dfeca594fd7a22e54548ed2269f7383f7a3eaf9f
[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 /*
209  * The memory controller data structure. The memory controller controls both
210  * page cache and RSS per cgroup. We would eventually like to provide
211  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
212  * to help the administrator determine what knobs to tune.
213  *
214  * TODO: Add a water mark for the memory controller. Reclaim will begin when
215  * we hit the water mark. May be even add a low water mark, such that
216  * no reclaim occurs from a cgroup at it's low water mark, this is
217  * a feature that will be implemented much later in the future.
218  */
219 struct mem_cgroup {
220         struct cgroup_subsys_state css;
221         /*
222          * the counter to account for memory usage
223          */
224         struct res_counter res;
225         /*
226          * the counter to account for mem+swap usage.
227          */
228         struct res_counter memsw;
229         /*
230          * Per cgroup active and inactive list, similar to the
231          * per zone LRU lists.
232          */
233         struct mem_cgroup_lru_info info;
234         /*
235          * While reclaiming in a hierarchy, we cache the last child we
236          * reclaimed from.
237          */
238         int last_scanned_child;
239         int last_scanned_node;
240 #if MAX_NUMNODES > 1
241         nodemask_t      scan_nodes;
242         atomic_t        numainfo_events;
243         atomic_t        numainfo_updating;
244 #endif
245         /*
246          * Should the accounting and control be hierarchical, per subtree?
247          */
248         bool use_hierarchy;
249
250         bool            oom_lock;
251         atomic_t        under_oom;
252
253         atomic_t        refcnt;
254
255         int     swappiness;
256         /* OOM-Killer disable */
257         int             oom_kill_disable;
258
259         /* set when res.limit == memsw.limit */
260         bool            memsw_is_minimum;
261
262         /* protect arrays of thresholds */
263         struct mutex thresholds_lock;
264
265         /* thresholds for memory usage. RCU-protected */
266         struct mem_cgroup_thresholds thresholds;
267
268         /* thresholds for mem+swap usage. RCU-protected */
269         struct mem_cgroup_thresholds memsw_thresholds;
270
271         /* For oom notifier event fd */
272         struct list_head oom_notify;
273
274         /*
275          * Should we move charges of a task when a task is moved into this
276          * mem_cgroup ? And what type of charges should we move ?
277          */
278         unsigned long   move_charge_at_immigrate;
279         /*
280          * percpu counter.
281          */
282         struct mem_cgroup_stat_cpu *stat;
283         /*
284          * used when a cpu is offlined or other synchronizations
285          * See mem_cgroup_read_stat().
286          */
287         struct mem_cgroup_stat_cpu nocpu_base;
288         spinlock_t pcp_counter_lock;
289 };
290
291 /* Stuffs for move charges at task migration. */
292 /*
293  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
294  * left-shifted bitmap of these types.
295  */
296 enum move_type {
297         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
298         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
299         NR_MOVE_TYPE,
300 };
301
302 /* "mc" and its members are protected by cgroup_mutex */
303 static struct move_charge_struct {
304         spinlock_t        lock; /* for from, to */
305         struct mem_cgroup *from;
306         struct mem_cgroup *to;
307         unsigned long precharge;
308         unsigned long moved_charge;
309         unsigned long moved_swap;
310         struct task_struct *moving_task;        /* a task moving charges */
311         wait_queue_head_t waitq;                /* a waitq for other context */
312 } mc = {
313         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
314         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
315 };
316
317 static bool move_anon(void)
318 {
319         return test_bit(MOVE_CHARGE_TYPE_ANON,
320                                         &mc.to->move_charge_at_immigrate);
321 }
322
323 static bool move_file(void)
324 {
325         return test_bit(MOVE_CHARGE_TYPE_FILE,
326                                         &mc.to->move_charge_at_immigrate);
327 }
328
329 /*
330  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
331  * limit reclaim to prevent infinite loops, if they ever occur.
332  */
333 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
334 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
335
336 enum charge_type {
337         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
338         MEM_CGROUP_CHARGE_TYPE_MAPPED,
339         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
340         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
341         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
342         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
343         NR_CHARGE_TYPE,
344 };
345
346 /* for encoding cft->private value on file */
347 #define _MEM                    (0)
348 #define _MEMSWAP                (1)
349 #define _OOM_TYPE               (2)
350 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
351 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
352 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
353 /* Used for OOM nofiier */
354 #define OOM_CONTROL             (0)
355
356 /*
357  * Reclaim flags for mem_cgroup_hierarchical_reclaim
358  */
359 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
360 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
361 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
362 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
363 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
364 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
365
366 static void mem_cgroup_get(struct mem_cgroup *mem);
367 static void mem_cgroup_put(struct mem_cgroup *mem);
368 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
369 static void drain_all_stock_async(struct mem_cgroup *mem);
370
371 static struct mem_cgroup_per_zone *
372 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
373 {
374         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
375 }
376
377 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
378 {
379         return &mem->css;
380 }
381
382 static struct mem_cgroup_per_zone *
383 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
384 {
385         int nid = page_to_nid(page);
386         int zid = page_zonenum(page);
387
388         return mem_cgroup_zoneinfo(mem, nid, zid);
389 }
390
391 static struct mem_cgroup_tree_per_zone *
392 soft_limit_tree_node_zone(int nid, int zid)
393 {
394         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
395 }
396
397 static struct mem_cgroup_tree_per_zone *
398 soft_limit_tree_from_page(struct page *page)
399 {
400         int nid = page_to_nid(page);
401         int zid = page_zonenum(page);
402
403         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
404 }
405
406 static void
407 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
408                                 struct mem_cgroup_per_zone *mz,
409                                 struct mem_cgroup_tree_per_zone *mctz,
410                                 unsigned long long new_usage_in_excess)
411 {
412         struct rb_node **p = &mctz->rb_root.rb_node;
413         struct rb_node *parent = NULL;
414         struct mem_cgroup_per_zone *mz_node;
415
416         if (mz->on_tree)
417                 return;
418
419         mz->usage_in_excess = new_usage_in_excess;
420         if (!mz->usage_in_excess)
421                 return;
422         while (*p) {
423                 parent = *p;
424                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
425                                         tree_node);
426                 if (mz->usage_in_excess < mz_node->usage_in_excess)
427                         p = &(*p)->rb_left;
428                 /*
429                  * We can't avoid mem cgroups that are over their soft
430                  * limit by the same amount
431                  */
432                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
433                         p = &(*p)->rb_right;
434         }
435         rb_link_node(&mz->tree_node, parent, p);
436         rb_insert_color(&mz->tree_node, &mctz->rb_root);
437         mz->on_tree = true;
438 }
439
440 static void
441 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
442                                 struct mem_cgroup_per_zone *mz,
443                                 struct mem_cgroup_tree_per_zone *mctz)
444 {
445         if (!mz->on_tree)
446                 return;
447         rb_erase(&mz->tree_node, &mctz->rb_root);
448         mz->on_tree = false;
449 }
450
451 static void
452 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
453                                 struct mem_cgroup_per_zone *mz,
454                                 struct mem_cgroup_tree_per_zone *mctz)
455 {
456         spin_lock(&mctz->lock);
457         __mem_cgroup_remove_exceeded(mem, mz, mctz);
458         spin_unlock(&mctz->lock);
459 }
460
461
462 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
463 {
464         unsigned long long excess;
465         struct mem_cgroup_per_zone *mz;
466         struct mem_cgroup_tree_per_zone *mctz;
467         int nid = page_to_nid(page);
468         int zid = page_zonenum(page);
469         mctz = soft_limit_tree_from_page(page);
470
471         /*
472          * Necessary to update all ancestors when hierarchy is used.
473          * because their event counter is not touched.
474          */
475         for (; mem; mem = parent_mem_cgroup(mem)) {
476                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
477                 excess = res_counter_soft_limit_excess(&mem->res);
478                 /*
479                  * We have to update the tree if mz is on RB-tree or
480                  * mem is over its softlimit.
481                  */
482                 if (excess || mz->on_tree) {
483                         spin_lock(&mctz->lock);
484                         /* if on-tree, remove it */
485                         if (mz->on_tree)
486                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
487                         /*
488                          * Insert again. mz->usage_in_excess will be updated.
489                          * If excess is 0, no tree ops.
490                          */
491                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
492                         spin_unlock(&mctz->lock);
493                 }
494         }
495 }
496
497 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
498 {
499         int node, zone;
500         struct mem_cgroup_per_zone *mz;
501         struct mem_cgroup_tree_per_zone *mctz;
502
503         for_each_node_state(node, N_POSSIBLE) {
504                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
505                         mz = mem_cgroup_zoneinfo(mem, node, zone);
506                         mctz = soft_limit_tree_node_zone(node, zone);
507                         mem_cgroup_remove_exceeded(mem, mz, mctz);
508                 }
509         }
510 }
511
512 static struct mem_cgroup_per_zone *
513 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
514 {
515         struct rb_node *rightmost = NULL;
516         struct mem_cgroup_per_zone *mz;
517
518 retry:
519         mz = NULL;
520         rightmost = rb_last(&mctz->rb_root);
521         if (!rightmost)
522                 goto done;              /* Nothing to reclaim from */
523
524         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
525         /*
526          * Remove the node now but someone else can add it back,
527          * we will to add it back at the end of reclaim to its correct
528          * position in the tree.
529          */
530         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
531         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
532                 !css_tryget(&mz->mem->css))
533                 goto retry;
534 done:
535         return mz;
536 }
537
538 static struct mem_cgroup_per_zone *
539 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
540 {
541         struct mem_cgroup_per_zone *mz;
542
543         spin_lock(&mctz->lock);
544         mz = __mem_cgroup_largest_soft_limit_node(mctz);
545         spin_unlock(&mctz->lock);
546         return mz;
547 }
548
549 /*
550  * Implementation Note: reading percpu statistics for memcg.
551  *
552  * Both of vmstat[] and percpu_counter has threshold and do periodic
553  * synchronization to implement "quick" read. There are trade-off between
554  * reading cost and precision of value. Then, we may have a chance to implement
555  * a periodic synchronizion of counter in memcg's counter.
556  *
557  * But this _read() function is used for user interface now. The user accounts
558  * memory usage by memory cgroup and he _always_ requires exact value because
559  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
560  * have to visit all online cpus and make sum. So, for now, unnecessary
561  * synchronization is not implemented. (just implemented for cpu hotplug)
562  *
563  * If there are kernel internal actions which can make use of some not-exact
564  * value, and reading all cpu value can be performance bottleneck in some
565  * common workload, threashold and synchonization as vmstat[] should be
566  * implemented.
567  */
568 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
569                                  enum mem_cgroup_stat_index idx)
570 {
571         long val = 0;
572         int cpu;
573
574         get_online_cpus();
575         for_each_online_cpu(cpu)
576                 val += per_cpu(mem->stat->count[idx], cpu);
577 #ifdef CONFIG_HOTPLUG_CPU
578         spin_lock(&mem->pcp_counter_lock);
579         val += mem->nocpu_base.count[idx];
580         spin_unlock(&mem->pcp_counter_lock);
581 #endif
582         put_online_cpus();
583         return val;
584 }
585
586 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
587                                          bool charge)
588 {
589         int val = (charge) ? 1 : -1;
590         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
591 }
592
593 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
594 {
595         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
596 }
597
598 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
599 {
600         this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
601 }
602
603 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
604                                             enum mem_cgroup_events_index idx)
605 {
606         unsigned long val = 0;
607         int cpu;
608
609         for_each_online_cpu(cpu)
610                 val += per_cpu(mem->stat->events[idx], cpu);
611 #ifdef CONFIG_HOTPLUG_CPU
612         spin_lock(&mem->pcp_counter_lock);
613         val += mem->nocpu_base.events[idx];
614         spin_unlock(&mem->pcp_counter_lock);
615 #endif
616         return val;
617 }
618
619 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
620                                          bool file, int nr_pages)
621 {
622         preempt_disable();
623
624         if (file)
625                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
626         else
627                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
628
629         /* pagein of a big page is an event. So, ignore page size */
630         if (nr_pages > 0)
631                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
632         else {
633                 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
634                 nr_pages = -nr_pages; /* for event */
635         }
636
637         __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
638
639         preempt_enable();
640 }
641
642 unsigned long
643 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
644                         unsigned int lru_mask)
645 {
646         struct mem_cgroup_per_zone *mz;
647         enum lru_list l;
648         unsigned long ret = 0;
649
650         mz = mem_cgroup_zoneinfo(mem, nid, zid);
651
652         for_each_lru(l) {
653                 if (BIT(l) & lru_mask)
654                         ret += MEM_CGROUP_ZSTAT(mz, l);
655         }
656         return ret;
657 }
658
659 static unsigned long
660 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
661                         int nid, unsigned int lru_mask)
662 {
663         u64 total = 0;
664         int zid;
665
666         for (zid = 0; zid < MAX_NR_ZONES; zid++)
667                 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
668
669         return total;
670 }
671
672 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
673                         unsigned int lru_mask)
674 {
675         int nid;
676         u64 total = 0;
677
678         for_each_node_state(nid, N_HIGH_MEMORY)
679                 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
680         return total;
681 }
682
683 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
684 {
685         unsigned long val, next;
686
687         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
688         next = this_cpu_read(mem->stat->targets[target]);
689         /* from time_after() in jiffies.h */
690         return ((long)next - (long)val < 0);
691 }
692
693 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
694 {
695         unsigned long val, next;
696
697         val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
698
699         switch (target) {
700         case MEM_CGROUP_TARGET_THRESH:
701                 next = val + THRESHOLDS_EVENTS_TARGET;
702                 break;
703         case MEM_CGROUP_TARGET_SOFTLIMIT:
704                 next = val + SOFTLIMIT_EVENTS_TARGET;
705                 break;
706         case MEM_CGROUP_TARGET_NUMAINFO:
707                 next = val + NUMAINFO_EVENTS_TARGET;
708                 break;
709         default:
710                 return;
711         }
712
713         this_cpu_write(mem->stat->targets[target], next);
714 }
715
716 /*
717  * Check events in order.
718  *
719  */
720 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
721 {
722         /* threshold event is triggered in finer grain than soft limit */
723         if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
724                 mem_cgroup_threshold(mem);
725                 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
726                 if (unlikely(__memcg_event_check(mem,
727                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
728                         mem_cgroup_update_tree(mem, page);
729                         __mem_cgroup_target_update(mem,
730                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
731                 }
732 #if MAX_NUMNODES > 1
733                 if (unlikely(__memcg_event_check(mem,
734                         MEM_CGROUP_TARGET_NUMAINFO))) {
735                         atomic_inc(&mem->numainfo_events);
736                         __mem_cgroup_target_update(mem,
737                                 MEM_CGROUP_TARGET_NUMAINFO);
738                 }
739 #endif
740         }
741 }
742
743 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
744 {
745         return container_of(cgroup_subsys_state(cont,
746                                 mem_cgroup_subsys_id), struct mem_cgroup,
747                                 css);
748 }
749
750 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
751 {
752         /*
753          * mm_update_next_owner() may clear mm->owner to NULL
754          * if it races with swapoff, page migration, etc.
