgroup *mem = NULL;
2800         unsigned int nr_pages = 1;
2801         struct page_cgroup *pc;
2802         bool oom = true;
2803         int ret;
2804
2805         if (PageTransHuge(page)) {
2806                 nr_pages <<= compound_order(page);
2807                 VM_BUG_ON(!PageTransHuge(page));
2808                 /*
2809                  * Never OOM-kill a process for a huge page.  The
2810                  * fault handler will fall back to regular pages.
2811                  */
2812                 oom = false;
2813         }
2814
2815         pc = lookup_page_cgroup(page);
2816         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2817
2818         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2819         if (ret || !mem)
2820                 return ret;
2821
2822         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2823         return 0;
2824 }
2825
2826 int mem_cgroup_newpage_charge(struct page *page,
2827                               struct mm_struct *mm, gfp_t gfp_mask)
2828 {
2829         if (mem_cgroup_disabled())
2830                 return 0;
2831         /*
2832          * If already mapped, we don't have to account.
2833          * If page cache, page->mapping has address_space.
2834          * But page->mapping may have out-of-use anon_vma pointer,
2835          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2836          * is NULL.
2837          */
2838         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2839                 return 0;
2840         if (unlikely(!mm))
2841                 mm = &init_mm;
2842         return mem_cgroup_charge_common(page, mm, gfp_mask,
2843                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2844 }
2845
2846 static void
2847 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2848                                         enum charge_type ctype);
2849
2850 static void
2851 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2852                                         enum charge_type ctype)
2853 {
2854         struct page_cgroup *pc = lookup_page_cgroup(page);
2855         /*
2856          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2857          * is already on LRU. It means the page may on some other page_cgroup's
2858          * LRU. Take care of it.
2859          */
2860         mem_cgroup_lru_del_before_commit(page);
2861         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2862         mem_cgroup_lru_add_after_commit(page);
2863         return;
2864 }
2865
2866 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2867                                 gfp_t gfp_mask)
2868 {
2869         struct mem_cgroup *mem = NULL;
2870         int ret;
2871
2872         if (mem_cgroup_disabled())
2873                 return 0;
2874         if (PageCompound(page))
2875                 return 0;
2876         /*
2877          * Corner case handling. This is called from add_to_page_cache()
2878          * in usual. But some FS (shmem) precharges this page before calling it
2879          * and call add_to_page_cache() with GFP_NOWAIT.
2880          *
2881          * For GFP_NOWAIT case, the page may be pre-charged before calling
2882          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2883          * charge twice. (It works but has to pay a bit larger cost.)
2884          * And when the page is SwapCache, it should take swap information
2885          * into account. This is under lock_page() now.
2886          */
2887         if (!(gfp_mask & __GFP_WAIT)) {
2888                 struct page_cgroup *pc;
2889
2890                 pc = lookup_page_cgroup(page);
2891                 if (!pc)
2892                         return 0;
2893                 lock_page_cgroup(pc);
2894                 if (PageCgroupUsed(pc)) {
2895                         unlock_page_cgroup(pc);
2896                         return 0;
2897                 }
2898                 unlock_page_cgroup(pc);
2899         }
2900
2901         if (unlikely(!mm))
2902                 mm = &init_mm;
2903
2904         if (page_is_file_cache(page)) {
2905                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2906                 if (ret || !mem)
2907                         return ret;
2908
2909                 /*
2910                  * FUSE reuses pages without going through the final
2911                  * put that would remove them from the LRU list, make
2912                  * sure that they get relinked properly.
2913                  */
2914                 __mem_cgroup_commit_charge_lrucare(page, mem,
2915                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2916                 return ret;
2917         }
2918         /* shmem */
2919         if (PageSwapCache(page)) {
2920                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2921                 if (!ret)
2922                         __mem_cgroup_commit_charge_swapin(page, mem,
2923                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2924         } else
2925                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2926                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2927
2928         return ret;
2929 }
2930
2931 /*
2932  * While swap-in, try_charge -> commit or cancel, the page is locked.
2933  * And when try_charge() successfully returns, one refcnt to memcg without
2934  * struct page_cgroup is acquired. This refcnt will be consumed by
2935  * "commit()" or removed by "cancel()"
2936  */
2937 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2938                                  struct page *page,
2939                                  gfp_t mask, struct mem_cgroup **ptr)
2940 {
2941         struct mem_cgroup *mem;
2942         int ret;
2943
2944         *ptr = NULL;
2945
2946         if (mem_cgroup_disabled())
2947                 return 0;
2948
2949         if (!do_swap_account)
2950                 goto charge_cur_mm;
2951         /*
2952          * A racing thread's fault, or swapoff, may have already updated
2953          * the pte, and even removed page from swap cache: in those cases
2954          * do_swap_page()'s pte_same() test will fail; but there's also a
2955          * KSM case which does need to charge the page.
2956          */
2957         if (!PageSwapCache(page))
2958                 goto charge_cur_mm;
2959         mem = try_get_mem_cgroup_from_page(page);
2960         if (!mem)
2961                 goto charge_cur_mm;
2962         *ptr = mem;
2963         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2964         css_put(&mem->css);
2965         return ret;
2966 charge_cur_mm:
2967         if (unlikely(!mm))
2968                 mm = &init_mm;
2969         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2970 }
2971
2972 static void
2973 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2974                                         enum charge_type ctype)
2975 {
2976         if (mem_cgroup_disabled())
2977                 return;
2978         if (!ptr)
2979                 return;
2980         cgroup_exclude_rmdir(&ptr->css);
2981
2982         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2983         /*
2984          * Now swap is on-memory. This means this page may be
2985          * counted both as mem and swap....double count.
2986          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2987          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2988          * may call delete_from_swap_cache() before reach here.
2989          */
2990         if (do_swap_account && PageSwapCache(page)) {
2991                 swp_entry_t ent = {.val = page_private(page)};
2992                 unsigned short id;
2993                 struct mem_cgroup *memcg;
2994
2995                 id = swap_cgroup_record(ent, 0);
2996                 rcu_read_lock();
2997                 memcg = mem_cgroup_lookup(id);
2998                 if (memcg) {
2999                         /*
3000                          * This recorded memcg can be obsolete one. So, avoid
3001                          * calling css_tryget
3002                          */
3003                         if (!mem_cgroup_is_root(memcg))
3004                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3005                         mem_cgroup_swap_statistics(memcg, false);
3006                         mem_cgroup_put(memcg);
3007                 }
3008                 rcu_read_unlock();
3009         }
3010         /*
3011          * At swapin, we may charge account against cgroup which has no tasks.
3012          * So, rmdir()->pre_destroy() can be called while we do this charge.
3013          * In that case, we need to call pre_destroy() again. check it here.
3014          */
3015         cgroup_release_and_wakeup_rmdir(&ptr->css);
3016 }
3017
3018 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
3019 {
3020         __mem_cgroup_commit_charge_swapin(page, ptr,
3021                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
3022 }
3023
3024 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3025 {
3026         if (mem_cgroup_disabled())
3027                 return;
3028         if (!mem)
3029                 return;
3030         __mem_cgroup_cancel_charge(mem, 1);
3031 }
3032
3033 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3034                                    unsigned int nr_pages,
3035                                    const enum charge_type ctype)
3036 {
3037         struct memcg_batch_info *batch = NULL;
3038         bool uncharge_memsw = true;
3039
3040         /* If swapout, usage of swap doesn't decrease */
3041         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3042                 uncharge_memsw = false;
3043
3044         batch = &current->memcg_batch;
3045         /*
3046          * In usual, we do css_get() when we remember memcg pointer.
3047          * But in this case, we keep res->usage until end of a series of
3048          * uncharges. Then, it's ok to ignore memcg's refcnt.
3049          */
3050         if (!batch->memcg)
3051                 batch->memcg = mem;
3052         /*
3053          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3054          * In those cases, all pages freed continuously can be expected to be in
3055          * the same cgroup and we have chance to coalesce uncharges.
3056          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3057          * because we want to do uncharge as soon as possible.
3058          */
3059
3060         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3061                 goto direct_uncharge;
3062
3063         if (nr_pages > 1)
3064                 goto direct_uncharge;
3065
3066         /*
3067          * In typical case, batch->memcg == mem. This means we can
3068          * merge a series of uncharges to an uncharge of res_counter.
3069          * If not, we uncharge res_counter ony by one.
3070          */
3071         if (batch->memcg != mem)
3072                 goto direct_uncharge;
3073         /* remember freed charge and uncharge it later */
3074         batch->nr_pages++;
3075         if (uncharge_memsw)
3076                 batch->memsw_nr_pages++;
3077         return;
3078 direct_uncharge:
3079         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3080         if (uncharge_memsw)
3081                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3082         if (unlikely(batch->memcg != mem))
3083                 memcg_oom_recover(mem);
3084         return;
3085 }
3086
3087 /*
3088  * uncharge if !page_mapped(page)
3089  */
3090 static struct mem_cgroup *
3091 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3092 {
3093         struct mem_cgroup *mem = NULL;
3094         unsigned int nr_pages = 1;
3095         struct page_cgroup *pc;
3096
3097         if (mem_cgroup_disabled())
3098                 return NULL;
3099
3100         if (PageSwapCache(page))
3101                 return NULL;
3102
3103         if (PageTransHuge(page)) {
3104                 nr_pages <<= compound_order(page);
3105                 VM_BUG_ON(!PageTransHuge(page));
3106         }
3107         /*
3108          * Check if our page_cgroup is valid
3109          */
3110         pc = lookup_page_cgroup(page);
3111         if (unlikely(!pc || !PageCgroupUsed(pc)))
3112                 return NULL;
3113
3114         lock_page_cgroup(pc);
3115
3116         mem = pc->mem_cgroup;
3117
3118         if (!PageCgroupUsed(pc))
3119                 goto unlock_out;
3120
3121         switch (ctype) {
3122         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3123         case MEM_CGROUP_CHARGE_TYPE_DROP:
3124                 /* See mem_cgroup_prepare_migration() */
3125                 if (page_mapped(page) || PageCgroupMigration(pc))
3126                         goto unlock_out;
3127                 break;
3128         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3129                 if (!PageAnon(page)) {  /* Shared memory */
3130                         if (page->mapping && !page_is_file_cache(page))
3131                                 goto unlock_out;
3132                 } else if (page_mapped(page)) /* Anon */
3133                                 goto unlock_out;
3134                 break;
3135         default:
3136                 break;
3137         }
3138
3139         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3140
3141         ClearPageCgroupUsed(pc);
3142         /*
3143          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3144          * freed from LRU. This is safe because uncharged page is expected not
3145          * to be reused (freed soon). Exception is SwapCache, it's handled by
3146          * special functions.
3147          */
3148
3149         unlock_page_cgroup(pc);
3150         /*
3151          * even after unlock, we have mem->res.usage here and this memcg
3152          * will never be freed.