755          * So this can be called with p == NULL.
756          */
757         if (unlikely(!p))
758                 return NULL;
759
760         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
761                                 struct mem_cgroup, css);
762 }
763
764 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
765 {
766         struct mem_cgroup *mem = NULL;
767
768         if (!mm)
769                 return NULL;
770         /*
771          * Because we have no locks, mm->owner's may be being moved to other
772          * cgroup. We use css_tryget() here even if this looks
773          * pessimistic (rather than adding locks here).
774          */
775         rcu_read_lock();
776         do {
777                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
778                 if (unlikely(!mem))
779                         break;
780         } while (!css_tryget(&mem->css));
781         rcu_read_unlock();
782         return mem;
783 }
784
785 /* The caller has to guarantee "mem" exists before calling this */
786 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
787 {
788         struct cgroup_subsys_state *css;
789         int found;
790
791         if (!mem) /* ROOT cgroup has the smallest ID */
792                 return root_mem_cgroup; /*css_put/get against root is ignored*/
793         if (!mem->use_hierarchy) {
794                 if (css_tryget(&mem->css))
795                         return mem;
796                 return NULL;
797         }
798         rcu_read_lock();
799         /*
800          * searching a memory cgroup which has the smallest ID under given
801          * ROOT cgroup. (ID >= 1)
802          */
803         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
804         if (css && css_tryget(css))
805                 mem = container_of(css, struct mem_cgroup, css);
806         else
807                 mem = NULL;
808         rcu_read_unlock();
809         return mem;
810 }
811
812 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
813                                         struct mem_cgroup *root,
814                                         bool cond)
815 {
816         int nextid = css_id(&iter->css) + 1;
817         int found;
818         int hierarchy_used;
819         struct cgroup_subsys_state *css;
820
821         hierarchy_used = iter->use_hierarchy;
822
823         css_put(&iter->css);
824         /* If no ROOT, walk all, ignore hierarchy */
825         if (!cond || (root && !hierarchy_used))
826                 return NULL;
827
828         if (!root)
829                 root = root_mem_cgroup;
830
831         do {
832                 iter = NULL;
833                 rcu_read_lock();
834
835                 css = css_get_next(&mem_cgroup_subsys, nextid,
836                                 &root->css, &found);
837                 if (css && css_tryget(css))
838                         iter = container_of(css, struct mem_cgroup, css);
839                 rcu_read_unlock();
840                 /* If css is NULL, no more cgroups will be found */
841                 nextid = found + 1;
842         } while (css && !iter);
843
844         return iter;
845 }
846 /*
847  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
848  * be careful that "break" loop is not allowed. We have reference count.
849  * Instead of that modify "cond" to be false and "continue" to exit the loop.
850  */
851 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
852         for (iter = mem_cgroup_start_loop(root);\
853              iter != NULL;\
854              iter = mem_cgroup_get_next(iter, root, cond))
855
856 #define for_each_mem_cgroup_tree(iter, root) \
857         for_each_mem_cgroup_tree_cond(iter, root, true)
858
859 #define for_each_mem_cgroup_all(iter) \
860         for_each_mem_cgroup_tree_cond(iter, NULL, true)
861
862
863 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
864 {
865         return (mem == root_mem_cgroup);
866 }
867
868 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
869 {
870         struct mem_cgroup *mem;
871
872         if (!mm)
873                 return;
874
875         rcu_read_lock();
876         mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
877         if (unlikely(!mem))
878                 goto out;
879
880         switch (idx) {
881         case PGMAJFAULT:
882                 mem_cgroup_pgmajfault(mem, 1);
883                 break;
884         case PGFAULT:
885                 mem_cgroup_pgfault(mem, 1);
886                 break;
887         default:
888                 BUG();
889         }
890 out:
891         rcu_read_unlock();
892 }
893 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
894
895 /*
896  * Following LRU functions are allowed to be used without PCG_LOCK.
897  * Operations are called by routine of global LRU independently from memcg.
898  * What we have to take care of here is validness of pc->mem_cgroup.
899  *
900  * Changes to pc->mem_cgroup happens when
901  * 1. charge
902  * 2. moving account
903  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
904  * It is added to LRU before charge.
905  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
906  * When moving account, the page is not on LRU. It's isolated.
907  */
908
909 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
910 {
911         struct page_cgroup *pc;
912         struct mem_cgroup_per_zone *mz;
913
914         if (mem_cgroup_disabled())
915                 return;
916         pc = lookup_page_cgroup(page);
917         /* can happen while we handle swapcache. */
918         if (!TestClearPageCgroupAcctLRU(pc))
919                 return;
920         VM_BUG_ON(!pc->mem_cgroup);
921         /*
922          * We don't check PCG_USED bit. It's cleared when the "page" is finally
923          * removed from global LRU.
924          */
925         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
926         /* huge page split is done under lru_lock. so, we have no races. */
927         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
928         if (mem_cgroup_is_root(pc->mem_cgroup))
929                 return;
930         VM_BUG_ON(list_empty(&pc->lru));
931         list_del_init(&pc->lru);
932 }
933
934 void mem_cgroup_del_lru(struct page *page)
935 {
936         mem_cgroup_del_lru_list(page, page_lru(page));
937 }
938
939 /*
940  * Writeback is about to end against a page which has been marked for immediate
941  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
942  * inactive list.
943  */
944 void mem_cgroup_rotate_reclaimable_page(struct page *page)
945 {
946         struct mem_cgroup_per_zone *mz;
947         struct page_cgroup *pc;
948         enum lru_list lru = page_lru(page);
949
950         if (mem_cgroup_disabled())
951                 return;
952
953         pc = lookup_page_cgroup(page);
954         /* unused or root page is not rotated. */
955         if (!PageCgroupUsed(pc))
956                 return;
957         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
958         smp_rmb();
959         if (mem_cgroup_is_root(pc->mem_cgroup))
960                 return;
961         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
962         list_move_tail(&pc->lru, &mz->lists[lru]);
963 }
964
965 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
966 {
967         struct mem_cgroup_per_zone *mz;
968         struct page_cgroup *pc;
969
970         if (mem_cgroup_disabled())
971                 return;
972
973         pc = lookup_page_cgroup(page);
974         /* unused or root page is not rotated. */
975         if (!PageCgroupUsed(pc))
976                 return;
977         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
978         smp_rmb();
979         if (mem_cgroup_is_root(pc->mem_cgroup))
980                 return;
981         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
982         list_move(&pc->lru, &mz->lists[lru]);
983 }
984
985 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
986 {
987         struct page_cgroup *pc;
988         struct mem_cgroup_per_zone *mz;
989
990         if (mem_cgroup_disabled())
991                 return;
992         pc = lookup_page_cgroup(page);
993         VM_BUG_ON(PageCgroupAcctLRU(pc));
994         if (!PageCgroupUsed(pc))
995                 return;
996         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
997         smp_rmb();
998         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
999         /* huge page split is done under lru_lock. so, we have no races. */
1000         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1001         SetPageCgroupAcctLRU(pc);
1002         if (mem_cgroup_is_root(pc->mem_cgroup))
1003                 return;
1004         list_add(&pc->lru, &mz->lists[lru]);
1005 }
1006
1007 /*
1008  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1009  * while it's linked to lru because the page may be reused after it's fully
1010  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1011  * It's done under lock_page and expected that zone->lru_lock isnever held.
1012  */
1013 static void mem_cgroup_lru_del_before_commit(struct page *page)
1014 {
1015         unsigned long flags;
1016         struct zone *zone = page_zone(page);
1017         struct page_cgroup *pc = lookup_page_cgroup(page);
1018
1019         /*
1020          * Doing this check without taking ->lru_lock seems wrong but this
1021          * is safe. Because if page_cgroup's USED bit is unset, the page
1022          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1023          * set, the commit after this will fail, anyway.
1024          * This all charge/uncharge is done under some mutual execustion.
1025          * So, we don't need to taking care of changes in USED bit.
1026          */
1027         if (likely(!PageLRU(page)))
1028                 return;
1029
1030         spin_lock_irqsave(&zone->lru_lock, flags);
1031         /*
1032          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1033          * is guarded by lock_page() because the page is SwapCache.
1034          */
1035         if (!PageCgroupUsed(pc))
1036                 mem_cgroup_del_lru_list(page, page_lru(page));
1037         spin_unlock_irqrestore(&zone->lru_lock, flags);
1038 }
1039
1040 static void mem_cgroup_lru_add_after_commit(struct page *page)
1041 {
1042         unsigned long flags;
1043         struct zone *zone = page_zone(page);
1044         struct page_cgroup *pc = lookup_page_cgroup(page);
1045
1046         /* taking care of that the page is added to LRU while we commit it */
1047         if (likely(!PageLRU(page)))
1048                 return;
1049         spin_lock_irqsave(&zone->lru_lock, flags);
1050         /* link when the page is linked to LRU but page_cgroup isn't */
1051         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1052                 mem_cgroup_add_lru_list(page, page_lru(page));
1053         spin_unlock_irqrestore(&zone->lru_lock, flags);
1054 }
1055
1056
1057 void mem_cgroup_move_lists(struct page *page,
1058                            enum lru_list from, enum lru_list to)
1059 {
1060         if (mem_cgroup_disabled())
1061                 return;
1062         mem_cgroup_del_lru_list(page, from);
1063         mem_cgroup_add_lru_list(page, to);
1064 }
1065
1066 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1067 {
1068         int ret;
1069         struct mem_cgroup *curr = NULL;
1070         struct task_struct *p;
1071
1072         p = find_lock_task_mm(task);
1073         if (!p)
1074                 return 0;
1075         curr = try_get_mem_cgroup_from_mm(p->mm);
1076         task_unlock(p);
1077         if (!curr)
1078                 return 0;
1079         /*
1080          * We should check use_hierarchy of "mem" not "curr". Because checking
1081          * use_hierarchy of "curr" here make this function true if hierarchy is
1082          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1083          * hierarchy(even if use_hierarchy is disabled in "mem").
1084          */
1085         if (mem->use_hierarchy)
1086                 ret = css_is_ancestor(&curr->css, &mem->css);
1087         else
1088                 ret = (curr == mem);
1089         css_put(&curr->css);
1090         return ret;
1091 }
1092
1093 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1094 {
1095         unsigned long active;
1096         unsigned long inactive;
1097         unsigned long gb;
1098         unsigned long inactive_ratio;
1099
1100         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1101         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1102
1103         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1104         if (gb)
1105                 inactive_ratio = int_sqrt(10 * gb);
1106         else
1107                 inactive_ratio = 1;
1108
1109         if (present_pages) {
1110                 present_pages[0] = inactive;
1111                 present_pages[1] = active;
1112         }
1113
1114         return inactive_ratio;
1115 }
1116
1117 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1118 {
1119         unsigned long active;
1120         unsigned long inactive;
1121         unsigned long present_pages[2];
1122         unsigned long inactive_ratio;
1123
1124         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1125
1126         inactive = present_pages[0];
1127         active = present_pages[1];
1128
1129         if (inactive * inactive_ratio < active)
1130                 return 1;
1131
1132         return 0;
1133 }
1134
1135 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1136 {
1137         unsigned long active;
1138         unsigned long inactive;
1139
1140         inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1141         active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1142
1143         return (active > inactive);
1144 }
1145
1146 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1147                                                       struct zone *zone)
1148 {
1149         int nid = zone_to_nid(zone);
1150         int zid = zone_idx(zone);
1151         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1152
1153         return &mz->reclaim_stat;
1154 }
1155
1156 struct zone_reclaim_stat *
1157 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1158 {
1159         struct page_cgroup *pc;
1160         struct mem_cgroup_per_zone *mz;
1161
1162         if (mem_cgroup_disabled())
1163                 return NULL;
1164
1165         pc = lookup_page_cgroup(page);
1166         if (!PageCgroupUsed(pc))
1167                 return NULL;
1168         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1169         smp_rmb();
1170         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1171         return &mz->reclaim_stat;
1172 }
1173
1174 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1175                                         struct list_head *dst,
1176                                         unsigned long *scanned, int order,
1177                                         int mode, struct zone *z,
1178                                         struct mem_cgroup *mem_cont,
1179                                         int active, int file)
1180 {
1181         unsigned long nr_taken = 0;
1182         struct page *page;
1183         unsigned long scan;
1184         LIST_HEAD(pc_list);
1185         struct list_head *src;
1186         struct page_cgroup *pc, *tmp;
1187         int nid = zone_to_nid(z);
1188         int zid = zone_idx(z);
1189         struct mem_cgroup_per_zone *mz;
1190         int lru = LRU_FILE * file + active;
1191         int ret;
1192
1193         BUG_ON(!mem_cont);
1194         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1195         src = &mz->lists[lru];
1196
1197         scan = 0;
1198         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1199                 if (scan >= nr_to_scan)
1200                         break;
1201
1202                 if (unlikely(!PageCgroupUsed(pc)))
1203                         continue;
1204
1205                 page = lookup_cgroup_page(pc);
1206
1207                 if (unlikely(!PageLRU(page)))
1208                         continue;
1209
1210                 scan++;
1211                 ret = __isolate_lru_page(page, mode, file);
1212                 switch (ret) {
1213                 case 0:
1214                         list_move(&page->lru, dst);
1215                         mem_cgroup_del_lru(page);
1216                         nr_taken += hpage_nr_pages(page);
1217                         break;
1218                 case -EBUSY:
1219                         /* we don't affect global LRU but rotate in our LRU */
1220                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1221                         break;
1222                 default:
1223                         break;
1224                 }
1225         }
1226
1227         *scanned = scan;
1228
1229         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1230                                       0, 0, 0, mode);
1231
1232         return nr_taken;
1233 }
1234
1235 #define mem_cgroup_from_res_counter(counter, member)    \
1236         container_of(counter, struct mem_cgroup, member)
1237
1238 /**
1239  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1240  * @mem: the memory cgroup
1241  *
1242  * Returns the maximum amount of memory @mem can be charged with, in
1243  * pages.
1244  */
1245 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1246 {
1247         unsigned long long margin;
1248
1249         margin = res_counter_margin(&mem->res);
1250         if (do_swap_account)
1251                 margin = min(margin, res_counter_margin(&mem->memsw));
1252         return margin >> PAGE_SHIFT;
1253 }
1254
1255 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1256 {
1257         struct cgroup *cgrp = memcg->css.cgroup;
1258
1259         /* root ? */
1260         if (cgrp->parent == NULL)
1261                 return vm_swappiness;
1262
1263         return memcg->swappiness;
1264 }
1265
1266 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1267 {
1268         int cpu;
1269
1270         get_online_cpus();
1271         spin_lock(&mem->pcp_counter_lock);
1272         for_each_online_cpu(cpu)
1273                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1274         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1275         spin_unlock(&mem->pcp_counter_lock);
1276         put_online_cpus();
1277
1278         synchronize_rcu();
1279 }
1280
1281 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1282 {
1283         int cpu;
1284
1285         if (!mem)
1286                 return;
1287         get_online_cpus();
1288         spin_lock(&mem->pcp_counter_lock);
1289         for_each_online_cpu(cpu)
1290                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1291         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1292         spin_unlock(&mem->pcp_counter_lock);
1293         put_online_cpus();
1294 }
1295 /*
1296  * 2 routines for checking "mem" is under move_account() or not.
1297  *
1298  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1299  *                        for avoiding race in accounting. If true,
1300  *                        pc->mem_cgroup may be overwritten.
1301  *
1302  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1303  *                        under hierarchy of moving cgroups. This is for
1304  *                        waiting at hith-memory prressure caused by "move".
1305  */
1306
1307 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1308 {
1309         VM_BUG_ON(!rcu_read_lock_held());
1310         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1311 }
1312
1313 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1314 {
1315         struct mem_cgroup *from;
1316         struct mem_cgroup *to;
1317         bool ret = false;
1318         /*
1319          * Unlike task_move routines, we access mc.to, mc.from not under
1320          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1321          */
1322         spin_lock(&mc.lock);
1323         from = mc.from;
1324         to = mc.to;
1325         if (!from)
1326                 goto unlock;
1327         if (from == mem || to == mem
1328             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1329             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1330                 ret = true;
1331 unlock:
1332         spin_unlock(&mc.lock);
1333         return ret;
1334 }
1335
1336 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1337 {
1338         if (mc.moving_task && current != mc.moving_task) {
1339                 if (mem_cgroup_under_move(mem)) {
1340                         DEFINE_WAIT(wait);
1341                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1342                         /* moving charge context might have finished. */
1343                         if (mc.moving_task)
1344                                 schedule();
1345                         finish_wait(&mc.waitq, &wait);
1346                         return true;
1347                 }
1348         }
1349         return false;
1350 }
1351
1352 /**
1353  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1354  * @memcg: The memory cgroup that went over limit
1355  * @p: Task that is going to be killed
1356  *
1357  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1358  * enabled
1359  */
1360 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1361 {
1362         struct cgroup *task_cgrp;
1363         struct cgroup *mem_cgrp;
1364         /*
1365          * Need a buffer in BSS, can't rely on allocations. The code relies
1366          * on the assumption that OOM is serialized for memory controller.
1367          * If this assumption is broken, revisit this code.
1368          */
1369         static char memcg_name[PATH_MAX];
1370         int ret;
1371
1372         if (!memcg || !p)
1373                 return;
1374
1375
1376         rcu_read_lock();
1377
1378         mem_cgrp = memcg->css.cgroup;
1379         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1380
1381         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1382         if (ret < 0) {
1383                 /*
1384                  * Unfortunately, we are unable to convert to a useful name
1385                  * But we'll still print out the usage information
1386                  */
1387                 rcu_read_unlock();
1388                 goto done;
1389         }
1390         rcu_read_unlock();
1391
1392         printk(KERN_INFO "Task in %s killed", memcg_name);
1393
1394         rcu_read_lock();
1395         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1396         if (ret < 0) {
1397                 rcu_read_unlock();
1398                 goto done;
1399         }
1400         rcu_read_unlock();
1401
1402         /*
1403          * Continues from above, so we don't need an KERN_ level
1404          */
1405         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1406 done:
1407
1408         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1409                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1410                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1411                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1412         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1413                 "failcnt %llu\n",
1414                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1415                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1416                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1417 }
1418
1419 /*
1420  * This function returns the number of memcg under hierarchy tree. Returns
1421  * 1(self count) if no children.
1422  */
1423 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1424 {
1425         int num = 0;
1426         struct mem_cgroup *iter;
1427
1428         for_each_mem_cgroup_tree(iter, mem)
1429                 num++;
1430         return num;
1431 }
1432
1433 /*
1434  * Return the memory (and swap, if configured) limit for a memcg.
1435  */
1436 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1437 {
1438         u64 limit;
1439         u64 memsw;
1440
1441         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1442         limit += total_swap_pages << PAGE_SHIFT;
1443
1444         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1445         /*
1446          * If memsw is finite and limits the amount of swap space available
1447          * to this memcg, return that limit.
1448          */
1449         return min(limit, memsw);
1450 }
1451
1452 /*
1453  * Visit the first child (need not be the first child as per the ordering
1454  * of the cgroup list, since we track last_scanned_child) of @mem and use
1455  * that to reclaim free pages from.
1456  */
1457 static struct mem_cgroup *
1458 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1459 {
1460         struct mem_cgroup *ret = NULL;
1461         struct cgroup_subsys_state *css;
1462         int nextid, found;
1463
1464         if (!root_mem->use_hierarchy) {
1465                 css_get(&root_mem->css);
1466                 ret = root_mem;
1467         }
1468
1469         while (!ret) {
1470                 rcu_read_lock();
1471                 nextid = root_mem->last_scanned_child + 1;
1472                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1473                                    &found);
1474                 if (css && css_tryget(css))
1475                         ret = container_of(css, struct mem_cgroup, css);
1476
1477                 rcu_read_unlock();
1478                 /* Updates scanning parameter */
1479                 if (!css) {
1480                         /* this means start scan from ID:1 */
1481                         root_mem->last_scanned_child = 0;
1482                 } else
1483                         root_mem->last_scanned_child = found;
1484         }
1485
1486         return ret;
1487 }
1488
1489 /**
1490  * test_mem_cgroup_node_reclaimable
1491  * @mem: the target memcg
1492  * @nid: the node ID to be checked.
1493  * @noswap : specify true here if the user wants flle only information.
1494  *
1495  * This function returns whether the specified memcg contains any
1496  * reclaimable pages on a node. Returns true if there are any reclaimable
1497  * pages in the node.
1498  */
1499 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1500                 int nid, bool noswap)
1501 {
1502         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1503                 return true;
1504         if (noswap || !total_swap_pages)
1505                 return false;
1506         if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1507                 return true;
1508         return false;
1509
1510 }
1511 #if MAX_NUMNODES > 1
1512
1513 /*
1514  * Always updating the nodemask is not very good - even if we have an empty
1515  * list or the wrong list here, we can start from some node and traverse all
1516  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1517  *
1518  */
1519 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1520 {
1521         int nid;
1522         /*
1523          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1524          * pagein/pageout changes since the last update.
1525          */
1526         if (!atomic_read(&mem->numainfo_events))
1527                 return;
1528         if (atomic_inc_return(&mem->numainfo_updating) > 1)
1529                 return;
1530
1531         /* make a nodemask where this memcg uses memory from */
1532         mem->scan_nodes = node_states[N_HIGH_MEMORY];
1533
1534         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1535
1536                 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1537                         node_clear(nid, mem->scan_nodes);
1538         }
1539
1540         atomic_set(&mem->numainfo_events, 0);
1541         atomic_set(&mem->numainfo_updating, 0);
1542 }
1543
1544 /*
1545  * Selecting a node where we start reclaim from. Because what we need is just
1546  * reducing usage counter, start from anywhere is O,K. Considering
1547  * memory reclaim from current node, there are pros. and cons.
1548  *
1549  * Freeing memory from current node means freeing memory from a node which
1550  * we'll use or we've used. So, it may make LRU bad. And if several threads
1551  * hit limits, it will see a contention on a node. But freeing from remote
1552  * node means more costs for memory reclaim because of memory latency.
1553  *
1554  * Now, we use round-robin. Better algorithm is welcomed.
1555  */
1556 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1557 {
1558         int node;
1559
1560         mem_cgroup_may_update_nodemask(mem);
1561         node = mem->last_scanned_node;
1562
1563         node = next_node(node, mem->scan_nodes);
1564         if (node == MAX_NUMNODES)
1565                 node = first_node(mem->scan_nodes);
1566         /*
1567          * We call this when we hit limit, not when pages are added to LRU.
1568          * No LRU may hold pages because all pages are UNEVICTABLE or
1569          * memcg is too small and all pages are not on LRU. In that case,
1570          * we use curret node.
1571          */
1572         if (unlikely(node == MAX_NUMNODES))
1573                 node = numa_node_id();
1574
1575         mem->last_scanned_node = node;
1576         return node;
1577 }
1578
1579 /*
1580  * Check all nodes whether it contains reclaimable pages or not.
1581  * For quick scan, we make use of scan_nodes. This will allow us to skip
1582  * unused nodes. But scan_nodes is lazily updated and may not cotain
1583  * enough new information. We need to do double check.
1584  */
1585 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1586 {
1587         int nid;
1588
1589         /*
1590          * quick check...making use of scan_node.
1591          * We can skip unused nodes.
1592          */
1593         if (!nodes_empty(mem->scan_nodes)) {
1594                 for (nid = first_node(mem->scan_nodes);
1595                      nid < MAX_NUMNODES;
1596                      nid = next_node(nid, mem->scan_nodes)) {
1597
1598                         if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1599                                 return true;
1600                 }
1601         }
1602         /*
1603          * Check rest of nodes.
1604          */
1605         for_each_node_state(nid, N_HIGH_MEMORY) {
1606                 if (node_isset(nid, mem->scan_nodes))
1607                         continue;
1608                 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1609                         return true;
1610         }
1611         return false;
1612 }
1613
1614 #else
1615 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1616 {
1617         return 0;
1618 }
1619
1620 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1621 {
1622         return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1623 }
1624 #endif
1625
1626 /*
1627  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1628  * we reclaimed from, so that we don't end up penalizing one child extensively
1629  * based on its position in the children list.
1630  *
1631  * root_mem is the original ancestor that we've been reclaim from.
1632  *
1633  * We give up and return to the caller when we visit root_mem twice.
1634  * (other groups can be removed while we're walking....)
1635  *
1636  * If shrink==true, for avoiding to free too much, this returns immedieately.
1637  */
1638 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1639                                                 struct zone *zone,
1640                                                 gfp_t gfp_mask,
1641                                                 unsigned long reclaim_options,
1642                                                 unsigned long *total_scanned)
1643 {
1644         struct mem_cgroup *victim;
1645         int ret, total = 0;
1646         int loop = 0;
1647         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1648         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1649         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1650         unsigned long excess;
1651         unsigned long nr_scanned;
1652
1653         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1654
1655         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1656         if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1657                 noswap = true;
1658
1659         while (1) {
1660                 victim = mem_cgroup_select_victim(root_mem);
1661                 if (victim == root_mem) {
1662                         loop++;
1663                         /*
1664                          * We are not draining per cpu cached charges during
1665                          * soft limit reclaim  because global reclaim doesn't
1666                          * care about charges. It tries to free some memory and
1667                          * charges will not give any.
1668                          */
1669                         if (!check_soft && loop >= 1)
1670                                 drain_all_stock_async(root_mem);
1671                         if (loop >= 2) {
1672                                 /*
1673                                  * If we have not been able to reclaim
1674                                  * anything, it might because there are
1675                                  * no reclaimable pages under this hierarchy
1676                                  */
1677                                 if (!check_soft || !total) {
1678                                         css_put(&victim->css);
1679                                         break;
1680                                 }
1681                                 /*
1682                                  * We want to do more targeted reclaim.
1683                                  * excess >> 2 is not to excessive so as to
1684                                  * reclaim too much, nor too less that we keep
1685                                  * coming back to reclaim from this cgroup
1686                                  */
1687                                 if (total >= (excess >> 2) ||
1688                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1689                                         css_put(&victim->css);
1690                                         break;
1691                                 }
1692                         }
1693                 }
1694                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1695                         /* this cgroup's local usage == 0 */
1696                         css_put(&victim->css);
1697                         continue;
1698                 }
1699                 /* we use swappiness of local cgroup */
1700                 if (check_soft) {
1701                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1702                                 noswap, zone, &nr_scanned);
1703                         *total_scanned += nr_scanned;
1704                 } else
1705                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1706                                                 noswap);
1707                 css_put(&victim->css);
1708                 /*
1709                  * At shrinking usage, we can't check we should stop here or
1710                  * reclaim more. It's depends on callers. last_scanned_child
1711                  * will work enough for keeping fairness under tree.
1712                  */
1713                 if (shrink)
1714                         return ret;
1715                 total += ret;
1716                 if (check_soft) {
1717                         if (!res_counter_soft_limit_excess(&root_mem->res))
1718                                 return total;
1719                 } else if (mem_cgroup_margin(root_mem))
1720                         return total;
1721         }
1722         return total;
1723 }
1724
1725 /*
1726  * Check OOM-Killer is already running under our hierarchy.
1727  * If someone is running, return false.
1728  * Has to be called with memcg_oom_lock
1729  */
1730 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1731 {
1732         int lock_count = -1;
1733         struct mem_cgroup *iter, *failed = NULL;
1734         bool cond = true;
1735
1736         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1737                 bool locked = iter->oom_lock;
1738
1739                 iter->oom_lock = true;
1740                 if (lock_count == -1)
1741                         lock_count = iter->oom_lock;
1742                 else if (lock_count != locked) {
1743                         /*
1744                          * this subtree of our hierarchy is already locked
1745                          * so we cannot give a lock.
1746                          */
1747                         lock_count = 0;
1748                         failed = iter;
1749                         cond = false;
1750                 }
1751         }
1752
1753         if (!failed)
1754                 goto done;
1755
1756         /*
1757          * OK, we failed to lock the whole subtree so we have to clean up
1758          * what we set up to the failing subtree
1759          */
1760         cond = true;
1761         for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1762                 if (iter == failed) {
1763                         cond = false;
1764                         continue;
1765                 }
1766                 iter->oom_lock = false;
1767         }
1768 done:
1769         return lock_count;
1770 }
1771
1772 /*
1773  * Has to be called with memcg_oom_lock
1774  */
1775 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1776 {
1777         struct mem_cgroup *iter;
1778
1779         for_each_mem_cgroup_tree(iter, mem)
1780                 iter->oom_lock = false;
1781         return 0;
1782 }
1783
1784 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1785 {
1786         struct mem_cgroup *iter;
1787
1788         for_each_mem_cgroup_tree(iter, mem)
1789                 atomic_inc(&iter->under_oom);
1790 }
1791
1792 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1793 {
1794         struct mem_cgroup *iter;
1795
1796         /*
1797          * When a new child is created while the hierarchy is under oom,
1798          * mem_cgroup_oom_lock() may not be called. We have to use
1799          * atomic_add_unless() here.