3153          */
3154         memcg_check_events(mem, page);
3155         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3156                 mem_cgroup_swap_statistics(mem, true);
3157                 mem_cgroup_get(mem);
3158         }
3159         if (!mem_cgroup_is_root(mem))
3160                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3161
3162         return mem;
3163
3164 unlock_out:
3165         unlock_page_cgroup(pc);
3166         return NULL;
3167 }
3168
3169 void mem_cgroup_uncharge_page(struct page *page)
3170 {
3171         /* early check. */
3172         if (page_mapped(page))
3173                 return;
3174         if (page->mapping && !PageAnon(page))
3175                 return;
3176         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3177 }
3178
3179 void mem_cgroup_uncharge_cache_page(struct page *page)
3180 {
3181         VM_BUG_ON(page_mapped(page));
3182         VM_BUG_ON(page->mapping);
3183         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3184 }
3185
3186 /*
3187  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3188  * In that cases, pages are freed continuously and we can expect pages
3189  * are in the same memcg. All these calls itself limits the number of
3190  * pages freed at once, then uncharge_start/end() is called properly.
3191  * This may be called prural(2) times in a context,
3192  */
3193
3194 void mem_cgroup_uncharge_start(void)
3195 {
3196         current->memcg_batch.do_batch++;
3197         /* We can do nest. */
3198         if (current->memcg_batch.do_batch == 1) {
3199                 current->memcg_batch.memcg = NULL;
3200                 current->memcg_batch.nr_pages = 0;
3201                 current->memcg_batch.memsw_nr_pages = 0;
3202         }
3203 }
3204
3205 void mem_cgroup_uncharge_end(void)
3206 {
3207         struct memcg_batch_info *batch = &current->memcg_batch;
3208
3209         if (!batch->do_batch)
3210                 return;
3211
3212         batch->do_batch--;
3213         if (batch->do_batch) /* If stacked, do nothing. */
3214                 return;
3215
3216         if (!batch->memcg)
3217                 return;
3218         /*
3219          * This "batch->memcg" is valid without any css_get/put etc...
3220          * bacause we hide charges behind us.
3221          */
3222         if (batch->nr_pages)
3223                 res_counter_uncharge(&batch->memcg->res,
3224                                      batch->nr_pages * PAGE_SIZE);
3225         if (batch->memsw_nr_pages)
3226                 res_counter_uncharge(&batch->memcg->memsw,
3227                                      batch->memsw_nr_pages * PAGE_SIZE);
3228         memcg_oom_recover(batch->memcg);
3229         /* forget this pointer (for sanity check) */
3230         batch->memcg = NULL;
3231 }
3232
3233 #ifdef CONFIG_SWAP
3234 /*
3235  * called after __delete_from_swap_cache() and drop "page" account.
3236  * memcg information is recorded to swap_cgroup of "ent"
3237  */
3238 void
3239 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3240 {
3241         struct mem_cgroup *memcg;
3242         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3243
3244         if (!swapout) /* this was a swap cache but the swap is unused ! */
3245                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3246
3247         memcg = __mem_cgroup_uncharge_common(page, ctype);
3248
3249         /*
3250          * record memcg information,  if swapout && memcg != NULL,
3251          * mem_cgroup_get() was called in uncharge().
3252          */
3253         if (do_swap_account && swapout && memcg)
3254                 swap_cgroup_record(ent, css_id(&memcg->css));
3255 }
3256 #endif
3257
3258 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3259 /*
3260  * called from swap_entry_free(). remove record in swap_cgroup and
3261  * uncharge "memsw" account.
3262  */
3263 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3264 {
3265         struct mem_cgroup *memcg;
3266         unsigned short id;
3267
3268         if (!do_swap_account)
3269                 return;
3270
3271         id = swap_cgroup_record(ent, 0);
3272         rcu_read_lock();
3273         memcg = mem_cgroup_lookup(id);
3274         if (memcg) {
3275                 /*
3276                  * We uncharge this because swap is freed.
3277                  * This memcg can be obsolete one. We avoid calling css_tryget
3278                  */
3279                 if (!mem_cgroup_is_root(memcg))
3280                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3281                 mem_cgroup_swap_statistics(memcg, false);
3282                 mem_cgroup_put(memcg);
3283         }
3284         rcu_read_unlock();
3285 }
3286
3287 /**
3288  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3289  * @entry: swap entry to be moved
3290  * @from:  mem_cgroup which the entry is moved from
3291  * @to:  mem_cgroup which the entry is moved to
3292  * @need_fixup: whether we should fixup res_counters and refcounts.
3293  *
3294  * It succeeds only when the swap_cgroup's record for this entry is the same
3295  * as the mem_cgroup's id of @from.
3296  *
3297  * Returns 0 on success, -EINVAL on failure.
3298  *
3299  * The caller must have charged to @to, IOW, called res_counter_charge() about
3300  * both res and memsw, and called css_get().
3301  */
3302 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3303                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3304 {
3305         unsigned short old_id, new_id;
3306
3307         old_id = css_id(&from->css);
3308         new_id = css_id(&to->css);
3309
3310         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3311                 mem_cgroup_swap_statistics(from, false);
3312                 mem_cgroup_swap_statistics(to, true);
3313                 /*
3314                  * This function is only called from task migration context now.
3315                  * It postpones res_counter and refcount handling till the end
3316                  * of task migration(mem_cgroup_clear_mc()) for performance
3317                  * improvement. But we cannot postpone mem_cgroup_get(to)
3318                  * because if the process that has been moved to @to does
3319                  * swap-in, the refcount of @to might be decreased to 0.
3320                  */
3321                 mem_cgroup_get(to);
3322                 if (need_fixup) {
3323                         if (!mem_cgroup_is_root(from))
3324                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3325                         mem_cgroup_put(from);
3326                         /*
3327                          * we charged both to->res and to->memsw, so we should
3328                          * uncharge to->res.
3329                          */
3330                         if (!mem_cgroup_is_root(to))
3331                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3332                 }
3333                 return 0;
3334         }
3335         return -EINVAL;
3336 }
3337 #else
3338 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3339                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3340 {
3341         return -EINVAL;
3342 }
3343 #endif
3344
3345 /*
3346  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3347  * page belongs to.
3348  */
3349 int mem_cgroup_prepare_migration(struct page *page,
3350         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3351 {
3352         struct mem_cgroup *mem = NULL;
3353         struct page_cgroup *pc;
3354         enum charge_type ctype;
3355         int ret = 0;
3356
3357         *ptr = NULL;
3358
3359         VM_BUG_ON(PageTransHuge(page));
3360         if (mem_cgroup_disabled())
3361                 return 0;
3362
3363         pc = lookup_page_cgroup(page);
3364         lock_page_cgroup(pc);
3365         if (PageCgroupUsed(pc)) {
3366                 mem = pc->mem_cgroup;
3367                 css_get(&mem->css);
3368                 /*
3369                  * At migrating an anonymous page, its mapcount goes down
3370                  * to 0 and uncharge() will be called. But, even if it's fully
3371                  * unmapped, migration may fail and this page has to be
3372                  * charged again. We set MIGRATION flag here and delay uncharge
3373                  * until end_migration() is called
3374                  *
3375                  * Corner Case Thinking
3376                  * A)
3377                  * When the old page was mapped as Anon and it's unmap-and-freed
3378                  * while migration was ongoing.
3379                  * If unmap finds the old page, uncharge() of it will be delayed
3380                  * until end_migration(). If unmap finds a new page, it's
3381                  * uncharged when it make mapcount to be 1->0. If unmap code
3382                  * finds swap_migration_entry, the new page will not be mapped
3383                  * and end_migration() will find it(mapcount==0).
3384                  *
3385                  * B)
3386                  * When the old page was mapped but migraion fails, the kernel
3387                  * remaps it. A charge for it is kept by MIGRATION flag even
3388                  * if mapcount goes down to 0. We can do remap successfully
3389                  * without charging it again.
3390                  *
3391                  * C)
3392                  * The "old" page is under lock_page() until the end of
3393                  * migration, so, the old page itself will not be swapped-out.
3394                  * If the new page is swapped out before end_migraton, our
3395                  * hook to usual swap-out path will catch the event.
3396                  */
3397                 if (PageAnon(page))
3398                         SetPageCgroupMigration(pc);
3399         }
3400         unlock_page_cgroup(pc);
3401         /*
3402          * If the page is not charged at this point,
3403          * we return here.
3404          */
3405         if (!mem)
3406                 return 0;
3407
3408         *ptr = mem;
3409         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3410         css_put(&mem->css);/* drop extra refcnt */
3411         if (ret || *ptr == NULL) {
3412                 if (PageAnon(page)) {
3413                         lock_page_cgroup(pc);
3414                         ClearPageCgroupMigration(pc);
3415                         unlock_page_cgroup(pc);
3416                         /*
3417                          * The old page may be fully unmapped while we kept it.
3418                          */
3419                         mem_cgroup_uncharge_page(page);
3420                 }
3421                 return -ENOMEM;
3422         }
3423         /*
3424          * We charge new page before it's used/mapped. So, even if unlock_page()
3425          * is called before end_migration, we can catch all events on this new
3426          * page. In the case new page is migrated but not remapped, new page's
3427          * mapcount will be finally 0 and we call uncharge in end_migration().
3428          */
3429         pc = lookup_page_cgroup(newpage);
3430         if (PageAnon(page))
3431                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3432         else if (page_is_file_cache(page))
3433                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3434         else
3435                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3436         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3437         return ret;
3438 }
3439
3440 /* remove redundant charge if migration failed*/
3441 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3442         struct page *oldpage, struct page *newpage, bool migration_ok)
3443 {
3444         struct page *used, *unused;
3445         struct page_cgroup *pc;
3446
3447         if (!mem)
3448                 return;
3449         /* blocks rmdir() */
3450         cgroup_exclude_rmdir(&mem->css);
3451         if (!migration_ok) {
3452                 used = oldpage;
3453                 unused = newpage;
3454         } else {
3455                 used = newpage;
3456                 unused = oldpage;
3457         }
3458         /*
3459          * We disallowed uncharge of pages under migration because mapcount
3460          * of the page goes down to zero, temporarly.
3461          * Clear the flag and check the page should be charged.
3462          */
3463         pc = lookup_page_cgroup(oldpage);
3464         lock_page_cgroup(pc);
3465         ClearPageCgroupMigration(pc);
3466         unlock_page_cgroup(pc);
3467
3468         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3469
3470         /*
3471          * If a page is a file cache, radix-tree replacement is very atomic
3472          * and we can skip this check. When it was an Anon page, its mapcount
3473          * goes down to 0. But because we added MIGRATION flage, it's not
3474          * uncharged yet. There are several case but page->mapcount check
3475          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3476          * check. (see prepare_charge() also)
3477          */
3478         if (PageAnon(used))
3479                 mem_cgroup_uncharge_page(used);
3480         /*
3481          * At migration, we may charge account against cgroup which has no
3482          * tasks.
3483          * So, rmdir()->pre_destroy() can be called while we do this charge.
3484          * In that case, we need to call pre_destroy() again. check it here.
3485          */
3486         cgroup_release_and_wakeup_rmdir(&mem->css);
3487 }
3488
3489 /*
3490  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3491  * Calling hierarchical_reclaim is not enough because we should update
3492  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3493  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3494  * not from the memcg which this page would be charged to.