1800          */
1801         for_each_mem_cgroup_tree(iter, mem)
1802                 atomic_add_unless(&iter->under_oom, -1, 0);
1803 }
1804
1805 static DEFINE_SPINLOCK(memcg_oom_lock);
1806 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1807
1808 struct oom_wait_info {
1809         struct mem_cgroup *mem;
1810         wait_queue_t    wait;
1811 };
1812
1813 static int memcg_oom_wake_function(wait_queue_t *wait,
1814         unsigned mode, int sync, void *arg)
1815 {
1816         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1817         struct oom_wait_info *oom_wait_info;
1818
1819         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1820
1821         if (oom_wait_info->mem == wake_mem)
1822                 goto wakeup;
1823         /* if no hierarchy, no match */
1824         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1825                 return 0;
1826         /*
1827          * Both of oom_wait_info->mem and wake_mem are stable under us.
1828          * Then we can use css_is_ancestor without taking care of RCU.
1829          */
1830         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1831             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1832                 return 0;
1833
1834 wakeup:
1835         return autoremove_wake_function(wait, mode, sync, arg);
1836 }
1837
1838 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1839 {
1840         /* for filtering, pass "mem" as argument. */
1841         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1842 }
1843
1844 static void memcg_oom_recover(struct mem_cgroup *mem)
1845 {
1846         if (mem && atomic_read(&mem->under_oom))
1847                 memcg_wakeup_oom(mem);
1848 }
1849
1850 /*
1851  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1852  */
1853 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1854 {
1855         struct oom_wait_info owait;
1856         bool locked, need_to_kill;
1857
1858         owait.mem = mem;
1859         owait.wait.flags = 0;
1860         owait.wait.func = memcg_oom_wake_function;
1861         owait.wait.private = current;
1862         INIT_LIST_HEAD(&owait.wait.task_list);
1863         need_to_kill = true;
1864         mem_cgroup_mark_under_oom(mem);
1865
1866         /* At first, try to OOM lock hierarchy under mem.*/
1867         spin_lock(&memcg_oom_lock);
1868         locked = mem_cgroup_oom_lock(mem);
1869         /*
1870          * Even if signal_pending(), we can't quit charge() loop without
1871          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1872          * under OOM is always welcomed, use TASK_KILLABLE here.
1873          */
1874         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1875         if (!locked || mem->oom_kill_disable)
1876                 need_to_kill = false;
1877         if (locked)
1878                 mem_cgroup_oom_notify(mem);
1879         spin_unlock(&memcg_oom_lock);
1880
1881         if (need_to_kill) {
1882                 finish_wait(&memcg_oom_waitq, &owait.wait);
1883                 mem_cgroup_out_of_memory(mem, mask);
1884         } else {
1885                 schedule();
1886                 finish_wait(&memcg_oom_waitq, &owait.wait);
1887         }
1888         spin_lock(&memcg_oom_lock);
1889         if (locked)
1890                 mem_cgroup_oom_unlock(mem);
1891         memcg_wakeup_oom(mem);
1892         spin_unlock(&memcg_oom_lock);
1893
1894         mem_cgroup_unmark_under_oom(mem);
1895
1896         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1897                 return false;
1898         /* Give chance to dying process */
1899         schedule_timeout(1);
1900         return true;
1901 }
1902
1903 /*
1904  * Currently used to update mapped file statistics, but the routine can be
1905  * generalized to update other statistics as well.
1906  *
1907  * Notes: Race condition
1908  *
1909  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1910  * it tends to be costly. But considering some conditions, we doesn't need
1911  * to do so _always_.
1912  *
1913  * Considering "charge", lock_page_cgroup() is not required because all
1914  * file-stat operations happen after a page is attached to radix-tree. There
1915  * are no race with "charge".
1916  *
1917  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1918  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1919  * if there are race with "uncharge". Statistics itself is properly handled
1920  * by flags.
1921  *
1922  * Considering "move", this is an only case we see a race. To make the race
1923  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1924  * possibility of race condition. If there is, we take a lock.
1925  */
1926
1927 void mem_cgroup_update_page_stat(struct page *page,
1928                                  enum mem_cgroup_page_stat_item idx, int val)
1929 {
1930         struct mem_cgroup *mem;
1931         struct page_cgroup *pc = lookup_page_cgroup(page);
1932         bool need_unlock = false;
1933         unsigned long uninitialized_var(flags);
1934
1935         if (unlikely(!pc))
1936                 return;
1937
1938         rcu_read_lock();
1939         mem = pc->mem_cgroup;
1940         if (unlikely(!mem || !PageCgroupUsed(pc)))
1941                 goto out;
1942         /* pc->mem_cgroup is unstable ? */
1943         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1944                 /* take a lock against to access pc->mem_cgroup */
1945                 move_lock_page_cgroup(pc, &flags);
1946                 need_unlock = true;
1947                 mem = pc->mem_cgroup;
1948                 if (!mem || !PageCgroupUsed(pc))
1949                         goto out;
1950         }
1951
1952         switch (idx) {
1953         case MEMCG_NR_FILE_MAPPED:
1954                 if (val > 0)
1955                         SetPageCgroupFileMapped(pc);
1956                 else if (!page_mapped(page))
1957                         ClearPageCgroupFileMapped(pc);
1958                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1959                 break;
1960         default:
1961                 BUG();
1962         }
1963
1964         this_cpu_add(mem->stat->count[idx], val);
1965
1966 out:
1967         if (unlikely(need_unlock))
1968                 move_unlock_page_cgroup(pc, &flags);
1969         rcu_read_unlock();
1970         return;
1971 }
1972 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1973
1974 /*
1975  * size of first charge trial. "32" comes from vmscan.c's magic value.
1976  * TODO: maybe necessary to use big numbers in big irons.
1977  */
1978 #define CHARGE_BATCH    32U
1979 struct memcg_stock_pcp {
1980         struct mem_cgroup *cached; /* this never be root cgroup */
1981         unsigned int nr_pages;
1982         struct work_struct work;
1983         unsigned long flags;
1984 #define FLUSHING_CACHED_CHARGE  (0)
1985 };
1986 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1987 static DEFINE_MUTEX(percpu_charge_mutex);
1988
1989 /*
1990  * Try to consume stocked charge on this cpu. If success, one page is consumed
1991  * from local stock and true is returned. If the stock is 0 or charges from a
1992  * cgroup which is not current target, returns false. This stock will be
1993  * refilled.
1994  */
1995 static bool consume_stock(struct mem_cgroup *mem)
1996 {
1997         struct memcg_stock_pcp *stock;
1998         bool ret = true;
1999
2000         stock = &get_cpu_var(memcg_stock);
2001         if (mem == stock->cached && stock->nr_pages)
2002                 stock->nr_pages--;
2003         else /* need to call res_counter_charge */
2004                 ret = false;
2005         put_cpu_var(memcg_stock);
2006         return ret;
2007 }
2008
2009 /*
2010  * Returns stocks cached in percpu to res_counter and reset cached information.
2011  */
2012 static void drain_stock(struct memcg_stock_pcp *stock)
2013 {
2014         struct mem_cgroup *old = stock->cached;
2015
2016         if (stock->nr_pages) {
2017                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2018
2019                 res_counter_uncharge(&old->res, bytes);
2020                 if (do_swap_account)
2021                         res_counter_uncharge(&old->memsw, bytes);
2022                 stock->nr_pages = 0;
2023         }
2024         stock->cached = NULL;
2025 }
2026
2027 /*
2028  * This must be called under preempt disabled or must be called by
2029  * a thread which is pinned to local cpu.
2030  */
2031 static void drain_local_stock(struct work_struct *dummy)
2032 {
2033         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2034         drain_stock(stock);
2035         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2036 }
2037
2038 /*
2039  * Cache charges(val) which is from res_counter, to local per_cpu area.
2040  * This will be consumed by consume_stock() function, later.
2041  */
2042 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2043 {
2044         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2045
2046         if (stock->cached != mem) { /* reset if necessary */
2047                 drain_stock(stock);
2048                 stock->cached = mem;
2049         }
2050         stock->nr_pages += nr_pages;
2051         put_cpu_var(memcg_stock);
2052 }
2053
2054 /*
2055  * Tries to drain stocked charges in other cpus. This function is asynchronous
2056  * and just put a work per cpu for draining localy on each cpu. Caller can
2057  * expects some charges will be back to res_counter later but cannot wait for
2058  * it.
2059  */
2060 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2061 {
2062         int cpu, curcpu;
2063         /*
2064          * If someone calls draining, avoid adding more kworker runs.
2065          */
2066         if (!mutex_trylock(&percpu_charge_mutex))
2067                 return;
2068         /* Notify other cpus that system-wide "drain" is running */
2069         get_online_cpus();
2070         /*
2071          * Get a hint for avoiding draining charges on the current cpu,
2072          * which must be exhausted by our charging.  It is not required that
2073          * this be a precise check, so we use raw_smp_processor_id() instead of
2074          * getcpu()/putcpu().
2075          */
2076         curcpu = raw_smp_processor_id();
2077         for_each_online_cpu(cpu) {
2078                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2079                 struct mem_cgroup *mem;
2080
2081                 if (cpu == curcpu)
2082                         continue;
2083
2084                 mem = stock->cached;
2085                 if (!mem)
2086                         continue;
2087                 if (mem != root_mem) {
2088                         if (!root_mem->use_hierarchy)
2089                                 continue;
2090                         /* check whether "mem" is under tree of "root_mem" */
2091                         if (!css_is_ancestor(&mem->css, &root_mem->css))
2092                                 continue;
2093                 }
2094                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2095                         schedule_work_on(cpu, &stock->work);
2096         }
2097         put_online_cpus();
2098         mutex_unlock(&percpu_charge_mutex);
2099         /* We don't wait for flush_work */
2100 }
2101
2102 /* This is a synchronous drain interface. */
2103 static void drain_all_stock_sync(void)
2104 {
2105         /* called when force_empty is called */
2106         mutex_lock(&percpu_charge_mutex);
2107         schedule_on_each_cpu(drain_local_stock);
2108         mutex_unlock(&percpu_charge_mutex);
2109 }
2110
2111 /*
2112  * This function drains percpu counter value from DEAD cpu and
2113  * move it to local cpu. Note that this function can be preempted.
2114  */
2115 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2116 {
2117         int i;
2118
2119         spin_lock(&mem->pcp_counter_lock);
2120         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2121                 long x = per_cpu(mem->stat->count[i], cpu);
2122
2123                 per_cpu(mem->stat->count[i], cpu) = 0;
2124                 mem->nocpu_base.count[i] += x;
2125         }
2126         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2127                 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2128
2129                 per_cpu(mem->stat->events[i], cpu) = 0;
2130                 mem->nocpu_base.events[i] += x;
2131         }
2132         /* need to clear ON_MOVE value, works as a kind of lock. */
2133         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2134         spin_unlock(&mem->pcp_counter_lock);
2135 }
2136
2137 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2138 {
2139         int idx = MEM_CGROUP_ON_MOVE;
2140
2141         spin_lock(&mem->pcp_counter_lock);
2142         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2143         spin_unlock(&mem->pcp_counter_lock);
2144 }
2145
2146 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2147                                         unsigned long action,
2148                                         void *hcpu)
2149 {
2150         int cpu = (unsigned long)hcpu;
2151         struct memcg_stock_pcp *stock;
2152         struct mem_cgroup *iter;
2153
2154         if ((action == CPU_ONLINE)) {
2155                 for_each_mem_cgroup_all(iter)
2156                         synchronize_mem_cgroup_on_move(iter, cpu);
2157                 return NOTIFY_OK;
2158         }
2159
2160         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2161                 return NOTIFY_OK;
2162
2163         for_each_mem_cgroup_all(iter)
2164                 mem_cgroup_drain_pcp_counter(iter, cpu);
2165
2166         stock = &per_cpu(memcg_stock, cpu);
2167         drain_stock(stock);
2168         return NOTIFY_OK;
2169 }
2170
2171
2172 /* See __mem_cgroup_try_charge() for details */
2173 enum {
2174         CHARGE_OK,              /* success */
2175         CHARGE_RETRY,           /* need to retry but retry is not bad */
2176         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2177         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2178         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2179 };
2180
2181 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2182                                 unsigned int nr_pages, bool oom_check)
2183 {
2184         unsigned long csize = nr_pages * PAGE_SIZE;
2185         struct mem_cgroup *mem_over_limit;
2186         struct res_counter *fail_res;
2187         unsigned long flags = 0;
2188         int ret;
2189
2190         ret = res_counter_charge(&mem->res, csize, &fail_res);
2191
2192         if (likely(!ret)) {
2193                 if (!do_swap_account)
2194                         return CHARGE_OK;
2195                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2196                 if (likely(!ret))
2197                         return CHARGE_OK;
2198
2199                 res_counter_uncharge(&mem->res, csize);
2200                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2201                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2202         } else
2203                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2204         /*
2205          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2206          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2207          *
2208          * Never reclaim on behalf of optional batching, retry with a
2209          * single page instead.
2210          */
2211         if (nr_pages == CHARGE_BATCH)
2212                 return CHARGE_RETRY;
2213
2214         if (!(gfp_mask & __GFP_WAIT))
2215                 return CHARGE_WOULDBLOCK;
2216
2217         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2218                                               gfp_mask, flags, NULL);
2219         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2220                 return CHARGE_RETRY;
2221         /*
2222          * Even though the limit is exceeded at this point, reclaim
2223          * may have been able to free some pages.  Retry the charge
2224          * before killing the task.
2225          *
2226          * Only for regular pages, though: huge pages are rather
2227          * unlikely to succeed so close to the limit, and we fall back
2228          * to regular pages anyway in case of failure.
2229          */
2230         if (nr_pages == 1 && ret)
2231                 return CHARGE_RETRY;
2232
2233         /*
2234          * At task move, charge accounts can be doubly counted. So, it's
2235          * better to wait until the end of task_move if something is going on.