3495  * try_charge_swapin does all of these works properly.
3496  */
3497 int mem_cgroup_shmem_charge_fallback(struct page *page,
3498                             struct mm_struct *mm,
3499                             gfp_t gfp_mask)
3500 {
3501         struct mem_cgroup *mem;
3502         int ret;
3503
3504         if (mem_cgroup_disabled())
3505                 return 0;
3506
3507         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3508         if (!ret)
3509                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3510
3511         return ret;
3512 }
3513
3514 #ifdef CONFIG_DEBUG_VM
3515 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3516 {
3517         struct page_cgroup *pc;
3518
3519         pc = lookup_page_cgroup(page);
3520         if (likely(pc) && PageCgroupUsed(pc))
3521                 return pc;
3522         return NULL;
3523 }
3524
3525 bool mem_cgroup_bad_page_check(struct page *page)
3526 {
3527         if (mem_cgroup_disabled())
3528                 return false;
3529
3530         return lookup_page_cgroup_used(page) != NULL;
3531 }
3532
3533 void mem_cgroup_print_bad_page(struct page *page)
3534 {
3535         struct page_cgroup *pc;
3536
3537         pc = lookup_page_cgroup_used(page);
3538         if (pc) {
3539                 int ret = -1;
3540                 char *path;
3541
3542                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3543                        pc, pc->flags, pc->mem_cgroup);
3544
3545                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3546                 if (path) {
3547                         rcu_read_lock();
3548                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3549                                                         path, PATH_MAX);
3550                         rcu_read_unlock();
3551                 }
3552
3553                 printk(KERN_CONT "(%s)\n",
3554                                 (ret < 0) ? "cannot get the path" : path);
3555                 kfree(path);
3556         }
3557 }
3558 #endif
3559
3560 static DEFINE_MUTEX(set_limit_mutex);
3561
3562 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3563                                 unsigned long long val)
3564 {
3565         int retry_count;
3566         u64 memswlimit, memlimit;
3567         int ret = 0;
3568         int children = mem_cgroup_count_children(memcg);
3569         u64 curusage, oldusage;
3570         int enlarge;
3571
3572         /*
3573          * For keeping hierarchical_reclaim simple, how long we should retry
3574          * is depends on callers. We set our retry-count to be function
3575          * of # of children which we should visit in this loop.
3576          */
3577         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3578
3579         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3580
3581         enlarge = 0;
3582         while (retry_count) {
3583                 if (signal_pending(current)) {
3584                         ret = -EINTR;
3585                         break;
3586                 }
3587                 /*
3588                  * Rather than hide all in some function, I do this in
3589                  * open coded manner. You see what this really does.
3590                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3591                  */
3592                 mutex_lock(&set_limit_mutex);
3593                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3594                 if (memswlimit < val) {
3595                         ret = -EINVAL;
3596                         mutex_unlock(&set_limit_mutex);
3597                         break;
3598                 }
3599
3600                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3601                 if (memlimit < val)
3602                         enlarge = 1;
3603
3604                 ret = res_counter_set_limit(&memcg->res, val);
3605                 if (!ret) {
3606                         if (memswlimit == val)
3607                                 memcg->memsw_is_minimum = true;
3608                         else
3609                                 memcg->memsw_is_minimum = false;
3610                 }
3611                 mutex_unlock(&set_limit_mutex);
3612
3613                 if (!ret)
3614                         break;
3615
3616                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3617                                                 MEM_CGROUP_RECLAIM_SHRINK,
3618                                                 NULL);
3619                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3620                 /* Usage is reduced ? */
3621                 if (curusage >= oldusage)
3622                         retry_count--;
3623                 else
3624                         oldusage = curusage;
3625         }
3626         if (!ret && enlarge)
3627                 memcg_oom_recover(memcg);
3628
3629         return ret;
3630 }
3631
3632 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3633                                         unsigned long long val)
3634 {
3635         int retry_count;
3636         u64 memlimit, memswlimit, oldusage, curusage;
3637         int children = mem_cgroup_count_children(memcg);
3638         int ret = -EBUSY;
3639         int enlarge = 0;
3640
3641         /* see mem_cgroup_resize_res_limit */
3642         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3643         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3644         while (retry_count) {
3645                 if (signal_pending(current)) {
3646                         ret = -EINTR;
3647                         break;
3648                 }
3649                 /*
3650                  * Rather than hide all in some function, I do this in
3651                  * open coded manner. You see what this really does.
3652                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3653                  */
3654                 mutex_lock(&set_limit_mutex);
3655                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3656                 if (memlimit > val) {
3657                         ret = -EINVAL;
3658                         mutex_unlock(&set_limit_mutex);
3659                         break;
3660                 }
3661                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3662                 if (memswlimit < val)
3663                         enlarge = 1;
3664                 ret = res_counter_set_limit(&memcg->memsw, val);
3665                 if (!ret) {
3666                         if (memlimit == val)
3667                                 memcg->memsw_is_minimum = true;
3668                         else
3669                                 memcg->memsw_is_minimum = false;
3670                 }
3671                 mutex_unlock(&set_limit_mutex);
3672
3673                 if (!ret)
3674                         break;
3675
3676                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3677                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3678                                                 MEM_CGROUP_RECLAIM_SHRINK,
3679                                                 NULL);
3680                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3681                 /* Usage is reduced ? */
3682                 if (curusage >= oldusage)
3683                         retry_count--;
3684                 else
3685                         oldusage = curusage;
3686         }
3687         if (!ret && enlarge)
3688                 memcg_oom_recover(memcg);
3689         return ret;
3690 }
3691
3692 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3693                                             gfp_t gfp_mask,
3694                                             unsigned long *total_scanned)
3695 {
3696         unsigned long nr_reclaimed = 0;
3697         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3698         unsigned long reclaimed;
3699         int loop = 0;
3700         struct mem_cgroup_tree_per_zone *mctz;
3701         unsigned long long excess;
3702         unsigned long nr_scanned;
3703
3704         if (order > 0)
3705                 return 0;
3706
3707         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3708         /*
3709          * This loop can run a while, specially if mem_cgroup's continuously
3710          * keep exceeding their soft limit and putting the system under
3711          * pressure
3712          */
3713         do {
3714                 if (next_mz)
3715                         mz = next_mz;
3716                 else
3717                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3718                 if (!mz)
3719                         break;
3720
3721                 nr_scanned = 0;
3722                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3723                                                 gfp_mask,
3724                                                 MEM_CGROUP_RECLAIM_SOFT,
3725                                                 &nr_scanned);
3726                 nr_reclaimed += reclaimed;
3727                 *total_scanned += nr_scanned;
3728                 spin_lock(&mctz->lock);
3729
3730                 /*
3731                  * If we failed to reclaim anything from this memory cgroup
3732                  * it is time to move on to the next cgroup
3733                  */
3734                 next_mz = NULL;
3735                 if (!reclaimed) {
3736                         do {
3737                                 /*
3738                                  * Loop until we find yet another one.
3739                                  *
3740                                  * By the time we get the soft_limit lock
3741                                  * again, someone might have aded the
3742                                  * group back on the RB tree. Iterate to
3743                                  * make sure we get a different mem.
3744                                  * mem_cgroup_largest_soft_limit_node returns
3745                                  * NULL if no other cgroup is present on
3746                                  * the tree
3747                                  */
3748                                 next_mz =
3749                                 __mem_cgroup_largest_soft_limit_node(mctz);
3750                                 if (next_mz == mz)
3751                                         css_put(&next_mz->mem->css);
3752                                 else /* next_mz == NULL or other memcg */
3753                                         break;
3754                         } while (1);
3755                 }
3756                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3757                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3758                 /*
3759                  * One school of thought says that we should not add
3760                  * back the node to the tree if reclaim returns 0.
3761                  * But our reclaim could return 0, simply because due
3762                  * to priority we are exposing a smaller subset of
3763                  * memory to reclaim from. Consider this as a longer
3764                  * term TODO.
3765                  */
3766                 /* If excess == 0, no tree ops */
3767                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3768                 spin_unlock(&mctz->lock);
3769                 css_put(&mz->mem->css);
3770                 loop++;
3771                 /*
3772                  * Could not reclaim anything and there are no more
3773                  * mem cgroups to try or we seem to be looping without
3774                  * reclaiming anything.
3775                  */
3776                 if (!nr_reclaimed &&
3777                         (next_mz == NULL ||
3778                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3779                         break;
3780         } while (!nr_reclaimed);
3781         if (next_mz)
3782                 css_put(&next_mz->mem->css);
3783         return nr_reclaimed;
3784 }
3785
3786 /*
3787  * This routine traverse page_cgroup in given list and drop them all.
3788  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3789  */
3790 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3791                                 int node, int zid, enum lru_list lru)
3792 {
3793         struct zone *zone;
3794         struct mem_cgroup_per_zone *mz;
3795         struct page_cgroup *pc, *busy;
3796         unsigned long flags, loop;
3797         struct list_head *list;
3798         int ret = 0;
3799
3800         zone = &NODE_DATA(node)->node_zones[zid];
3801         mz = mem_cgroup_zoneinfo(mem, node, zid);
3802         list = &mz->lists[lru];
3803
3804         loop = MEM_CGROUP_ZSTAT(mz, lru);
3805         /* give some margin against EBUSY etc...*/
3806         loop += 256;
3807         busy = NULL;
3808         while (loop--) {
3809                 struct page *page;
3810
3811                 ret = 0;
3812                 spin_lock_irqsave(&zone->lru_lock, flags);
3813                 if (list_empty(list)) {
3814                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3815                         break;
3816                 }
3817                 pc = list_entry(list->prev, struct page_cgroup, lru);
3818                 if (busy == pc) {
3819                         list_move(&pc->lru, list);
3820                         busy = NULL;
3821                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3822                         continue;
3823                 }
3824                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3825
3826                 page = lookup_cgroup_page(pc);
3827
3828                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3829                 if (ret == -ENOMEM)
3830                         break;
3831
3832                 if (ret == -EBUSY || ret == -EINVAL) {
3833                         /* found lock contention or "pc" is obsolete. */
3834                         busy = pc;
3835                         cond_resched();
3836                 } else
3837                         busy = NULL;
3838         }
3839
3840         if (!ret && !list_empty(list))
3841                 return -EBUSY;
3842         return ret;
3843 }
3844
3845 /*
3846  * make mem_cgroup's charge to be 0 if there is no task.
3847  * This enables deleting this mem_cgroup.
3848  */
3849 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3850 {
3851         int ret;
3852         int node, zid, shrink;
3853         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3854         struct cgroup *cgrp = mem->css.cgroup;
3855
3856         css_get(&mem->css);
3857
3858         shrink = 0;
3859         /* should free all ? */
3860         if (free_all)
3861                 goto try_to_free;
3862 move_account:
3863         do {
3864                 ret = -EBUSY;
3865                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3866                         goto out;
3867                 ret = -EINTR;
3868                 if (signal_pending(current))
3869                         goto out;
3870                 /* This is for making all *used* pages to be on LRU. */
3871                 lru_add_drain_all();
3872                 drain_all_stock_sync(mem);
3873                 ret = 0;
3874                 mem_cgroup_start_move(mem);
3875                 for_each_node_state(node, N_HIGH_MEMORY) {
3876                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3877                                 enum lru_list l;
3878                                 for_each_lru(l) {
3879                                         ret = mem_cgroup_force_empty_list(mem,
3880                                                         node, zid, l);
3881                                         if (ret)
3882                                                 break;
3883                                 }
3884                         }
3885                         if (ret)
3886                                 break;
3887                 }
3888                 mem_cgroup_end_move(mem);
3889                 memcg_oom_recover(mem);
3890                 /* it seems parent cgroup doesn't have enough mem */
3891                 if (ret == -ENOMEM)
3892                         goto try_to_free;
3893                 cond_resched();
3894         /* "ret" should also be checked to ensure all lists are empty. */
3895         } while (mem->res.usage > 0 || ret);
3896 out:
3897         css_put(&mem->css);
3898         return ret;
3899
3900 try_to_free:
3901         /* returns EBUSY if there is a task or if we come here twice. */
3902         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3903                 ret = -EBUSY;
3904                 goto out;
3905         }
3906         /* we call try-to-free pages for make this cgroup empty */
3907         lru_add_drain_all();
3908         /* try to free all pages in this cgroup */
3909         shrink = 1;
3910         while (nr_retries && mem->res.usage > 0) {
3911                 struct memcg_scanrecord rec;
3912                 int progress;
3913
3914                 if (signal_pending(current)) {
3915                         ret = -EINTR;
3916                         goto out;
3917                 }
3918                 rec.context = SCAN_BY_SHRINK;
3919                 rec.mem = mem;
3920                 rec.root = mem;
3921                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3922                                                 false, &rec);
3923                 if (!progress) {
3924                         nr_retries--;
3925                         /* maybe some writeback is necessary */
3926                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3927                 }
3928
3929         }
3930         lru_add_drain();
3931         /* try move_account...there may be some *locked* pages. */
3932         goto move_account;
3933 }
3934
3935 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3936 {
3937         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3938 }
3939
3940
3941 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3942 {
3943         return mem_cgroup_from_cont(cont)->use_hierarchy;
3944 }
3945
3946 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3947                                         u64 val)
3948 {
3949         int retval = 0;
3950         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3951         struct cgroup *parent = cont->parent;
3952         struct mem_cgroup *parent_mem = NULL;
3953
3954         if (parent)
3955                 parent_mem = mem_cgroup_from_cont(parent);
3956
3957         cgroup_lock();
3958         /*
3959          * If parent's use_hierarchy is set, we can't make any modifications
3960          * in the child subtrees. If it is unset, then the change can
3961          * occur, provided the current cgroup has no children.
3962          *
3963          * For the root cgroup, parent_mem is NULL, we allow value to be
3964          * set if there are no children.
3965          */
3966         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3967                                 (val == 1 || val == 0)) {
3968                 if (list_empty(&cont->children))
3969                         mem->use_hierarchy = val;
3970                 else
3971                         retval = -EBUSY;
3972         } else
3973                 retval = -EINVAL;
3974         cgroup_unlock();
3975
3976         return retval;
3977 }
3978
3979
3980 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3981                                                enum mem_cgroup_stat_index idx)
3982 {
3983         struct mem_cgroup *iter;
3984         long val = 0;
3985
3986         /* Per-cpu values can be negative, use a signed accumulator */
3987         for_each_mem_cgroup_tree(iter, mem)
3988                 val += mem_cgroup_read_stat(iter, idx);
3989
3990         if (val < 0) /* race ? */
3991                 val = 0;
3992         return val;
3993 }
3994
3995 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3996 {
3997         u64 val;
3998
3999         if (!mem_cgroup_is_root(mem)) {
4000                 if (!swap)
4001                         return res_counter_read_u64(&mem->res, RES_USAGE);
4002                 else
4003                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
4004         }
4005
4006         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
4007         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
4008
4009         if (swap)
4010                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4011
4012         return val << PAGE_SHIFT;
4013 }
4014
4015 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
4016 {
4017         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4018         u64 val;
4019         int type, name;
4020
4021         type = MEMFILE_TYPE(cft->private);
4022         name = MEMFILE_ATTR(cft->private);
4023         switch (type) {
4024         case _MEM:
4025                 if (name == RES_USAGE)
4026                         val = mem_cgroup_usage(mem, false);
4027                 else
4028                         val = res_counter_read_u64(&mem->res, name);
4029                 break;
4030         case _MEMSWAP:
4031                 if (name == RES_USAGE)
4032                         val = mem_cgroup_usage(mem, true);
4033                 else
4034                         val = res_counter_read_u64(&mem->memsw, name);
4035                 break;
4036         default:
4037                 BUG();
4038                 break;
4039         }
4040         return val;
4041 }
4042 /*
4043  * The user of this function is...
4044  * RES_LIMIT.
4045  */
4046 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4047                             const char *buffer)
4048 {
4049         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4050         int type, name;
4051         unsigned long long val;
4052         int ret;
4053
4054         type = MEMFILE_TYPE(cft->private);
4055         name = MEMFILE_ATTR(cft->private);
4056         switch (name) {
4057         case RES_LIMIT:
4058                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4059                         ret = -EINVAL;
4060                         break;
4061                 }
4062                 /* This function does all necessary parse...reuse it */
4063                 ret = res_counter_memparse_write_strategy(buffer, &val);
4064                 if (ret)
4065                         break;
4066                 if (type == _MEM)
4067                         ret = mem_cgroup_resize_limit(memcg, val);
4068                 else
4069                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
4070                 break;
4071         case RES_SOFT_LIMIT:
4072                 ret = res_counter_memparse_write_strategy(buffer, &val);
4073                 if (ret)
4074                         break;
4075                 /*
4076                  * For memsw, soft limits are hard to implement in terms
4077                  * of semantics, for now, we support soft limits for
4078                  * control without swap
4079                  */
4080                 if (type == _MEM)
4081                         ret = res_counter_set_soft_limit(&memcg->res, val);
4082                 else
4083                         ret = -EINVAL;
4084                 break;
4085         default:
4086                 ret = -EINVAL; /* should be BUG() ? */
4087                 break;
4088         }
4089         return ret;
4090 }
4091
4092 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4093                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4094 {
4095         struct cgroup *cgroup;
4096         unsigned long long min_limit, min_memsw_limit, tmp;
4097
4098         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4099         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4100         cgroup = memcg->css.cgroup;
4101         if (!memcg->use_hierarchy)
4102                 goto out;
4103
4104         while (cgroup->parent) {
4105                 cgroup = cgroup->parent;
4106                 memcg = mem_cgroup_from_cont(cgroup);
4107                 if (!memcg->use_hierarchy)
4108                         break;
4109                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4110                 min_limit = min(min_limit, tmp);
4111                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4112                 min_memsw_limit = min(min_memsw_limit, tmp);
4113         }
4114 out:
4115         *mem_limit = min_limit;
4116         *memsw_limit = min_memsw_limit;
4117         return;
4118 }
4119
4120 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4121 {
4122         struct mem_cgroup *mem;
4123         int type, name;
4124
4125         mem = mem_cgroup_from_cont(cont);
4126         type = MEMFILE_TYPE(event);
4127         name = MEMFILE_ATTR(event);
4128         switch (name) {
4129         case RES_MAX_USAGE:
4130                 if (type == _MEM)
4131                         res_counter_reset_max(&mem->res);
4132                 else
4133                         res_counter_reset_max(&mem->memsw);
4134                 break;
4135         case RES_FAILCNT:
4136                 if (type == _MEM)
4137                         res_counter_reset_failcnt(&mem->res);
4138                 else
4139                         res_counter_reset_failcnt(&mem->memsw);
4140                 break;
4141         }
4142
4143         return 0;
4144 }
4145
4146 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4147                                         struct cftype *cft)
4148 {
4149         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4150 }
4151
4152 #ifdef CONFIG_MMU
4153 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4154                                         struct cftype *cft, u64 val)
4155 {
4156         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4157
4158         if (val >= (1 << NR_MOVE_TYPE))
4159                 return -EINVAL;
4160         /*
4161          * We check this value several times in both in can_attach() and
4162          * attach(), so we need cgroup lock to prevent this value from being
4163          * inconsistent.