2236          */
2237         if (mem_cgroup_wait_acct_move(mem_over_limit))
2238                 return CHARGE_RETRY;
2239
2240         /* If we don't need to call oom-killer at el, return immediately */
2241         if (!oom_check)
2242                 return CHARGE_NOMEM;
2243         /* check OOM */
2244         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2245                 return CHARGE_OOM_DIE;
2246
2247         return CHARGE_RETRY;
2248 }
2249
2250 /*
2251  * Unlike exported interface, "oom" parameter is added. if oom==true,
2252  * oom-killer can be invoked.
2253  */
2254 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2255                                    gfp_t gfp_mask,
2256                                    unsigned int nr_pages,
2257                                    struct mem_cgroup **memcg,
2258                                    bool oom)
2259 {
2260         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2261         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2262         struct mem_cgroup *mem = NULL;
2263         int ret;
2264
2265         /*
2266          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2267          * in system level. So, allow to go ahead dying process in addition to
2268          * MEMDIE process.
2269          */
2270         if (unlikely(test_thread_flag(TIF_MEMDIE)
2271                      || fatal_signal_pending(current)))
2272                 goto bypass;
2273
2274         /*
2275          * We always charge the cgroup the mm_struct belongs to.
2276          * The mm_struct's mem_cgroup changes on task migration if the
2277          * thread group leader migrates. It's possible that mm is not
2278          * set, if so charge the init_mm (happens for pagecache usage).
2279          */
2280         if (!*memcg && !mm)
2281                 goto bypass;
2282 again:
2283         if (*memcg) { /* css should be a valid one */
2284                 mem = *memcg;
2285                 VM_BUG_ON(css_is_removed(&mem->css));
2286                 if (mem_cgroup_is_root(mem))
2287                         goto done;
2288                 if (nr_pages == 1 && consume_stock(mem))
2289                         goto done;
2290                 css_get(&mem->css);
2291         } else {
2292                 struct task_struct *p;
2293
2294                 rcu_read_lock();
2295                 p = rcu_dereference(mm->owner);
2296                 /*
2297                  * Because we don't have task_lock(), "p" can exit.
2298                  * In that case, "mem" can point to root or p can be NULL with
2299                  * race with swapoff. Then, we have small risk of mis-accouning.
2300                  * But such kind of mis-account by race always happens because
2301                  * we don't have cgroup_mutex(). It's overkill and we allo that
2302                  * small race, here.
2303                  * (*) swapoff at el will charge against mm-struct not against
2304                  * task-struct. So, mm->owner can be NULL.
2305                  */
2306                 mem = mem_cgroup_from_task(p);
2307                 if (!mem || mem_cgroup_is_root(mem)) {
2308                         rcu_read_unlock();
2309                         goto done;
2310                 }
2311                 if (nr_pages == 1 && consume_stock(mem)) {
2312                         /*
2313                          * It seems dagerous to access memcg without css_get().
2314                          * But considering how consume_stok works, it's not
2315                          * necessary. If consume_stock success, some charges
2316                          * from this memcg are cached on this cpu. So, we
2317                          * don't need to call css_get()/css_tryget() before
2318                          * calling consume_stock().
2319                          */
2320                         rcu_read_unlock();
2321                         goto done;
2322                 }
2323                 /* after here, we may be blocked. we need to get refcnt */
2324                 if (!css_tryget(&mem->css)) {
2325                         rcu_read_unlock();
2326                         goto again;
2327                 }
2328                 rcu_read_unlock();
2329         }
2330
2331         do {
2332                 bool oom_check;
2333
2334                 /* If killed, bypass charge */
2335                 if (fatal_signal_pending(current)) {
2336                         css_put(&mem->css);
2337                         goto bypass;
2338                 }
2339
2340                 oom_check = false;
2341                 if (oom && !nr_oom_retries) {
2342                         oom_check = true;
2343                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2344                 }
2345
2346                 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2347                 switch (ret) {
2348                 case CHARGE_OK:
2349                         break;
2350                 case CHARGE_RETRY: /* not in OOM situation but retry */
2351                         batch = nr_pages;
2352                         css_put(&mem->css);
2353                         mem = NULL;
2354                         goto again;
2355                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2356                         css_put(&mem->css);
2357                         goto nomem;
2358                 case CHARGE_NOMEM: /* OOM routine works */
2359                         if (!oom) {
2360                                 css_put(&mem->css);
2361                                 goto nomem;
2362                         }
2363                         /* If oom, we never return -ENOMEM */
2364                         nr_oom_retries--;
2365                         break;
2366                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2367                         css_put(&mem->css);
2368                         goto bypass;
2369                 }
2370         } while (ret != CHARGE_OK);
2371
2372         if (batch > nr_pages)
2373                 refill_stock(mem, batch - nr_pages);
2374         css_put(&mem->css);
2375 done:
2376         *memcg = mem;
2377         return 0;
2378 nomem:
2379         *memcg = NULL;
2380         return -ENOMEM;
2381 bypass:
2382         *memcg = NULL;
2383         return 0;
2384 }
2385
2386 /*
2387  * Somemtimes we have to undo a charge we got by try_charge().
2388  * This function is for that and do uncharge, put css's refcnt.
2389  * gotten by try_charge().
2390  */
2391 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2392                                        unsigned int nr_pages)
2393 {
2394         if (!mem_cgroup_is_root(mem)) {
2395                 unsigned long bytes = nr_pages * PAGE_SIZE;
2396
2397                 res_counter_uncharge(&mem->res, bytes);
2398                 if (do_swap_account)
2399                         res_counter_uncharge(&mem->memsw, bytes);
2400         }
2401 }
2402
2403 /*
2404  * A helper function to get mem_cgroup from ID. must be called under
2405  * rcu_read_lock(). The caller must check css_is_removed() or some if
2406  * it's concern. (dropping refcnt from swap can be called against removed
2407  * memcg.)
2408  */
2409 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2410 {
2411         struct cgroup_subsys_state *css;
2412
2413         /* ID 0 is unused ID */
2414         if (!id)
2415                 return NULL;
2416         css = css_lookup(&mem_cgroup_subsys, id);
2417         if (!css)
2418                 return NULL;
2419         return container_of(css, struct mem_cgroup, css);
2420 }
2421
2422 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2423 {
2424         struct mem_cgroup *mem = NULL;
2425         struct page_cgroup *pc;
2426         unsigned short id;
2427         swp_entry_t ent;
2428
2429         VM_BUG_ON(!PageLocked(page));
2430
2431         pc = lookup_page_cgroup(page);
2432         lock_page_cgroup(pc);
2433         if (PageCgroupUsed(pc)) {
2434                 mem = pc->mem_cgroup;
2435                 if (mem && !css_tryget(&mem->css))
2436                         mem = NULL;
2437         } else if (PageSwapCache(page)) {
2438                 ent.val = page_private(page);
2439                 id = lookup_swap_cgroup(ent);
2440                 rcu_read_lock();
2441                 mem = mem_cgroup_lookup(id);
2442                 if (mem && !css_tryget(&mem->css))
2443                         mem = NULL;
2444                 rcu_read_unlock();
2445         }
2446         unlock_page_cgroup(pc);
2447         return mem;
2448 }
2449
2450 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2451                                        struct page *page,
2452                                        unsigned int nr_pages,
2453                                        struct page_cgroup *pc,
2454                                        enum charge_type ctype)
2455 {
2456         lock_page_cgroup(pc);
2457         if (unlikely(PageCgroupUsed(pc))) {
2458                 unlock_page_cgroup(pc);
2459                 __mem_cgroup_cancel_charge(mem, nr_pages);
2460                 return;
2461         }
2462         /*
2463          * we don't need page_cgroup_lock about tail pages, becase they are not
2464          * accessed by any other context at this point.
2465          */
2466         pc->mem_cgroup = mem;
2467         /*
2468          * We access a page_cgroup asynchronously without lock_page_cgroup().
2469          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2470          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2471          * before USED bit, we need memory barrier here.
2472          * See mem_cgroup_add_lru_list(), etc.
2473          */
2474         smp_wmb();
2475         switch (ctype) {
2476         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2477         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2478                 SetPageCgroupCache(pc);
2479                 SetPageCgroupUsed(pc);
2480                 break;
2481         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2482                 ClearPageCgroupCache(pc);
2483                 SetPageCgroupUsed(pc);
2484                 break;
2485         default:
2486                 break;
2487         }
2488
2489         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2490         unlock_page_cgroup(pc);
2491         /*
2492          * "charge_statistics" updated event counter. Then, check it.
2493          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2494          * if they exceeds softlimit.
2495          */
2496         memcg_check_events(mem, page);
2497 }
2498
2499 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2500
2501 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2502                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2503 /*
2504  * Because tail pages are not marked as "used", set it. We're under
2505  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2506  */
2507 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2508 {
2509         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2510         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2511         unsigned long flags;
2512
2513         if (mem_cgroup_disabled())
2514                 return;
2515         /*
2516          * We have no races with charge/uncharge but will have races with
2517          * page state accounting.
2518          */
2519         move_lock_page_cgroup(head_pc, &flags);
2520
2521         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2522         smp_wmb(); /* see __commit_charge() */
2523         if (PageCgroupAcctLRU(head_pc)) {
2524                 enum lru_list lru;
2525                 struct mem_cgroup_per_zone *mz;
2526
2527                 /*
2528                  * LRU flags cannot be copied because we need to add tail
2529                  *.page to LRU by generic call and our hook will be called.
2530                  * We hold lru_lock, then, reduce counter directly.
2531                  */
2532                 lru = page_lru(head);
2533                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2534                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2535         }
2536         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2537         move_unlock_page_cgroup(head_pc, &flags);
2538 }
2539 #endif
2540
2541 /**
2542  * mem_cgroup_move_account - move account of the page
2543  * @page: the page
2544  * @nr_pages: number of regular pages (>1 for huge pages)
2545  * @pc: page_cgroup of the page.
2546  * @from: mem_cgroup which the page is moved from.
2547  * @to: mem_cgroup which the page is moved to. @from != @to.
2548  * @uncharge: whether we should call uncharge and css_put against @from.
2549  *
2550  * The caller must confirm following.
2551  * - page is not on LRU (isolate_page() is useful.)
2552  * - compound_lock is held when nr_pages > 1
2553  *
2554  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2555  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2556  * true, this function does "uncharge" from old cgroup, but it doesn't if
2557  * @uncharge is false, so a caller should do "uncharge".
2558  */
2559 static int mem_cgroup_move_account(struct page *page,
2560                                    unsigned int nr_pages,
2561                                    struct page_cgroup *pc,
2562                                    struct mem_cgroup *from,
2563                                    struct mem_cgroup *to,
2564                                    bool uncharge)
2565 {
2566         unsigned long flags;
2567         int ret;
2568
2569         VM_BUG_ON(from == to);
2570         VM_BUG_ON(PageLRU(page));
2571         /*
2572          * The page is isolated from LRU. So, collapse function
2573          * will not handle this page. But page splitting can happen.
2574          * Do this check under compound_page_lock(). The caller should
2575          * hold it.
2576          */
2577         ret = -EBUSY;
2578         if (nr_pages > 1 && !PageTransHuge(page))
2579                 goto out;
2580
2581         lock_page_cgroup(pc);
2582
2583         ret = -EINVAL;
2584         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2585                 goto unlock;
2586
2587         move_lock_page_cgroup(pc, &flags);
2588
2589         if (PageCgroupFileMapped(pc)) {
2590                 /* Update mapped_file data for mem_cgroup */
2591                 preempt_disable();
2592                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2593                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2594                 preempt_enable();
2595         }
2596         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2597         if (uncharge)
2598                 /* This is not "cancel", but cancel_charge does all we need. */
2599                 __mem_cgroup_cancel_charge(from, nr_pages);
2600
2601         /* caller should have done css_get */
2602         pc->mem_cgroup = to;
2603         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2604         /*
2605          * We charges against "to" which may not have any tasks. Then, "to"
2606          * can be under rmdir(). But in current implementation, caller of
2607          * this function is just force_empty() and move charge, so it's
2608          * guaranteed that "to" is never removed. So, we don't check rmdir
2609          * status here.
2610          */
2611         move_unlock_page_cgroup(pc, &flags);
2612         ret = 0;
2613 unlock:
2614         unlock_page_cgroup(pc);
2615         /*
2616          * check events
2617          */
2618         memcg_check_events(to, page);
2619         memcg_check_events(from, page);
2620 out:
2621         return ret;
2622 }
2623
2624 /*
2625  * move charges to its parent.
2626  */
2627
2628 static int mem_cgroup_move_parent(struct page *page,
2629                                   struct page_cgroup *pc,
2630                                   struct mem_cgroup *child,
2631                                   gfp_t gfp_mask)
2632 {
2633         struct cgroup *cg = child->css.cgroup;
2634         struct cgroup *pcg = cg->parent;
2635         struct mem_cgroup *parent;
2636         unsigned int nr_pages;
2637         unsigned long uninitialized_var(flags);
2638         int ret;
2639
2640         /* Is ROOT ? */
2641         if (!pcg)
2642                 return -EINVAL;
2643
2644         ret = -EBUSY;
2645         if (!get_page_unless_zero(page))
2646                 goto out;
2647         if (isolate_lru_page(page))
2648                 goto put;
2649
2650         nr_pages = hpage_nr_pages(page);
2651
2652         parent = mem_cgroup_from_cont(pcg);
2653         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2654         if (ret || !parent)
2655                 goto put_back;
2656
2657         if (nr_pages > 1)
2658                 flags = compound_lock_irqsave(page);
2659
2660         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2661         if (ret)
2662                 __mem_cgroup_cancel_charge(parent, nr_pages);
2663
2664         if (nr_pages > 1)
2665                 compound_unlock_irqrestore(page, flags);
2666 put_back:
2667         putback_lru_page(page);
2668 put:
2669         put_page(page);
2670 out:
2671         return ret;
2672 }
2673
2674 /*
2675  * Charge the memory controller for page usage.
2676  * Return
2677  * 0 if the charge was successful
2678  * < 0 if the cgroup is over its limit
2679  */
2680 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2681                                 gfp_t gfp_mask, enum charge_type ctype)
2682 {
2683         struct mem_cgroup *mem = NULL;
2684         unsigned int nr_pages = 1;
2685         struct page_cgroup *pc;
2686         bool oom = true;
2687         int ret;
2688
2689         if (PageTransHuge(page)) {
2690                 nr_pages <<= compound_order(page);
2691                 VM_BUG_ON(!PageTransHuge(page));
2692                 /*
2693                  * Never OOM-kill a process for a huge page.  The
2694                  * fault handler will fall back to regular pages.
2695                  */
2696                 oom = false;
2697         }
2698
2699         pc = lookup_page_cgroup(page);
2700         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2701
2702         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2703         if (ret || !mem)
2704                 return ret;
2705
2706         __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2707         return 0;
2708 }
2709
2710 int mem_cgroup_newpage_charge(struct page *page,
2711                               struct mm_struct *mm, gfp_t gfp_mask)
2712 {
2713         if (mem_cgroup_disabled())
2714                 return 0;
2715         /*
2716          * If already mapped, we don't have to account.
2717          * If page cache, page->mapping has address_space.
2718          * But page->mapping may have out-of-use anon_vma pointer,
2719          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2720          * is NULL.
2721          */
2722         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2723                 return 0;
2724         if (unlikely(!mm))
2725                 mm = &init_mm;
2726         return mem_cgroup_charge_common(page, mm, gfp_mask,
2727                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2728 }
2729
2730 static void
2731 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2732                                         enum charge_type ctype);
2733
2734 static void
2735 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2736                                         enum charge_type ctype)
2737 {
2738         struct page_cgroup *pc = lookup_page_cgroup(page);
2739         /*
2740          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2741          * is already on LRU. It means the page may on some other page_cgroup's
2742          * LRU. Take care of it.
2743          */
2744         mem_cgroup_lru_del_before_commit(page);
2745         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2746         mem_cgroup_lru_add_after_commit(page);
2747         return;
2748 }
2749
2750 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2751                                 gfp_t gfp_mask)
2752 {
2753         struct mem_cgroup *mem = NULL;
2754         int ret;
2755
2756         if (mem_cgroup_disabled())
2757                 return 0;
2758         if (PageCompound(page))
2759                 return 0;
2760         /*
2761          * Corner case handling. This is called from add_to_page_cache()
2762          * in usual. But some FS (shmem) precharges this page before calling it
2763          * and call add_to_page_cache() with GFP_NOWAIT.
2764          *
2765          * For GFP_NOWAIT case, the page may be pre-charged before calling
2766          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2767          * charge twice. (It works but has to pay a bit larger cost.)
2768          * And when the page is SwapCache, it should take swap information
2769          * into account. This is under lock_page() now.
2770          */
2771         if (!(gfp_mask & __GFP_WAIT)) {
2772                 struct page_cgroup *pc;
2773
2774                 pc = lookup_page_cgroup(page);
2775                 if (!pc)
2776                         return 0;
2777                 lock_page_cgroup(pc);
2778                 if (PageCgroupUsed(pc)) {
2779                         unlock_page_cgroup(pc);
2780                         return 0;
2781                 }
2782                 unlock_page_cgroup(pc);
2783         }
2784
2785         if (unlikely(!mm))
2786                 mm = &init_mm;
2787
2788         if (page_is_file_cache(page)) {
2789                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2790                 if (ret || !mem)
2791                         return ret;
2792
2793                 /*
2794                  * FUSE reuses pages without going through the final
2795                  * put that would remove them from the LRU list, make
2796                  * sure that they get relinked properly.
2797                  */
2798                 __mem_cgroup_commit_charge_lrucare(page, mem,
2799                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2800                 return ret;
2801         }
2802         /* shmem */
2803         if (PageSwapCache(page)) {
2804                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2805                 if (!ret)
2806                         __mem_cgroup_commit_charge_swapin(page, mem,
2807                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2808         } else
2809                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2810                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2811
2812         return ret;
2813 }
2814
2815 /*
2816  * While swap-in, try_charge -> commit or cancel, the page is locked.
2817  * And when try_charge() successfully returns, one refcnt to memcg without
2818  * struct page_cgroup is acquired. This refcnt will be consumed by
2819  * "commit()" or removed by "cancel()"
2820  */
2821 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2822                                  struct page *page,
2823                                  gfp_t mask, struct mem_cgroup **ptr)
2824 {
2825         struct mem_cgroup *mem;
2826         int ret;
2827
2828         *ptr = NULL;
2829
2830         if (mem_cgroup_disabled())
2831                 return 0;
2832
2833         if (!do_swap_account)
2834                 goto charge_cur_mm;
2835         /*
2836          * A racing thread's fault, or swapoff, may have already updated
2837          * the pte, and even removed page from swap cache: in those cases
2838          * do_swap_page()'s pte_same() test will fail; but there's also a
2839          * KSM case which does need to charge the page.
2840          */
2841         if (!PageSwapCache(page))
2842                 goto charge_cur_mm;
2843         mem = try_get_mem_cgroup_from_page(page);
2844         if (!mem)
2845                 goto charge_cur_mm;
2846         *ptr = mem;
2847         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2848         css_put(&mem->css);
2849         return ret;
2850 charge_cur_mm:
2851         if (unlikely(!mm))
2852                 mm = &init_mm;
2853         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2854 }
2855
2856 static void
2857 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2858                                         enum charge_type ctype)
2859 {
2860         if (mem_cgroup_disabled())
2861                 return;
2862         if (!ptr)
2863                 return;
2864         cgroup_exclude_rmdir(&ptr->css);
2865
2866         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2867         /*
2868          * Now swap is on-memory. This means this page may be
2869          * counted both as mem and swap....double count.
2870          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2871          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2872          * may call delete_from_swap_cache() before reach here.
2873          */
2874         if (do_swap_account && PageSwapCache(page)) {
2875                 swp_entry_t ent = {.val = page_private(page)};
2876                 unsigned short id;
2877                 struct mem_cgroup *memcg;
2878
2879                 id = swap_cgroup_record(ent, 0);
2880                 rcu_read_lock();
2881                 memcg = mem_cgroup_lookup(id);
2882                 if (memcg) {
2883                         /*
2884                          * This recorded memcg can be obsolete one. So, avoid
2885                          * calling css_tryget
2886                          */
2887                         if (!mem_cgroup_is_root(memcg))
2888                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2889                         mem_cgroup_swap_statistics(memcg, false);
2890                         mem_cgroup_put(memcg);
2891                 }
2892                 rcu_read_unlock();
2893         }
2894         /*
2895          * At swapin, we may charge account against cgroup which has no tasks.
2896          * So, rmdir()->pre_destroy() can be called while we do this charge.
2897          * In that case, we need to call pre_destroy() again. check it here.
2898          */
2899         cgroup_release_and_wakeup_rmdir(&ptr->css);
2900 }
2901
2902 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2903 {
2904         __mem_cgroup_commit_charge_swapin(page, ptr,
2905                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2906 }
2907
2908 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2909 {
2910         if (mem_cgroup_disabled())
2911                 return;
2912         if (!mem)
2913                 return;
2914         __mem_cgroup_cancel_charge(mem, 1);
2915 }
2916
2917 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2918                                    unsigned int nr_pages,
2919                                    const enum charge_type ctype)
2920 {
2921         struct memcg_batch_info *batch = NULL;
2922         bool uncharge_memsw = true;
2923
2924         /* If swapout, usage of swap doesn't decrease */
2925         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2926                 uncharge_memsw = false;
2927
2928         batch = &current->memcg_batch;
2929         /*
2930          * In usual, we do css_get() when we remember memcg pointer.
2931          * But in this case, we keep res->usage until end of a series of
2932          * uncharges. Then, it's ok to ignore memcg's refcnt.
2933          */
2934         if (!batch->memcg)
2935                 batch->memcg = mem;
2936         /*
2937          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2938          * In those cases, all pages freed continuously can be expected to be in
2939          * the same cgroup and we have chance to coalesce uncharges.
2940          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2941          * because we want to do uncharge as soon as possible.
2942          */
2943
2944         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2945                 goto direct_uncharge;
2946
2947         if (nr_pages > 1)
2948                 goto direct_uncharge;
2949
2950         /*
2951          * In typical case, batch->memcg == mem. This means we can
2952          * merge a series of uncharges to an uncharge of res_counter.
2953          * If not, we uncharge res_counter ony by one.
2954          */
2955         if (batch->memcg != mem)
2956                 goto direct_uncharge;
2957         /* remember freed charge and uncharge it later */
2958         batch->nr_pages++;
2959         if (uncharge_memsw)
2960                 batch->memsw_nr_pages++;
2961         return;
2962 direct_uncharge:
2963         res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2964         if (uncharge_memsw)
2965                 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2966         if (unlikely(batch->memcg != mem))
2967                 memcg_oom_recover(mem);
2968         return;
2969 }
2970
2971 /*
2972  * uncharge if !page_mapped(page)
2973  */
2974 static struct mem_cgroup *
2975 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2976 {
2977         struct mem_cgroup *mem = NULL;
2978         unsigned int nr_pages = 1;
2979         struct page_cgroup *pc;
2980
2981         if (mem_cgroup_disabled())
2982                 return NULL;
2983
2984         if (PageSwapCache(page))
2985                 return NULL;
2986
2987         if (PageTransHuge(page)) {
2988                 nr_pages <<= compound_order(page);
2989                 VM_BUG_ON(!PageTransHuge(page));
2990         }
2991         /*
2992          * Check if our page_cgroup is valid
2993          */
2994         pc = lookup_page_cgroup(page);
2995         if (unlikely(!pc || !PageCgroupUsed(pc)))
2996                 return NULL;
2997
2998         lock_page_cgroup(pc);
2999
3000         mem = pc->mem_cgroup;
3001
3002         if (!PageCgroupUsed(pc))
3003                 goto unlock_out;
3004
3005         switch (ctype) {
3006         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3007         case MEM_CGROUP_CHARGE_TYPE_DROP:
3008                 /* See mem_cgroup_prepare_migration() */
3009                 if (page_mapped(page) || PageCgroupMigration(pc))
3010                         goto unlock_out;
3011                 break;
3012         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3013                 if (!PageAnon(page)) {  /* Shared memory */
3014                         if (page->mapping && !page_is_file_cache(page))
3015                                 goto unlock_out;
3016                 } else if (page_mapped(page)) /* Anon */
3017                                 goto unlock_out;
3018                 break;
3019         default:
3020                 break;
3021         }
3022
3023         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3024
3025         ClearPageCgroupUsed(pc);
3026         /*
3027          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3028          * freed from LRU. This is safe because uncharged page is expected not
3029          * to be reused (freed soon). Exception is SwapCache, it's handled by
3030          * special functions.
3031          */
3032
3033         unlock_page_cgroup(pc);
3034         /*
3035          * even after unlock, we have mem->res.usage here and this memcg
3036          * will never be freed.
3037          */
3038         memcg_check_events(mem, page);
3039         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3040                 mem_cgroup_swap_statistics(mem, true);
3041                 mem_cgroup_get(mem);
3042         }
3043         if (!mem_cgroup_is_root(mem))
3044                 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3045
3046         return mem;
3047
3048 unlock_out:
3049         unlock_page_cgroup(pc);
3050         return NULL;
3051 }
3052
3053 void mem_cgroup_uncharge_page(struct page *page)
3054 {
3055         /* early check. */
3056         if (page_mapped(page))
3057                 return;
3058         if (page->mapping && !PageAnon(page))
3059                 return;
3060         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3061 }
3062
3063 void mem_cgroup_uncharge_cache_page(struct page *page)
3064 {
3065         VM_BUG_ON(page_mapped(page));
3066         VM_BUG_ON(page->mapping);
3067         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3068 }
3069
3070 /*
3071  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3072  * In that cases, pages are freed continuously and we can expect pages
3073  * are in the same memcg. All these calls itself limits the number of
3074  * pages freed at once, then uncharge_start/end() is called properly.
3075  * This may be called prural(2) times in a context,
3076  */
3077
3078 void mem_cgroup_uncharge_start(void)
3079 {
3080         current->memcg_batch.do_batch++;
3081         /* We can do nest. */
3082         if (current->memcg_batch.do_batch == 1) {
3083                 current->memcg_batch.memcg = NULL;
3084                 current->memcg_batch.nr_pages = 0;
3085                 current->memcg_batch.memsw_nr_pages = 0;
3086         }
3087 }
3088
3089 void mem_cgroup_uncharge_end(void)
3090 {
3091         struct memcg_batch_info *batch = &current->memcg_batch;
3092
3093         if (!batch->do_batch)
3094                 return;
3095
3096         batch->do_batch--;
3097         if (batch->do_batch) /* If stacked, do nothing. */
3098                 return;
3099
3100         if (!batch->memcg)
3101                 return;
3102         /*
3103          * This "batch->memcg" is valid without any css_get/put etc...
3104          * bacause we hide charges behind us.
3105          */
3106         if (batch->nr_pages)
3107                 res_counter_uncharge(&batch->memcg->res,
3108                                      batch->nr_pages * PAGE_SIZE);
3109         if (batch->memsw_nr_pages)
3110                 res_counter_uncharge(&batch->memcg->memsw,
3111                                      batch->memsw_nr_pages * PAGE_SIZE);
3112         memcg_oom_recover(batch->memcg);
3113         /* forget this pointer (for sanity check) */
3114         batch->memcg = NULL;
3115 }
3116
3117 #ifdef CONFIG_SWAP
3118 /*
3119  * called after __delete_from_swap_cache() and drop "page" account.
3120  * memcg information is recorded to swap_cgroup of "ent"
3121  */
3122 void
3123 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3124 {
3125         struct mem_cgroup *memcg;
3126         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3127
3128         if (!swapout) /* this was a swap cache but the swap is unused ! */
3129                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3130
3131         memcg = __mem_cgroup_uncharge_common(page, ctype);
3132
3133         /*
3134          * record memcg information,  if swapout && memcg != NULL,
3135          * mem_cgroup_get() was called in uncharge().
3136          */
3137         if (do_swap_account && swapout && memcg)
3138                 swap_cgroup_record(ent, css_id(&memcg->css));
3139 }
3140 #endif
3141
3142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3143 /*
3144  * called from swap_entry_free(). remove record in swap_cgroup and
3145  * uncharge "memsw" account.
3146  */
3147 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3148 {
3149         struct mem_cgroup *memcg;
3150         unsigned short id;
3151
3152         if (!do_swap_account)
3153                 return;
3154
3155         id = swap_cgroup_record(ent, 0);
3156         rcu_read_lock();
3157         memcg = mem_cgroup_lookup(id);
3158         if (memcg) {
3159                 /*
3160                  * We uncharge this because swap is freed.
3161                  * This memcg can be obsolete one. We avoid calling css_tryget
3162                  */
3163                 if (!mem_cgroup_is_root(memcg))
3164                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3165                 mem_cgroup_swap_statistics(memcg, false);
3166                 mem_cgroup_put(memcg);
3167         }
3168         rcu_read_unlock();
3169 }
3170
3171 /**
3172  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3173  * @entry: swap entry to be moved
3174  * @from:  mem_cgroup which the entry is moved from
3175  * @to:  mem_cgroup which the entry is moved to
3176  * @need_fixup: whether we should fixup res_counters and refcounts.