4164          */
4165         cgroup_lock();
4166         mem->move_charge_at_immigrate = val;
4167         cgroup_unlock();
4168
4169         return 0;
4170 }
4171 #else
4172 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4173                                         struct cftype *cft, u64 val)
4174 {
4175         return -ENOSYS;
4176 }
4177 #endif
4178
4179
4180 /* For read statistics */
4181 enum {
4182         MCS_CACHE,
4183         MCS_RSS,
4184         MCS_FILE_MAPPED,
4185         MCS_PGPGIN,
4186         MCS_PGPGOUT,
4187         MCS_SWAP,
4188         MCS_PGFAULT,
4189         MCS_PGMAJFAULT,
4190         MCS_INACTIVE_ANON,
4191         MCS_ACTIVE_ANON,
4192         MCS_INACTIVE_FILE,
4193         MCS_ACTIVE_FILE,
4194         MCS_UNEVICTABLE,
4195         NR_MCS_STAT,
4196 };
4197
4198 struct mcs_total_stat {
4199         s64 stat[NR_MCS_STAT];
4200 };
4201
4202 struct {
4203         char *local_name;
4204         char *total_name;
4205 } memcg_stat_strings[NR_MCS_STAT] = {
4206         {"cache", "total_cache"},
4207         {"rss", "total_rss"},
4208         {"mapped_file", "total_mapped_file"},
4209         {"pgpgin", "total_pgpgin"},
4210         {"pgpgout", "total_pgpgout"},
4211         {"swap", "total_swap"},
4212         {"pgfault", "total_pgfault"},
4213         {"pgmajfault", "total_pgmajfault"},
4214         {"inactive_anon", "total_inactive_anon"},
4215         {"active_anon", "total_active_anon"},
4216         {"inactive_file", "total_inactive_file"},
4217         {"active_file", "total_active_file"},
4218         {"unevictable", "total_unevictable"}
4219 };
4220
4221
4222 static void
4223 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4224 {
4225         s64 val;
4226
4227         /* per cpu stat */
4228         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4229         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4230         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4231         s->stat[MCS_RSS] += val * PAGE_SIZE;
4232         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4233         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4234         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4235         s->stat[MCS_PGPGIN] += val;
4236         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4237         s->stat[MCS_PGPGOUT] += val;
4238         if (do_swap_account) {
4239                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4240                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4241         }
4242         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4243         s->stat[MCS_PGFAULT] += val;
4244         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4245         s->stat[MCS_PGMAJFAULT] += val;
4246
4247         /* per zone stat */
4248         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4249         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4250         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4251         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4252         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4253         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4254         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4255         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4256         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4257         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4258 }
4259
4260 static void
4261 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4262 {
4263         struct mem_cgroup *iter;
4264
4265         for_each_mem_cgroup_tree(iter, mem)
4266                 mem_cgroup_get_local_stat(iter, s);
4267 }
4268
4269 #ifdef CONFIG_NUMA
4270 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4271 {
4272         int nid;
4273         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4274         unsigned long node_nr;
4275         struct cgroup *cont = m->private;
4276         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4277
4278         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4279         seq_printf(m, "total=%lu", total_nr);
4280         for_each_node_state(nid, N_HIGH_MEMORY) {
4281                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4282                 seq_printf(m, " N%d=%lu", nid, node_nr);
4283         }
4284         seq_putc(m, '\n');
4285
4286         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4287         seq_printf(m, "file=%lu", file_nr);
4288         for_each_node_state(nid, N_HIGH_MEMORY) {
4289                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4290                                 LRU_ALL_FILE);
4291                 seq_printf(m, " N%d=%lu", nid, node_nr);
4292         }
4293         seq_putc(m, '\n');
4294
4295         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4296         seq_printf(m, "anon=%lu", anon_nr);
4297         for_each_node_state(nid, N_HIGH_MEMORY) {
4298                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4299                                 LRU_ALL_ANON);
4300                 seq_printf(m, " N%d=%lu", nid, node_nr);
4301         }
4302         seq_putc(m, '\n');
4303
4304         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4305         seq_printf(m, "unevictable=%lu", unevictable_nr);
4306         for_each_node_state(nid, N_HIGH_MEMORY) {
4307                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4308                                 BIT(LRU_UNEVICTABLE));
4309                 seq_printf(m, " N%d=%lu", nid, node_nr);
4310         }
4311         seq_putc(m, '\n');
4312         return 0;
4313 }
4314 #endif /* CONFIG_NUMA */
4315
4316 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4317                                  struct cgroup_map_cb *cb)
4318 {
4319         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4320         struct mcs_total_stat mystat;
4321         int i;
4322
4323         memset(&mystat, 0, sizeof(mystat));
4324         mem_cgroup_get_local_stat(mem_cont, &mystat);
4325
4326
4327         for (i = 0; i < NR_MCS_STAT; i++) {
4328                 if (i == MCS_SWAP && !do_swap_account)
4329                         continue;
4330                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4331         }
4332
4333         /* Hierarchical information */
4334         {
4335                 unsigned long long limit, memsw_limit;
4336                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4337                 cb->fill(cb, "hierarchical_memory_limit", limit);
4338                 if (do_swap_account)
4339                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4340         }
4341
4342         memset(&mystat, 0, sizeof(mystat));
4343         mem_cgroup_get_total_stat(mem_cont, &mystat);
4344         for (i = 0; i < NR_MCS_STAT; i++) {
4345                 if (i == MCS_SWAP && !do_swap_account)
4346                         continue;
4347                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4348         }
4349
4350 #ifdef CONFIG_DEBUG_VM
4351         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4352
4353         {
4354                 int nid, zid;
4355                 struct mem_cgroup_per_zone *mz;
4356                 unsigned long recent_rotated[2] = {0, 0};
4357                 unsigned long recent_scanned[2] = {0, 0};
4358
4359                 for_each_online_node(nid)
4360                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4361                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4362
4363                                 recent_rotated[0] +=
4364                                         mz->reclaim_stat.recent_rotated[0];
4365                                 recent_rotated[1] +=
4366                                         mz->reclaim_stat.recent_rotated[1];
4367                                 recent_scanned[0] +=
4368                                         mz->reclaim_stat.recent_scanned[0];
4369                                 recent_scanned[1] +=
4370                                         mz->reclaim_stat.recent_scanned[1];
4371                         }
4372                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4373                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4374                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4375                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4376         }
4377 #endif
4378
4379         return 0;
4380 }
4381
4382 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4383 {
4384         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4385
4386         return mem_cgroup_swappiness(memcg);
4387 }
4388
4389 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4390                                        u64 val)
4391 {
4392         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4393         struct mem_cgroup *parent;
4394
4395         if (val > 100)
4396                 return -EINVAL;
4397
4398         if (cgrp->parent == NULL)
4399                 return -EINVAL;
4400
4401         parent = mem_cgroup_from_cont(cgrp->parent);
4402
4403         cgroup_lock();
4404
4405         /* If under hierarchy, only empty-root can set this value */
4406         if ((parent->use_hierarchy) ||
4407             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4408                 cgroup_unlock();
4409                 return -EINVAL;
4410         }
4411
4412         memcg->swappiness = val;
4413
4414         cgroup_unlock();
4415
4416         return 0;
4417 }
4418
4419 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4420 {
4421         struct mem_cgroup_threshold_ary *t;
4422         u64 usage;
4423         int i;
4424
4425         rcu_read_lock();
4426         if (!swap)
4427                 t = rcu_dereference(memcg->thresholds.primary);
4428         else
4429                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4430
4431         if (!t)
4432                 goto unlock;
4433
4434         usage = mem_cgroup_usage(memcg, swap);
4435
4436         /*
4437          * current_threshold points to threshold just below usage.
4438          * If it's not true, a threshold was crossed after last
4439          * call of __mem_cgroup_threshold().
4440          */
4441         i = t->current_threshold;
4442
4443         /*
4444          * Iterate backward over array of thresholds starting from
4445          * current_threshold and check if a threshold is crossed.
4446          * If none of thresholds below usage is crossed, we read
4447          * only one element of the array here.
4448          */
4449         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4450                 eventfd_signal(t->entries[i].eventfd, 1);
4451
4452         /* i = current_threshold + 1 */
4453         i++;
4454
4455         /*
4456          * Iterate forward over array of thresholds starting from
4457          * current_threshold+1 and check if a threshold is crossed.
4458          * If none of thresholds above usage is crossed, we read
4459          * only one element of the array here.
4460          */
4461         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4462                 eventfd_signal(t->entries[i].eventfd, 1);
4463
4464         /* Update current_threshold */
4465         t->current_threshold = i - 1;
4466 unlock:
4467         rcu_read_unlock();
4468 }
4469
4470 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4471 {
4472         while (memcg) {
4473                 __mem_cgroup_threshold(memcg, false);
4474                 if (do_swap_account)
4475                         __mem_cgroup_threshold(memcg, true);
4476
4477                 memcg = parent_mem_cgroup(memcg);
4478         }
4479 }
4480
4481 static int compare_thresholds(const void *a, const void *b)
4482 {
4483         const struct mem_cgroup_threshold *_a = a;
4484         const struct mem_cgroup_threshold *_b = b;
4485
4486         return _a->threshold - _b->threshold;
4487 }
4488
4489 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4490 {
4491         struct mem_cgroup_eventfd_list *ev;
4492
4493         list_for_each_entry(ev, &mem->oom_notify, list)
4494                 eventfd_signal(ev->eventfd, 1);
4495         return 0;
4496 }
4497
4498 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4499 {
4500         struct mem_cgroup *iter;
4501
4502         for_each_mem_cgroup_tree(iter, mem)
4503                 mem_cgroup_oom_notify_cb(iter);
4504 }
4505
4506 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4507         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4508 {
4509         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4510         struct mem_cgroup_thresholds *thresholds;
4511         struct mem_cgroup_threshold_ary *new;
4512         int type = MEMFILE_TYPE(cft->private);
4513         u64 threshold, usage;
4514         int i, size, ret;
4515
4516         ret = res_counter_memparse_write_strategy(args, &threshold);
4517         if (ret)
4518                 return ret;
4519
4520         mutex_lock(&memcg->thresholds_lock);
4521
4522         if (type == _MEM)
4523                 thresholds = &memcg->thresholds;
4524         else if (type == _MEMSWAP)
4525                 thresholds = &memcg->memsw_thresholds;
4526         else
4527                 BUG();
4528
4529         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4530
4531         /* Check if a threshold crossed before adding a new one */
4532         if (thresholds->primary)
4533                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4534
4535         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4536
4537         /* Allocate memory for new array of thresholds */
4538         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4539                         GFP_KERNEL);
4540         if (!new) {
4541                 ret = -ENOMEM;
4542                 goto unlock;
4543         }
4544         new->size = size;
4545
4546         /* Copy thresholds (if any) to new array */
4547         if (thresholds->primary) {
4548                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4549                                 sizeof(struct mem_cgroup_threshold));
4550         }
4551
4552         /* Add new threshold */
4553         new->entries[size - 1].eventfd = eventfd;
4554         new->entries[size - 1].threshold = threshold;
4555
4556         /* Sort thresholds. Registering of new threshold isn't time-critical */
4557         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4558                         compare_thresholds, NULL);
4559
4560         /* Find current threshold */
4561         new->current_threshold = -1;
4562         for (i = 0; i < size; i++) {
4563                 if (new->entries[i].threshold < usage) {
4564                         /*
4565                          * new->current_threshold will not be used until
4566                          * rcu_assign_pointer(), so it's safe to increment
4567                          * it here.
4568                          */
4569                         ++new->current_threshold;
4570                 }
4571         }
4572
4573         /* Free old spare buffer and save old primary buffer as spare */
4574         kfree(thresholds->spare);
4575         thresholds->spare = thresholds->primary;
4576
4577         rcu_assign_pointer(thresholds->primary, new);
4578
4579         /* To be sure that nobody uses thresholds */
4580         synchronize_rcu();
4581
4582 unlock:
4583         mutex_unlock(&memcg->thresholds_lock);
4584
4585         return ret;
4586 }
4587
4588 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4589         struct cftype *cft, struct eventfd_ctx *eventfd)
4590 {
4591         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4592         struct mem_cgroup_thresholds *thresholds;
4593         struct mem_cgroup_threshold_ary *new;
4594         int type = MEMFILE_TYPE(cft->private);
4595         u64 usage;
4596         int i, j, size;
4597
4598         mutex_lock(&memcg->thresholds_lock);
4599         if (type == _MEM)
4600                 thresholds = &memcg->thresholds;
4601         else if (type == _MEMSWAP)
4602                 thresholds = &memcg->memsw_thresholds;
4603         else
4604                 BUG();
4605
4606         /*
4607          * Something went wrong if we trying to unregister a threshold
4608          * if we don't have thresholds
4609          */
4610         BUG_ON(!thresholds);
4611
4612         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4613
4614         /* Check if a threshold crossed before removing */
4615         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4616
4617         /* Calculate new number of threshold */
4618         size = 0;
4619         for (i = 0; i < thresholds->primary->size; i++) {
4620                 if (thresholds->primary->entries[i].eventfd != eventfd)
4621                         size++;
4622         }
4623
4624         new = thresholds->spare;
4625
4626         /* Set thresholds array to NULL if we don't have thresholds */
4627         if (!size) {
4628                 kfree(new);
4629                 new = NULL;
4630                 goto swap_buffers;
4631         }
4632
4633         new->size = size;
4634
4635         /* Copy thresholds and find current threshold */
4636         new->current_threshold = -1;
4637         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4638                 if (thresholds->primary->entries[i].eventfd == eventfd)
4639                         continue;
4640
4641                 new->entries[j] = thresholds->primary->entries[i];
4642                 if (new->entries[j].threshold < usage) {
4643                         /*
4644                          * new->current_threshold will not be used
4645                          * until rcu_assign_pointer(), so it's safe to increment
4646                          * it here.