3177  *
3178  * It succeeds only when the swap_cgroup's record for this entry is the same
3179  * as the mem_cgroup's id of @from.
3180  *
3181  * Returns 0 on success, -EINVAL on failure.
3182  *
3183  * The caller must have charged to @to, IOW, called res_counter_charge() about
3184  * both res and memsw, and called css_get().
3185  */
3186 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3187                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3188 {
3189         unsigned short old_id, new_id;
3190
3191         old_id = css_id(&from->css);
3192         new_id = css_id(&to->css);
3193
3194         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3195                 mem_cgroup_swap_statistics(from, false);
3196                 mem_cgroup_swap_statistics(to, true);
3197                 /*
3198                  * This function is only called from task migration context now.
3199                  * It postpones res_counter and refcount handling till the end
3200                  * of task migration(mem_cgroup_clear_mc()) for performance
3201                  * improvement. But we cannot postpone mem_cgroup_get(to)
3202                  * because if the process that has been moved to @to does
3203                  * swap-in, the refcount of @to might be decreased to 0.
3204                  */
3205                 mem_cgroup_get(to);
3206                 if (need_fixup) {
3207                         if (!mem_cgroup_is_root(from))
3208                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3209                         mem_cgroup_put(from);
3210                         /*
3211                          * we charged both to->res and to->memsw, so we should
3212                          * uncharge to->res.
3213                          */
3214                         if (!mem_cgroup_is_root(to))
3215                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3216                 }
3217                 return 0;
3218         }
3219         return -EINVAL;
3220 }
3221 #else
3222 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3223                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3224 {
3225         return -EINVAL;
3226 }
3227 #endif
3228
3229 /*
3230  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3231  * page belongs to.
3232  */
3233 int mem_cgroup_prepare_migration(struct page *page,
3234         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3235 {
3236         struct mem_cgroup *mem = NULL;
3237         struct page_cgroup *pc;
3238         enum charge_type ctype;
3239         int ret = 0;
3240
3241         *ptr = NULL;
3242
3243         VM_BUG_ON(PageTransHuge(page));
3244         if (mem_cgroup_disabled())
3245                 return 0;
3246
3247         pc = lookup_page_cgroup(page);
3248         lock_page_cgroup(pc);
3249         if (PageCgroupUsed(pc)) {
3250                 mem = pc->mem_cgroup;
3251                 css_get(&mem->css);
3252                 /*
3253                  * At migrating an anonymous page, its mapcount goes down
3254                  * to 0 and uncharge() will be called. But, even if it's fully
3255                  * unmapped, migration may fail and this page has to be
3256                  * charged again. We set MIGRATION flag here and delay uncharge
3257                  * until end_migration() is called
3258                  *
3259                  * Corner Case Thinking
3260                  * A)
3261                  * When the old page was mapped as Anon and it's unmap-and-freed
3262                  * while migration was ongoing.
3263                  * If unmap finds the old page, uncharge() of it will be delayed
3264                  * until end_migration(). If unmap finds a new page, it's
3265                  * uncharged when it make mapcount to be 1->0. If unmap code
3266                  * finds swap_migration_entry, the new page will not be mapped
3267                  * and end_migration() will find it(mapcount==0).
3268                  *
3269                  * B)
3270                  * When the old page was mapped but migraion fails, the kernel
3271                  * remaps it. A charge for it is kept by MIGRATION flag even
3272                  * if mapcount goes down to 0. We can do remap successfully
3273                  * without charging it again.
3274                  *
3275                  * C)
3276                  * The "old" page is under lock_page() until the end of
3277                  * migration, so, the old page itself will not be swapped-out.
3278                  * If the new page is swapped out before end_migraton, our
3279                  * hook to usual swap-out path will catch the event.
3280                  */
3281                 if (PageAnon(page))
3282                         SetPageCgroupMigration(pc);
3283         }
3284         unlock_page_cgroup(pc);
3285         /*
3286          * If the page is not charged at this point,
3287          * we return here.
3288          */
3289         if (!mem)
3290                 return 0;
3291
3292         *ptr = mem;
3293         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3294         css_put(&mem->css);/* drop extra refcnt */
3295         if (ret || *ptr == NULL) {
3296                 if (PageAnon(page)) {
3297                         lock_page_cgroup(pc);
3298                         ClearPageCgroupMigration(pc);
3299                         unlock_page_cgroup(pc);
3300                         /*
3301                          * The old page may be fully unmapped while we kept it.
3302                          */
3303                         mem_cgroup_uncharge_page(page);
3304                 }
3305                 return -ENOMEM;
3306         }
3307         /*
3308          * We charge new page before it's used/mapped. So, even if unlock_page()
3309          * is called before end_migration, we can catch all events on this new
3310          * page. In the case new page is migrated but not remapped, new page's
3311          * mapcount will be finally 0 and we call uncharge in end_migration().
3312          */
3313         pc = lookup_page_cgroup(newpage);
3314         if (PageAnon(page))
3315                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3316         else if (page_is_file_cache(page))
3317                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3318         else
3319                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3320         __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3321         return ret;
3322 }
3323
3324 /* remove redundant charge if migration failed*/
3325 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3326         struct page *oldpage, struct page *newpage, bool migration_ok)
3327 {
3328         struct page *used, *unused;
3329         struct page_cgroup *pc;
3330
3331         if (!mem)
3332                 return;
3333         /* blocks rmdir() */
3334         cgroup_exclude_rmdir(&mem->css);
3335         if (!migration_ok) {
3336                 used = oldpage;
3337                 unused = newpage;
3338         } else {
3339                 used = newpage;
3340                 unused = oldpage;
3341         }
3342         /*
3343          * We disallowed uncharge of pages under migration because mapcount
3344          * of the page goes down to zero, temporarly.
3345          * Clear the flag and check the page should be charged.
3346          */
3347         pc = lookup_page_cgroup(oldpage);
3348         lock_page_cgroup(pc);
3349         ClearPageCgroupMigration(pc);
3350         unlock_page_cgroup(pc);
3351
3352         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3353
3354         /*
3355          * If a page is a file cache, radix-tree replacement is very atomic
3356          * and we can skip this check. When it was an Anon page, its mapcount
3357          * goes down to 0. But because we added MIGRATION flage, it's not
3358          * uncharged yet. There are several case but page->mapcount check
3359          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3360          * check. (see prepare_charge() also)
3361          */
3362         if (PageAnon(used))
3363                 mem_cgroup_uncharge_page(used);
3364         /*
3365          * At migration, we may charge account against cgroup which has no
3366          * tasks.
3367          * So, rmdir()->pre_destroy() can be called while we do this charge.
3368          * In that case, we need to call pre_destroy() again. check it here.
3369          */
3370         cgroup_release_and_wakeup_rmdir(&mem->css);
3371 }
3372
3373 /*
3374  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3375  * Calling hierarchical_reclaim is not enough because we should update
3376  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3377  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3378  * not from the memcg which this page would be charged to.
3379  * try_charge_swapin does all of these works properly.
3380  */
3381 int mem_cgroup_shmem_charge_fallback(struct page *page,
3382                             struct mm_struct *mm,
3383                             gfp_t gfp_mask)
3384 {
3385         struct mem_cgroup *mem;
3386         int ret;
3387
3388         if (mem_cgroup_disabled())
3389                 return 0;
3390
3391         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3392         if (!ret)
3393                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3394
3395         return ret;
3396 }
3397
3398 #ifdef CONFIG_DEBUG_VM
3399 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3400 {
3401         struct page_cgroup *pc;
3402
3403         pc = lookup_page_cgroup(page);
3404         if (likely(pc) && PageCgroupUsed(pc))
3405                 return pc;
3406         return NULL;
3407 }
3408
3409 bool mem_cgroup_bad_page_check(struct page *page)
3410 {
3411         if (mem_cgroup_disabled())
3412                 return false;
3413
3414         return lookup_page_cgroup_used(page) != NULL;
3415 }
3416
3417 void mem_cgroup_print_bad_page(struct page *page)
3418 {
3419         struct page_cgroup *pc;
3420
3421         pc = lookup_page_cgroup_used(page);
3422         if (pc) {
3423                 int ret = -1;
3424                 char *path;
3425
3426                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3427                        pc, pc->flags, pc->mem_cgroup);
3428
3429                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3430                 if (path) {
3431                         rcu_read_lock();
3432                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3433                                                         path, PATH_MAX);
3434                         rcu_read_unlock();
3435                 }
3436
3437                 printk(KERN_CONT "(%s)\n",
3438                                 (ret < 0) ? "cannot get the path" : path);
3439                 kfree(path);
3440         }
3441 }
3442 #endif
3443
3444 static DEFINE_MUTEX(set_limit_mutex);
3445
3446 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3447                                 unsigned long long val)
3448 {
3449         int retry_count;
3450         u64 memswlimit, memlimit;
3451         int ret = 0;
3452         int children = mem_cgroup_count_children(memcg);
3453         u64 curusage, oldusage;
3454         int enlarge;
3455
3456         /*
3457          * For keeping hierarchical_reclaim simple, how long we should retry
3458          * is depends on callers. We set our retry-count to be function
3459          * of # of children which we should visit in this loop.
3460          */
3461         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3462
3463         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3464
3465         enlarge = 0;
3466         while (retry_count) {
3467                 if (signal_pending(current)) {
3468                         ret = -EINTR;
3469                         break;
3470                 }
3471                 /*
3472                  * Rather than hide all in some function, I do this in
3473                  * open coded manner. You see what this really does.
3474                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3475                  */
3476                 mutex_lock(&set_limit_mutex);
3477                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3478                 if (memswlimit < val) {
3479                         ret = -EINVAL;
3480                         mutex_unlock(&set_limit_mutex);
3481                         break;
3482                 }
3483
3484                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3485                 if (memlimit < val)
3486                         enlarge = 1;
3487
3488                 ret = res_counter_set_limit(&memcg->res, val);
3489                 if (!ret) {
3490                         if (memswlimit == val)
3491                                 memcg->memsw_is_minimum = true;
3492                         else
3493                                 memcg->memsw_is_minimum = false;
3494                 }
3495                 mutex_unlock(&set_limit_mutex);
3496
3497                 if (!ret)
3498                         break;
3499
3500                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3501                                                 MEM_CGROUP_RECLAIM_SHRINK,
3502                                                 NULL);
3503                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3504                 /* Usage is reduced ? */
3505                 if (curusage >= oldusage)
3506                         retry_count--;
3507                 else
3508                         oldusage = curusage;
3509         }
3510         if (!ret && enlarge)
3511                 memcg_oom_recover(memcg);
3512
3513         return ret;
3514 }
3515
3516 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3517                                         unsigned long long val)
3518 {
3519         int retry_count;
3520         u64 memlimit, memswlimit, oldusage, curusage;
3521         int children = mem_cgroup_count_children(memcg);
3522         int ret = -EBUSY;
3523         int enlarge = 0;
3524
3525         /* see mem_cgroup_resize_res_limit */
3526         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3527         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3528         while (retry_count) {
3529                 if (signal_pending(current)) {
3530                         ret = -EINTR;
3531                         break;
3532                 }
3533                 /*
3534                  * Rather than hide all in some function, I do this in
3535                  * open coded manner. You see what this really does.
3536                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3537                  */
3538                 mutex_lock(&set_limit_mutex);
3539                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3540                 if (memlimit > val) {
3541                         ret = -EINVAL;
3542                         mutex_unlock(&set_limit_mutex);
3543                         break;
3544                 }
3545                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3546                 if (memswlimit < val)
3547                         enlarge = 1;
3548                 ret = res_counter_set_limit(&memcg->memsw, val);
3549                 if (!ret) {
3550                         if (memlimit == val)
3551                                 memcg->memsw_is_minimum = true;
3552                         else
3553                                 memcg->memsw_is_minimum = false;
3554                 }
3555                 mutex_unlock(&set_limit_mutex);
3556
3557                 if (!ret)
3558                         break;
3559
3560                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3561                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3562                                                 MEM_CGROUP_RECLAIM_SHRINK,
3563                                                 NULL);
3564                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3565                 /* Usage is reduced ? */
3566                 if (curusage >= oldusage)
3567                         retry_count--;
3568                 else
3569                         oldusage = curusage;
3570         }
3571         if (!ret && enlarge)
3572                 memcg_oom_recover(memcg);
3573         return ret;
3574 }
3575
3576 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3577                                             gfp_t gfp_mask,
3578                                             unsigned long *total_scanned)
3579 {
3580         unsigned long nr_reclaimed = 0;
3581         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3582         unsigned long reclaimed;
3583         int loop = 0;
3584         struct mem_cgroup_tree_per_zone *mctz;
3585         unsigned long long excess;
3586         unsigned long nr_scanned;
3587
3588         if (order > 0)
3589                 return 0;
3590
3591         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3592         /*
3593          * This loop can run a while, specially if mem_cgroup's continuously
3594          * keep exceeding their soft limit and putting the system under
3595          * pressure
3596          */
3597         do {
3598                 if (next_mz)
3599                         mz = next_mz;
3600                 else
3601                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3602                 if (!mz)
3603                         break;
3604
3605                 nr_scanned = 0;
3606                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3607                                                 gfp_mask,
3608                                                 MEM_CGROUP_RECLAIM_SOFT,
3609                                                 &nr_scanned);
3610                 nr_reclaimed += reclaimed;
3611                 *total_scanned += nr_scanned;
3612                 spin_lock(&mctz->lock);
3613
3614                 /*
3615                  * If we failed to reclaim anything from this memory cgroup
3616                  * it is time to move on to the next cgroup
3617                  */
3618                 next_mz = NULL;
3619                 if (!reclaimed) {
3620                         do {
3621                                 /*
3622                                  * Loop until we find yet another one.
3623                                  *
3624                                  * By the time we get the soft_limit lock
3625                                  * again, someone might have aded the
3626                                  * group back on the RB tree. Iterate to
3627                                  * make sure we get a different mem.
3628                                  * mem_cgroup_largest_soft_limit_node returns
3629                                  * NULL if no other cgroup is present on
3630                                  * the tree
3631                                  */
3632                                 next_mz =
3633                                 __mem_cgroup_largest_soft_limit_node(mctz);
3634                                 if (next_mz == mz)
3635                                         css_put(&next_mz->mem->css);
3636                                 else /* next_mz == NULL or other memcg */
3637                                         break;
3638                         } while (1);
3639                 }
3640                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3641                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3642                 /*
3643                  * One school of thought says that we should not add
3644                  * back the node to the tree if reclaim returns 0.
3645                  * But our reclaim could return 0, simply because due
3646                  * to priority we are exposing a smaller subset of
3647                  * memory to reclaim from. Consider this as a longer
3648                  * term TODO.