4647                          */
4648                         ++new->current_threshold;
4649                 }
4650                 j++;
4651         }
4652
4653 swap_buffers:
4654         /* Swap primary and spare array */
4655         thresholds->spare = thresholds->primary;
4656         rcu_assign_pointer(thresholds->primary, new);
4657
4658         /* To be sure that nobody uses thresholds */
4659         synchronize_rcu();
4660
4661         mutex_unlock(&memcg->thresholds_lock);
4662 }
4663
4664 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4665         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4666 {
4667         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4668         struct mem_cgroup_eventfd_list *event;
4669         int type = MEMFILE_TYPE(cft->private);
4670
4671         BUG_ON(type != _OOM_TYPE);
4672         event = kmalloc(sizeof(*event), GFP_KERNEL);
4673         if (!event)
4674                 return -ENOMEM;
4675
4676         spin_lock(&memcg_oom_lock);
4677
4678         event->eventfd = eventfd;
4679         list_add(&event->list, &memcg->oom_notify);
4680
4681         /* already in OOM ? */
4682         if (atomic_read(&memcg->under_oom))
4683                 eventfd_signal(eventfd, 1);
4684         spin_unlock(&memcg_oom_lock);
4685
4686         return 0;
4687 }
4688
4689 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4690         struct cftype *cft, struct eventfd_ctx *eventfd)
4691 {
4692         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4693         struct mem_cgroup_eventfd_list *ev, *tmp;
4694         int type = MEMFILE_TYPE(cft->private);
4695
4696         BUG_ON(type != _OOM_TYPE);
4697
4698         spin_lock(&memcg_oom_lock);
4699
4700         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4701                 if (ev->eventfd == eventfd) {
4702                         list_del(&ev->list);
4703                         kfree(ev);
4704                 }
4705         }
4706
4707         spin_unlock(&memcg_oom_lock);
4708 }
4709
4710 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4711         struct cftype *cft,  struct cgroup_map_cb *cb)
4712 {
4713         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4714
4715         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4716
4717         if (atomic_read(&mem->under_oom))
4718                 cb->fill(cb, "under_oom", 1);
4719         else
4720                 cb->fill(cb, "under_oom", 0);
4721         return 0;
4722 }
4723
4724 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4725         struct cftype *cft, u64 val)
4726 {
4727         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4728         struct mem_cgroup *parent;
4729
4730         /* cannot set to root cgroup and only 0 and 1 are allowed */
4731         if (!cgrp->parent || !((val == 0) || (val == 1)))
4732                 return -EINVAL;
4733
4734         parent = mem_cgroup_from_cont(cgrp->parent);
4735
4736         cgroup_lock();
4737         /* oom-kill-disable is a flag for subhierarchy. */
4738         if ((parent->use_hierarchy) ||
4739             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4740                 cgroup_unlock();
4741                 return -EINVAL;
4742         }
4743         mem->oom_kill_disable = val;
4744         if (!val)
4745                 memcg_oom_recover(mem);
4746         cgroup_unlock();
4747         return 0;
4748 }
4749
4750 #ifdef CONFIG_NUMA
4751 static const struct file_operations mem_control_numa_stat_file_operations = {
4752         .read = seq_read,
4753         .llseek = seq_lseek,
4754         .release = single_release,
4755 };
4756
4757 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4758 {
4759         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4760
4761         file->f_op = &mem_control_numa_stat_file_operations;
4762         return single_open(file, mem_control_numa_stat_show, cont);
4763 }
4764 #endif /* CONFIG_NUMA */
4765
4766 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4767                                 struct cftype *cft,
4768                                 struct cgroup_map_cb *cb)
4769 {
4770         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4771         char string[64];
4772         int i;
4773
4774         for (i = 0; i < NR_SCANSTATS; i++) {
4775                 strcpy(string, scanstat_string[i]);
4776                 strcat(string, SCANSTAT_WORD_LIMIT);
4777                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4778         }
4779
4780         for (i = 0; i < NR_SCANSTATS; i++) {
4781                 strcpy(string, scanstat_string[i]);
4782                 strcat(string, SCANSTAT_WORD_SYSTEM);
4783                 cb->fill(cb, string,  mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4784         }
4785
4786         for (i = 0; i < NR_SCANSTATS; i++) {
4787                 strcpy(string, scanstat_string[i]);
4788                 strcat(string, SCANSTAT_WORD_LIMIT);
4789                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4790                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4791         }
4792         for (i = 0; i < NR_SCANSTATS; i++) {
4793                 strcpy(string, scanstat_string[i]);
4794                 strcat(string, SCANSTAT_WORD_SYSTEM);
4795                 strcat(string, SCANSTAT_WORD_HIERARCHY);
4796                 cb->fill(cb, string,  mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4797         }
4798         return 0;
4799 }
4800
4801 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4802                                 unsigned int event)
4803 {
4804         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4805
4806         spin_lock(&mem->scanstat.lock);
4807         memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4808         memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4809         spin_unlock(&mem->scanstat.lock);
4810         return 0;
4811 }
4812
4813
4814 static struct cftype mem_cgroup_files[] = {
4815         {
4816                 .name = "usage_in_bytes",
4817                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4818                 .read_u64 = mem_cgroup_read,
4819                 .register_event = mem_cgroup_usage_register_event,
4820                 .unregister_event = mem_cgroup_usage_unregister_event,
4821         },
4822         {
4823                 .name = "max_usage_in_bytes",
4824                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4825                 .trigger = mem_cgroup_reset,
4826                 .read_u64 = mem_cgroup_read,
4827         },
4828         {
4829                 .name = "limit_in_bytes",
4830                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4831                 .write_string = mem_cgroup_write,
4832                 .read_u64 = mem_cgroup_read,
4833         },
4834         {
4835                 .name = "soft_limit_in_bytes",
4836                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4837                 .write_string = mem_cgroup_write,
4838                 .read_u64 = mem_cgroup_read,
4839         },
4840         {
4841                 .name = "failcnt",
4842                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4843                 .trigger = mem_cgroup_reset,
4844                 .read_u64 = mem_cgroup_read,
4845         },
4846         {
4847                 .name = "stat",
4848                 .read_map = mem_control_stat_show,
4849         },
4850         {
4851                 .name = "force_empty",
4852                 .trigger = mem_cgroup_force_empty_write,
4853         },
4854         {
4855                 .name = "use_hierarchy",
4856                 .write_u64 = mem_cgroup_hierarchy_write,
4857                 .read_u64 = mem_cgroup_hierarchy_read,
4858         },
4859         {
4860                 .name = "swappiness",
4861                 .read_u64 = mem_cgroup_swappiness_read,
4862                 .write_u64 = mem_cgroup_swappiness_write,
4863         },
4864         {
4865                 .name = "move_charge_at_immigrate",
4866                 .read_u64 = mem_cgroup_move_charge_read,
4867                 .write_u64 = mem_cgroup_move_charge_write,
4868         },
4869         {
4870                 .name = "oom_control",
4871                 .read_map = mem_cgroup_oom_control_read,
4872                 .write_u64 = mem_cgroup_oom_control_write,
4873                 .register_event = mem_cgroup_oom_register_event,
4874                 .unregister_event = mem_cgroup_oom_unregister_event,
4875                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4876         },
4877 #ifdef CONFIG_NUMA
4878         {
4879                 .name = "numa_stat",
4880                 .open = mem_control_numa_stat_open,
4881                 .mode = S_IRUGO,
4882         },
4883 #endif
4884         {
4885                 .name = "vmscan_stat",
4886                 .read_map = mem_cgroup_vmscan_stat_read,
4887                 .trigger = mem_cgroup_reset_vmscan_stat,
4888         },
4889 };
4890
4891 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4892 static struct cftype memsw_cgroup_files[] = {
4893         {
4894                 .name = "memsw.usage_in_bytes",
4895                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4896                 .read_u64 = mem_cgroup_read,
4897                 .register_event = mem_cgroup_usage_register_event,
4898                 .unregister_event = mem_cgroup_usage_unregister_event,
4899         },
4900         {
4901                 .name = "memsw.max_usage_in_bytes",
4902                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4903                 .trigger = mem_cgroup_reset,
4904                 .read_u64 = mem_cgroup_read,
4905         },
4906         {
4907                 .name = "memsw.limit_in_bytes",
4908                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4909                 .write_string = mem_cgroup_write,
4910                 .read_u64 = mem_cgroup_read,
4911         },
4912         {
4913                 .name = "memsw.failcnt",
4914                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4915                 .trigger = mem_cgroup_reset,
4916                 .read_u64 = mem_cgroup_read,
4917         },
4918 };
4919
4920 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4921 {
4922         if (!do_swap_account)
4923                 return 0;
4924         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4925                                 ARRAY_SIZE(memsw_cgroup_files));
4926 };
4927 #else
4928 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4929 {
4930         return 0;
4931 }
4932 #endif
4933
4934 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4935 {
4936         struct mem_cgroup_per_node *pn;
4937         struct mem_cgroup_per_zone *mz;
4938         enum lru_list l;
4939         int zone, tmp = node;
4940         /*
4941          * This routine is called against possible nodes.
4942          * But it's BUG to call kmalloc() against offline node.
4943          *
4944          * TODO: this routine can waste much memory for nodes which will
4945          *       never be onlined. It's better to use memory hotplug callback
4946          *       function.
4947          */
4948         if (!node_state(node, N_NORMAL_MEMORY))
4949                 tmp = -1;
4950         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4951         if (!pn)
4952                 return 1;
4953
4954         mem->info.nodeinfo[node] = pn;
4955         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4956                 mz = &pn->zoneinfo[zone];
4957                 for_each_lru(l)
4958                         INIT_LIST_HEAD(&mz->lists[l]);
4959                 mz->usage_in_excess = 0;
4960                 mz->on_tree = false;
4961                 mz->mem = mem;
4962         }
4963         return 0;
4964 }
4965
4966 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4967 {
4968         kfree(mem->info.nodeinfo[node]);
4969 }
4970
4971 static struct mem_cgroup *mem_cgroup_alloc(void)
4972 {
4973         struct mem_cgroup *mem;
4974         int size = sizeof(struct mem_cgroup);
4975
4976         /* Can be very big if MAX_NUMNODES is very big */
4977         if (size < PAGE_SIZE)
4978                 mem = kzalloc(size, GFP_KERNEL);
4979         else
4980                 mem = vzalloc(size);
4981
4982         if (!mem)
4983                 return NULL;
4984
4985         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4986         if (!mem->stat)
4987                 goto out_free;
4988         spin_lock_init(&mem->pcp_counter_lock);
4989         return mem;
4990
4991 out_free:
4992         if (size < PAGE_SIZE)
4993                 kfree(mem);
4994         else
4995                 vfree(mem);
4996         return NULL;
4997 }
4998
4999 /*
5000  * At destroying mem_cgroup, references from swap_cgroup can remain.
5001  * (scanning all at force_empty is too costly...)
5002  *
5003  * Instead of clearing all references at force_empty, we remember
5004  * the number of reference from swap_cgroup and free mem_cgroup when
5005  * it goes down to 0.
5006  *
5007  * Removal of cgroup itself succeeds regardless of refs from swap.
5008  */
5009
5010 static void __mem_cgroup_free(struct mem_cgroup *mem)
5011 {
5012         int node;
5013
5014         mem_cgroup_remove_from_trees(mem);
5015         free_css_id(&mem_cgroup_subsys, &mem->css);
5016
5017         for_each_node_state(node, N_POSSIBLE)
5018                 free_mem_cgroup_per_zone_info(mem, node);
5019
5020         free_percpu(mem->stat);
5021         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
5022                 kfree(mem);
5023         else
5024                 vfree(mem);
5025 }
5026
5027 static void mem_cgroup_get(struct mem_cgroup *mem)
5028 {
5029         atomic_inc(&mem->refcnt);
5030 }
5031
5032 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
5033 {
5034         if (atomic_sub_and_test(count, &mem->refcnt)) {
5035                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
5036                 __mem_cgroup_free(mem);
5037                 if (parent)
5038                         mem_cgroup_put(parent);
5039         }
5040 }
5041
5042 static void mem_cgroup_put(struct mem_cgroup *mem)
5043 {
5044         __mem_cgroup_put(mem, 1);
5045 }
5046
5047 /*
5048  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5049  */
5050 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5051 {
5052         if (!mem->res.parent)
5053                 return NULL;
5054         return mem_cgroup_from_res_counter(mem->res.parent, res);
5055 }
5056
5057 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5058 static void __init enable_swap_cgroup(void)
5059 {
5060         if (!mem_cgroup_disabled() && really_do_swap_account)
5061                 do_swap_account = 1;
5062 }
5063 #else
5064 static void __init enable_swap_cgroup(void)
5065 {
5066 }
5067 #endif
5068
5069 static int mem_cgroup_soft_limit_tree_init(void)
5070 {
5071         struct mem_cgroup_tree_per_node *rtpn;
5072         struct mem_cgroup_tree_per_zone *rtpz;
5073         int tmp, node, zone;
5074
5075         for_each_node_state(node, N_POSSIBLE) {
5076                 tmp = node;
5077                 if (!node_state(node, N_NORMAL_MEMORY))
5078                         tmp = -1;
5079                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5080                 if (!rtpn)
5081                         return 1;
5082
5083                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5084
5085                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5086                         rtpz = &rtpn->rb_tree_per_zone[zone];
5087                         rtpz->rb_root = RB_ROOT;
5088                         spin_lock_init(&rtpz->lock);
5089                 }
5090         }
5091         return 0;
5092 }
5093
5094 static struct cgroup_subsys_state * __ref
5095 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5096 {
5097         struct mem_cgroup *mem, *parent;
5098         long error = -ENOMEM;
5099         int node;
5100
5101         mem = mem_cgroup_alloc();
5102         if (!mem)
5103                 return ERR_PTR(error);
5104
5105         for_each_node_state(node, N_POSSIBLE)
5106                 if (alloc_mem_cgroup_per_zone_info(mem, node))
5107                         goto free_out;
5108
5109         /* root ? */
5110         if (cont->parent == NULL) {
5111                 int cpu;
5112                 enable_swap_cgroup();
5113                 parent = NULL;
5114                 root_mem_cgroup = mem;
5115                 if (mem_cgroup_soft_limit_tree_init())
5116                         goto free_out;
5117                 for_each_possible_cpu(cpu) {
5118                         struct memcg_stock_pcp *stock =
5119                                                 &per_cpu(memcg_stock, cpu);
5120                         INIT_WORK(&stock->work, drain_local_stock);
5121                 }
5122                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5123         } else {
5124                 parent = mem_cgroup_from_cont(cont->parent);
5125                 mem->use_hierarchy = parent->use_hierarchy;
5126                 mem->oom_kill_disable = parent->oom_kill_disable;
5127         }
5128
5129         if (parent && parent->use_hierarchy) {
5130                 res_counter_init(&mem->res, &parent->res);
5131                 res_counter_init(&mem->memsw, &parent->memsw);
5132                 /*
5133                  * We increment refcnt of the parent to ensure that we can
5134                  * safely access it on res_counter_charge/uncharge.
5135                  * This refcnt will be decremented when freeing this
5136                  * mem_cgroup(see mem_cgroup_put).
5137                  */
5138                 mem_cgroup_get(parent);
5139         } else {
5140                 res_counter_init(&mem->res, NULL);
5141                 res_counter_init(&mem->memsw, NULL);
5142         }
5143         mem->last_scanned_child = 0;
5144         mem->last_scanned_node = MAX_NUMNODES;
5145         INIT_LIST_HEAD(&mem->oom_notify);
5146
5147         if (parent)
5148                 mem->swappiness = mem_cgroup_swappiness(parent);
5149         atomic_set(&mem->refcnt, 1);
5150         mem->move_charge_at_immigrate = 0;
5151         mutex_init(&mem->thresholds_lock);
5152         spin_lock_init(&mem->scanstat.lock);
5153         return &mem->css;
5154 free_out:
5155         __mem_cgroup_free(mem);
5156         root_mem_cgroup = NULL;
5157         return ERR_PTR(error);
5158 }
5159
5160 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5161                                         struct cgroup *cont)
5162 {
5163         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5164
5165         return mem_cgroup_force_empty(mem, false);
5166 }
5167
5168 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5169                                 struct cgroup *cont)
5170 {
5171         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5172
5173         mem_cgroup_put(mem);
5174 }
5175
5176 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5177                                 struct cgroup *cont)
5178 {
5179         int ret;
5180
5181         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5182                                 ARRAY_SIZE(mem_cgroup_files));
5183
5184         if (!ret)
5185                 ret = register_memsw_files(cont, ss);
5186         return ret;
5187 }
5188
5189 #ifdef CONFIG_MMU
5190 /* Handlers for move charge at task migration. */
5191 #define PRECHARGE_COUNT_AT_ONCE 256
5192 static int mem_cgroup_do_precharge(unsigned long count)
5193 {
5194         int ret = 0;
5195         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5196         struct mem_cgroup *mem = mc.to;
5197
5198         if (mem_cgroup_is_root(mem)) {
5199                 mc.precharge += count;
5200                 /* we don't need css_get for root */
5201                 return ret;
5202         }
5203         /* try to charge at once */
5204         if (count > 1) {
5205                 struct res_counter *dummy;
5206                 /*
5207                  * "mem" cannot be under rmdir() because we've already checked
5208                  * by cgroup_lock_live_cgroup() that it is not removed and we
5209                  * are still under the same cgroup_mutex. So we can postpone
5210                  * css_get().
5211                  */
5212                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5213                         goto one_by_one;
5214                 if (do_swap_account && res_counter_charge(&mem->memsw,
5215                                                 PAGE_SIZE * count, &dummy)) {
5216                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5217                         goto one_by_one;
5218                 }
5219                 mc.precharge += count;
5220                 return ret;
5221         }
5222 one_by_one:
5223         /* fall back to one by one charge */
5224         while (count--) {
5225                 if (signal_pending(current)) {
5226                         ret = -EINTR;
5227                         break;
5228                 }
5229                 if (!batch_count--) {
5230                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5231                         cond_resched();
5232                 }
5233                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5234                 if (ret || !mem)
5235                         /* mem_cgroup_clear_mc() will do uncharge later */
5236                         return -ENOMEM;
5237                 mc.precharge++;
5238         }
5239         return ret;
5240 }
5241
5242 /**
5243  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5244  * @vma: the vma the pte to be checked belongs
5245  * @addr: the address corresponding to the pte to be checked
5246  * @ptent: the pte to be checked
5247  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5248  *
5249  * Returns
5250  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5251  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5252  *     move charge. if @target is not NULL, the page is stored in target->page
5253  *     with extra refcnt got(Callers should handle it).
5254  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5255  *     target for charge migration. if @target is not NULL, the entry is stored
5256  *     in target->ent.
5257  *
5258  * Called with pte lock held.
5259  */
5260 union mc_target {
5261         struct page     *page;
5262         swp_entry_t     ent;
5263 };
5264
5265 enum mc_target_type {
5266         MC_TARGET_NONE, /* not used */
5267         MC_TARGET_PAGE,
5268         MC_TARGET_SWAP,
5269 };
5270
5271 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5272                                                 unsigned long addr, pte_t ptent)
5273 {
5274         struct page *page = vm_normal_page(vma, addr, ptent);
5275
5276         if (!page || !page_mapped(page))
5277                 return NULL;
5278         if (PageAnon(page)) {
5279                 /* we don't move shared anon */
5280                 if (!move_anon() || page_mapcount(page) > 2)
5281                         return NULL;
5282         } else if (!move_file())
5283                 /* we ignore mapcount for file pages */
5284                 return NULL;
5285         if (!get_page_unless_zero(page))
5286                 return NULL;
5287
5288         return page;
5289 }
5290
5291 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5292                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5293 {
5294         int usage_count;
5295         struct page *page = NULL;
5296         swp_entry_t ent = pte_to_swp_entry(ptent);
5297
5298         if (!move_anon() || non_swap_entry(ent))
5299                 return NULL;
5300         usage_count = mem_cgroup_count_swap_user(ent, &page);
5301         if (usage_count > 1) { /* we don't move shared anon */
5302                 if (page)
5303                         put_page(page);
5304                 return NULL;
5305         }
5306         if (do_swap_account)
5307                 entry->val = ent.val;
5308
5309         return page;
5310 }
5311
5312 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5313                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5314 {
5315         struct page *page = NULL;
5316         struct inode *inode;
5317         struct address_space *mapping;
5318         pgoff_t pgoff;
5319
5320         if (!vma->vm_file) /* anonymous vma */
5321                 return NULL;
5322         if (!move_file())
5323                 return NULL;
5324
5325         inode = vma->vm_file->f_path.dentry->d_inode;
5326         mapping = vma->vm_file->f_mapping;
5327         if (pte_none(ptent))
5328                 pgoff = linear_page_index(vma, addr);
5329         else /* pte_file(ptent) is true */
5330                 pgoff = pte_to_pgoff(ptent);
5331
5332         /* page is moved even if it's not RSS of this task(page-faulted). */
5333         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5334                 page = find_get_page(mapping, pgoff);
5335         } else { /* shmem/tmpfs file. we should take account of swap too. */
5336                 swp_entry_t ent;
5337                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5338                 if (do_swap_account)
5339                         entry->val = ent.val;
5340         }
5341
5342         return page;
5343 }
5344
5345 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5346                 unsigned long addr, pte_t ptent, union mc_target *target)
5347 {
5348         struct page *page = NULL;
5349         struct page_cgroup *pc;
5350         int ret = 0;
5351         swp_entry_t ent = { .val = 0 };
5352
5353         if (pte_present(ptent))
5354                 page = mc_handle_present_pte(vma, addr, ptent);
5355         else if (is_swap_pte(ptent))
5356                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5357         else if (pte_none(ptent) || pte_file(ptent))
5358                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5359
5360         if (!page && !ent.val)
5361                 return 0;
5362         if (page) {
5363                 pc = lookup_page_cgroup(page);
5364                 /*
5365                  * Do only loose check w/o page_cgroup lock.
5366                  * mem_cgroup_move_account() checks the pc is valid or not under
5367                  * the lock.
5368                  */
5369                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5370                         ret = MC_TARGET_PAGE;
5371                         if (target)
5372                                 target->page = page;
5373                 }
5374                 if (!ret || !target)
5375                         put_page(page);
5376         }
5377         /* There is a swap entry and a page doesn't exist or isn't charged */
5378         if (ent.val && !ret &&
5379                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5380                 ret = MC_TARGET_SWAP;
5381                 if (target)
5382                         target->ent = ent;
5383         }
5384         return ret;
5385 }
5386
5387 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5388                                         unsigned long addr, unsigned long end,
5389                                         struct mm_walk *walk)
5390 {
5391         struct vm_area_struct *vma = walk->private;
5392         pte_t *pte;
5393         spinlock_t *ptl;
5394
5395         split_huge_page_pmd(walk->mm, pmd);
5396
5397         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5398         for (; addr != end; pte++, addr += PAGE_SIZE)
5399                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5400                         mc.precharge++; /* increment precharge temporarily */
5401         pte_unmap_unlock(pte - 1, ptl);
5402         cond_resched();
5403
5404         return 0;
5405 }
5406
5407 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5408 {
5409         unsigned long precharge;
5410         struct vm_area_struct *vma;
5411
5412         down_read(&mm->mmap_sem);
5413         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5414                 struct mm_walk mem_cgroup_count_precharge_walk = {
5415                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5416                         .mm = mm,
5417                         .private = vma,
5418                 };
5419                 if (is_vm_hugetlb_page(vma))
5420                         continue;
5421                 walk_page_range(vma->vm_start, vma->vm_end,
5422                                         &mem_cgroup_count_precharge_walk);
5423         }
5424         up_read(&mm->mmap_sem);
5425
5426         precharge = mc.precharge;
5427         mc.precharge = 0;
5428
5429         return precharge;
5430 }
5431
5432 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5433 {
5434         unsigned long precharge = mem_cgroup_count_precharge(mm);
5435
5436         VM_BUG_ON(mc.moving_task);
5437         mc.moving_task = current;
5438         return mem_cgroup_do_precharge(precharge);
5439 }
5440
5441 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5442 static void __mem_cgroup_clear_mc(void)
5443 {
5444         struct mem_cgroup *from = mc.from;
5445         struct mem_cgroup *to = mc.to;
5446
5447         /* we must uncharge all the leftover precharges from mc.to */
5448         if (mc.precharge) {
5449                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5450                 mc.precharge = 0;
5451         }
5452         /*
5453          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5454          * we must uncharge here.
5455          */
5456         if (mc.moved_charge) {
5457                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5458                 mc.moved_charge = 0;
5459         }
5460         /* we must fixup refcnts and charges */
5461         if (mc.moved_swap) {
5462                 /* uncharge swap account from the old cgroup */
5463                 if (!mem_cgroup_is_root(mc.from))
5464                         res_counter_uncharge(&mc.from->memsw,
5465                                                 PAGE_SIZE * mc.moved_swap);
5466                 __mem_cgroup_put(mc.from, mc.moved_swap);
5467
5468                 if (!mem_cgroup_is_root(mc.to)) {
5469                         /*
5470                          * we charged both to->res and to->memsw, so we should
5471                          * uncharge to->res.
5472                          */
5473                         res_counter_uncharge(&mc.to->res,
5474                                                 PAGE_SIZE * mc.moved_swap);
5475                 }
5476                 /* we've already done mem_cgroup_get(mc.to) */
5477                 mc.moved_swap = 0;
5478         }
5479         memcg_oom_recover(from);
5480         memcg_oom_recover(to);
5481         wake_up_all(&mc.waitq);
5482 }
5483
5484 static void mem_cgroup_clear_mc(void)
5485 {
5486         struct mem_cgroup *from = mc.from;
5487
5488         /*
5489          * we must clear moving_task before waking up waiters at the end of
5490          * task migration.
5491          */
5492         mc.moving_task = NULL;
5493         __mem_cgroup_clear_mc();
5494         spin_lock(&mc.lock);
5495         mc.from = NULL;
5496         mc.to = NULL;
5497         spin_unlock(&mc.lock);
5498         mem_cgroup_end_move(from);
5499 }
5500
5501 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5502                                 struct cgroup *cgroup,
5503                                 struct task_struct *p)
5504 {
5505         int ret = 0;
5506         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5507
5508         if (mem->move_charge_at_immigrate) {
5509                 struct mm_struct *mm;
5510                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5511
5512                 VM_BUG_ON(from == mem);
5513
5514                 mm = get_task_mm(p);
5515                 if (!mm)
5516                         return 0;
5517                 /* We move charges only when we move a owner of the mm */
5518                 if (mm->owner == p) {
5519                         VM_BUG_ON(mc.from);
5520                         VM_BUG_ON(mc.to);
5521                         VM_BUG_ON(mc.precharge);
5522                         VM_BUG_ON(mc.moved_charge);
5523                         VM_BUG_ON(mc.moved_swap);
5524                         mem_cgroup_start_move(from);
5525                         spin_lock(&mc.lock);
5526                         mc.from = from;
5527                         mc.to = mem;
5528                         spin_unlock(&mc.lock);
5529                         /* We set mc.moving_task later */
5530
5531                         ret = mem_cgroup_precharge_mc(mm);
5532                         if (ret)
5533                                 mem_cgroup_clear_mc();
5534                 }
5535                 mmput(mm);
5536         }
5537         return ret;
5538 }
5539
5540 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5541                                 struct cgroup *cgroup,
5542                                 struct task_struct *p)
5543 {
5544         mem_cgroup_clear_mc();
5545 }
5546
5547 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5548                                 unsigned long addr, unsigned long end,
5549                                 struct mm_walk *walk)
5550 {
5551         int ret = 0;
5552         struct vm_area_struct *vma = walk->private;
5553         pte_t *pte;
5554         spinlock_t *ptl;
5555
5556         split_huge_page_pmd(walk->mm, pmd);
5557 retry:
5558         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5559         for (; addr != end; addr += PAGE_SIZE) {
5560                 pte_t ptent = *(pte++);
5561                 union mc_target target;
5562                 int type;
5563                 struct page *page;
5564                 struct page_cgroup *pc;
5565                 swp_entry_t ent;
5566
5567                 if (!mc.precharge)
5568                         break;
5569
5570                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5571                 switch (type) {
5572                 case MC_TARGET_PAGE:
5573                         page = target.page;
5574                         if (isolate_lru_page(page))
5575                                 goto put;
5576                         pc = lookup_page_cgroup(page);
5577                         if (!mem_cgroup_move_account(page, 1, pc,
5578                                                      mc.from, mc.to, false)) {
5579                                 mc.precharge--;
5580                                 /* we uncharge from mc.from later. */
5581                                 mc.moved_charge++;
5582                         }
5583                         putback_lru_page(page);
5584 put:                    /* is_target_pte_for_mc() gets the page */
5585                         put_page(page);
5586                         break;
5587                 case MC_TARGET_SWAP:
5588                         ent = target.ent;
5589                         if (!mem_cgroup_move_swap_account(ent,
5590                                                 mc.from, mc.to, false)) {
5591                                 mc.precharge--;
5592                                 /* we fixup refcnts and charges later. */
5593                                 mc.moved_swap++;
5594                         }
5595                         break;
5596                 default:
5597                         break;
5598                 }
5599         }
5600         pte_unmap_unlock(pte - 1, ptl);
5601         cond_resched();
5602
5603         if (addr != end) {
5604                 /*
5605                  * We have consumed all precharges we got in can_attach().
5606                  * We try charge one by one, but don't do any additional
5607                  * charges to mc.to if we have failed in charge once in attach()
5608                  * phase.
5609                  */
5610                 ret = mem_cgroup_do_precharge(1);
5611                 if (!ret)
5612                         goto retry;
5613         }
5614
5615         return ret;
5616 }
5617
5618 static void mem_cgroup_move_charge(struct mm_struct *mm)
5619 {
5620         struct vm_area_struct *vma;
5621
5622         lru_add_drain_all();
5623 retry:
5624         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5625                 /*
5626                  * Someone who are holding the mmap_sem might be waiting in
5627                  * waitq. So we cancel all extra charges, wake up all waiters,
5628                  * and retry. Because we cancel precharges, we might not be able
5629                  * to move enough charges, but moving charge is a best-effort
5630                  * feature anyway, so it wouldn't be a big problem.
5631                  */
5632                 __mem_cgroup_clear_mc();
5633                 cond_resched();
5634                 goto retry;
5635         }
5636         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5637                 int ret;
5638                 struct mm_walk mem_cgroup_move_charge_walk = {
5639                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5640                         .mm = mm,
5641                         .private = vma,
5642                 };
5643                 if (is_vm_hugetlb_page(vma))
5644                         continue;
5645                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5646                                                 &mem_cgroup_move_charge_walk);
5647                 if (ret)
5648                         /*
5649                          * means we have consumed all precharges and failed in
5650                          * doing additional charge. Just abandon here.
5651                          */
5652                         break;
5653         }
5654         up_read(&mm->mmap_sem);
5655 }
5656
5657 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5658                                 struct cgroup *cont,
5659                                 struct cgroup *old_cont,
5660                                 struct task_struct *p)
5661 {
5662         struct mm_struct *mm = get_task_mm(p);
5663
5664         if (mm) {
5665                 if (mc.to)
5666                         mem_cgroup_move_charge(mm);
5667                 put_swap_token(mm);
5668                 mmput(mm);
5669         }
5670         if (mc.to)
5671                 mem_cgroup_clear_mc();
5672 }
5673 #else   /* !CONFIG_MMU */
5674 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5675                                 struct cgroup *cgroup,
5676                                 struct task_struct *p)
5677 {
5678         return 0;
5679 }
5680 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5681                                 struct cgroup *cgroup,
5682                                 struct task_struct *p)
5683 {
5684 }
5685 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5686                                 struct cgroup *cont,
5687                                 struct cgroup *old_cont,
5688                                 struct task_struct *p)
5689 {
5690 }
5691 #endif
5692
5693 struct cgroup_subsys mem_cgroup_subsys = {
5694         .name = "memory",
5695         .subsys_id = mem_cgroup_subsys_id,
5696         .create = mem_cgroup_create,
5697         .pre_destroy = mem_cgroup_pre_destroy,
5698         .destroy = mem_cgroup_destroy,
5699         .populate = mem_cgroup_populate,
5700         .can_attach = mem_cgroup_can_attach,
5701         .cancel_attach = mem_cgroup_cancel_attach,
5702         .attach = mem_cgroup_move_task,
5703         .early_init = 0,
5704         .use_id = 1,
5705 };
5706
5707 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5708 static int __init enable_swap_account(char *s)
5709 {
5710         /* consider enabled if no parameter or 1 is given */
5711         if (!strcmp(s, "1"))
5712                 really_do_swap_account = 1;
5713         else if (!strcmp(s, "0"))
5714                 really_do_swap_account = 0;
5715         return 1;
5716 }
5717 __setup("swapaccount=", enable_swap_account);
5718
5719 #endif