3649                  */
3650                 /* If excess == 0, no tree ops */
3651                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3652                 spin_unlock(&mctz->lock);
3653                 css_put(&mz->mem->css);
3654                 loop++;
3655                 /*
3656                  * Could not reclaim anything and there are no more
3657                  * mem cgroups to try or we seem to be looping without
3658                  * reclaiming anything.
3659                  */
3660                 if (!nr_reclaimed &&
3661                         (next_mz == NULL ||
3662                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3663                         break;
3664         } while (!nr_reclaimed);
3665         if (next_mz)
3666                 css_put(&next_mz->mem->css);
3667         return nr_reclaimed;
3668 }
3669
3670 /*
3671  * This routine traverse page_cgroup in given list and drop them all.
3672  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3673  */
3674 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3675                                 int node, int zid, enum lru_list lru)
3676 {
3677         struct zone *zone;
3678         struct mem_cgroup_per_zone *mz;
3679         struct page_cgroup *pc, *busy;
3680         unsigned long flags, loop;
3681         struct list_head *list;
3682         int ret = 0;
3683
3684         zone = &NODE_DATA(node)->node_zones[zid];
3685         mz = mem_cgroup_zoneinfo(mem, node, zid);
3686         list = &mz->lists[lru];
3687
3688         loop = MEM_CGROUP_ZSTAT(mz, lru);
3689         /* give some margin against EBUSY etc...*/
3690         loop += 256;
3691         busy = NULL;
3692         while (loop--) {
3693                 struct page *page;
3694
3695                 ret = 0;
3696                 spin_lock_irqsave(&zone->lru_lock, flags);
3697                 if (list_empty(list)) {
3698                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3699                         break;
3700                 }
3701                 pc = list_entry(list->prev, struct page_cgroup, lru);
3702                 if (busy == pc) {
3703                         list_move(&pc->lru, list);
3704                         busy = NULL;
3705                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3706                         continue;
3707                 }
3708                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3709
3710                 page = lookup_cgroup_page(pc);
3711
3712                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3713                 if (ret == -ENOMEM)
3714                         break;
3715
3716                 if (ret == -EBUSY || ret == -EINVAL) {
3717                         /* found lock contention or "pc" is obsolete. */
3718                         busy = pc;
3719                         cond_resched();
3720                 } else
3721                         busy = NULL;
3722         }
3723
3724         if (!ret && !list_empty(list))
3725                 return -EBUSY;
3726         return ret;
3727 }
3728
3729 /*
3730  * make mem_cgroup's charge to be 0 if there is no task.
3731  * This enables deleting this mem_cgroup.
3732  */
3733 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3734 {
3735         int ret;
3736         int node, zid, shrink;
3737         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3738         struct cgroup *cgrp = mem->css.cgroup;
3739
3740         css_get(&mem->css);
3741
3742         shrink = 0;
3743         /* should free all ? */
3744         if (free_all)
3745                 goto try_to_free;
3746 move_account:
3747         do {
3748                 ret = -EBUSY;
3749                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3750                         goto out;
3751                 ret = -EINTR;
3752                 if (signal_pending(current))
3753                         goto out;
3754                 /* This is for making all *used* pages to be on LRU. */
3755                 lru_add_drain_all();
3756                 drain_all_stock_sync();
3757                 ret = 0;
3758                 mem_cgroup_start_move(mem);
3759                 for_each_node_state(node, N_HIGH_MEMORY) {
3760                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3761                                 enum lru_list l;
3762                                 for_each_lru(l) {
3763                                         ret = mem_cgroup_force_empty_list(mem,
3764                                                         node, zid, l);
3765                                         if (ret)
3766                                                 break;
3767                                 }
3768                         }
3769                         if (ret)
3770                                 break;
3771                 }
3772                 mem_cgroup_end_move(mem);
3773                 memcg_oom_recover(mem);
3774                 /* it seems parent cgroup doesn't have enough mem */
3775                 if (ret == -ENOMEM)
3776                         goto try_to_free;
3777                 cond_resched();
3778         /* "ret" should also be checked to ensure all lists are empty. */
3779         } while (mem->res.usage > 0 || ret);
3780 out:
3781         css_put(&mem->css);
3782         return ret;
3783
3784 try_to_free:
3785         /* returns EBUSY if there is a task or if we come here twice. */
3786         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3787                 ret = -EBUSY;
3788                 goto out;
3789         }
3790         /* we call try-to-free pages for make this cgroup empty */
3791         lru_add_drain_all();
3792         /* try to free all pages in this cgroup */
3793         shrink = 1;
3794         while (nr_retries && mem->res.usage > 0) {
3795                 int progress;
3796
3797                 if (signal_pending(current)) {
3798                         ret = -EINTR;
3799                         goto out;
3800                 }
3801                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3802                                                 false);
3803                 if (!progress) {
3804                         nr_retries--;
3805                         /* maybe some writeback is necessary */
3806                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3807                 }
3808
3809         }
3810         lru_add_drain();
3811         /* try move_account...there may be some *locked* pages. */
3812         goto move_account;
3813 }
3814
3815 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3816 {
3817         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3818 }
3819
3820
3821 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3822 {
3823         return mem_cgroup_from_cont(cont)->use_hierarchy;
3824 }
3825
3826 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3827                                         u64 val)
3828 {
3829         int retval = 0;
3830         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3831         struct cgroup *parent = cont->parent;
3832         struct mem_cgroup *parent_mem = NULL;
3833
3834         if (parent)
3835                 parent_mem = mem_cgroup_from_cont(parent);
3836
3837         cgroup_lock();
3838         /*
3839          * If parent's use_hierarchy is set, we can't make any modifications
3840          * in the child subtrees. If it is unset, then the change can
3841          * occur, provided the current cgroup has no children.
3842          *
3843          * For the root cgroup, parent_mem is NULL, we allow value to be
3844          * set if there are no children.
3845          */
3846         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3847                                 (val == 1 || val == 0)) {
3848                 if (list_empty(&cont->children))
3849                         mem->use_hierarchy = val;
3850                 else
3851                         retval = -EBUSY;
3852         } else
3853                 retval = -EINVAL;
3854         cgroup_unlock();
3855
3856         return retval;
3857 }
3858
3859
3860 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3861                                                enum mem_cgroup_stat_index idx)
3862 {
3863         struct mem_cgroup *iter;
3864         long val = 0;
3865
3866         /* Per-cpu values can be negative, use a signed accumulator */
3867         for_each_mem_cgroup_tree(iter, mem)
3868                 val += mem_cgroup_read_stat(iter, idx);
3869
3870         if (val < 0) /* race ? */
3871                 val = 0;
3872         return val;
3873 }
3874
3875 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3876 {
3877         u64 val;
3878
3879         if (!mem_cgroup_is_root(mem)) {
3880                 if (!swap)
3881                         return res_counter_read_u64(&mem->res, RES_USAGE);
3882                 else
3883                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3884         }
3885
3886         val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3887         val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3888
3889         if (swap)
3890                 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3891
3892         return val << PAGE_SHIFT;
3893 }
3894
3895 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3896 {
3897         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3898         u64 val;
3899         int type, name;
3900
3901         type = MEMFILE_TYPE(cft->private);
3902         name = MEMFILE_ATTR(cft->private);
3903         switch (type) {
3904         case _MEM:
3905                 if (name == RES_USAGE)
3906                         val = mem_cgroup_usage(mem, false);
3907                 else
3908                         val = res_counter_read_u64(&mem->res, name);
3909                 break;
3910         case _MEMSWAP:
3911                 if (name == RES_USAGE)
3912                         val = mem_cgroup_usage(mem, true);
3913                 else
3914                         val = res_counter_read_u64(&mem->memsw, name);
3915                 break;
3916         default:
3917                 BUG();
3918                 break;
3919         }
3920         return val;
3921 }
3922 /*
3923  * The user of this function is...
3924  * RES_LIMIT.
3925  */
3926 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3927                             const char *buffer)
3928 {
3929         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3930         int type, name;
3931         unsigned long long val;
3932         int ret;
3933
3934         type = MEMFILE_TYPE(cft->private);
3935         name = MEMFILE_ATTR(cft->private);
3936         switch (name) {
3937         case RES_LIMIT:
3938                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3939                         ret = -EINVAL;
3940                         break;
3941                 }
3942                 /* This function does all necessary parse...reuse it */
3943                 ret = res_counter_memparse_write_strategy(buffer, &val);
3944                 if (ret)
3945                         break;
3946                 if (type == _MEM)
3947                         ret = mem_cgroup_resize_limit(memcg, val);
3948                 else
3949                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3950                 break;
3951         case RES_SOFT_LIMIT:
3952                 ret = res_counter_memparse_write_strategy(buffer, &val);
3953                 if (ret)
3954                         break;
3955                 /*
3956                  * For memsw, soft limits are hard to implement in terms
3957                  * of semantics, for now, we support soft limits for
3958                  * control without swap
3959                  */
3960                 if (type == _MEM)
3961                         ret = res_counter_set_soft_limit(&memcg->res, val);
3962                 else
3963                         ret = -EINVAL;
3964                 break;
3965         default:
3966                 ret = -EINVAL; /* should be BUG() ? */
3967                 break;
3968         }
3969         return ret;
3970 }
3971
3972 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3973                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3974 {
3975         struct cgroup *cgroup;
3976         unsigned long long min_limit, min_memsw_limit, tmp;
3977
3978         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3979         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3980         cgroup = memcg->css.cgroup;
3981         if (!memcg->use_hierarchy)
3982                 goto out;
3983
3984         while (cgroup->parent) {
3985                 cgroup = cgroup->parent;
3986                 memcg = mem_cgroup_from_cont(cgroup);
3987                 if (!memcg->use_hierarchy)
3988                         break;
3989                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3990                 min_limit = min(min_limit, tmp);
3991                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3992                 min_memsw_limit = min(min_memsw_limit, tmp);
3993         }
3994 out:
3995         *mem_limit = min_limit;
3996         *memsw_limit = min_memsw_limit;
3997         return;
3998 }
3999
4000 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4001 {
4002         struct mem_cgroup *mem;
4003         int type, name;
4004
4005         mem = mem_cgroup_from_cont(cont);
4006         type = MEMFILE_TYPE(event);
4007         name = MEMFILE_ATTR(event);
4008         switch (name) {
4009         case RES_MAX_USAGE:
4010                 if (type == _MEM)
4011                         res_counter_reset_max(&mem->res);
4012                 else
4013                         res_counter_reset_max(&mem->memsw);
4014                 break;
4015         case RES_FAILCNT:
4016                 if (type == _MEM)
4017                         res_counter_reset_failcnt(&mem->res);
4018                 else
4019                         res_counter_reset_failcnt(&mem->memsw);
4020                 break;
4021         }
4022
4023         return 0;
4024 }
4025
4026 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4027                                         struct cftype *cft)
4028 {
4029         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4030 }
4031
4032 #ifdef CONFIG_MMU
4033 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4034                                         struct cftype *cft, u64 val)
4035 {
4036         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4037
4038         if (val >= (1 << NR_MOVE_TYPE))
4039                 return -EINVAL;
4040         /*
4041          * We check this value several times in both in can_attach() and
4042          * attach(), so we need cgroup lock to prevent this value from being
4043          * inconsistent.
4044          */
4045         cgroup_lock();
4046         mem->move_charge_at_immigrate = val;
4047         cgroup_unlock();
4048
4049         return 0;
4050 }
4051 #else
4052 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4053                                         struct cftype *cft, u64 val)
4054 {
4055         return -ENOSYS;
4056 }
4057 #endif
4058
4059
4060 /* For read statistics */
4061 enum {
4062         MCS_CACHE,
4063         MCS_RSS,
4064         MCS_FILE_MAPPED,
4065         MCS_PGPGIN,
4066         MCS_PGPGOUT,
4067         MCS_SWAP,
4068         MCS_PGFAULT,
4069         MCS_PGMAJFAULT,
4070         MCS_INACTIVE_ANON,
4071         MCS_ACTIVE_ANON,
4072         MCS_INACTIVE_FILE,
4073         MCS_ACTIVE_FILE,
4074         MCS_UNEVICTABLE,
4075         NR_MCS_STAT,
4076 };
4077
4078 struct mcs_total_stat {
4079         s64 stat[NR_MCS_STAT];
4080 };
4081
4082 struct {
4083         char *local_name;
4084         char *total_name;
4085 } memcg_stat_strings[NR_MCS_STAT] = {
4086         {"cache", "total_cache"},
4087         {"rss", "total_rss"},
4088         {"mapped_file", "total_mapped_file"},
4089         {"pgpgin", "total_pgpgin"},
4090         {"pgpgout", "total_pgpgout"},
4091         {"swap", "total_swap"},
4092         {"pgfault", "total_pgfault"},
4093         {"pgmajfault", "total_pgmajfault"},
4094         {"inactive_anon", "total_inactive_anon"},
4095         {"active_anon", "total_active_anon"},
4096         {"inactive_file", "total_inactive_file"},
4097         {"active_file", "total_active_file"},
4098         {"unevictable", "total_unevictable"}
4099 };
4100
4101
4102 static void
4103 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4104 {
4105         s64 val;
4106
4107         /* per cpu stat */
4108         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4109         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4110         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4111         s->stat[MCS_RSS] += val * PAGE_SIZE;
4112         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4113         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4114         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4115         s->stat[MCS_PGPGIN] += val;
4116         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4117         s->stat[MCS_PGPGOUT] += val;
4118         if (do_swap_account) {
4119                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4120                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4121         }
4122         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4123         s->stat[MCS_PGFAULT] += val;
4124         val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4125         s->stat[MCS_PGMAJFAULT] += val;
4126
4127         /* per zone stat */
4128         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4129         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4130         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4131         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4132         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4133         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4134         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4135         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4136         val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4137         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4138 }
4139
4140 static void
4141 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4142 {
4143         struct mem_cgroup *iter;
4144
4145         for_each_mem_cgroup_tree(iter, mem)
4146                 mem_cgroup_get_local_stat(iter, s);
4147 }
4148
4149 #ifdef CONFIG_NUMA
4150 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4151 {
4152         int nid;
4153         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4154         unsigned long node_nr;
4155         struct cgroup *cont = m->private;
4156         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4157
4158         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4159         seq_printf(m, "total=%lu", total_nr);
4160         for_each_node_state(nid, N_HIGH_MEMORY) {
4161                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4162                 seq_printf(m, " N%d=%lu", nid, node_nr);
4163         }
4164         seq_putc(m, '\n');
4165
4166         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4167         seq_printf(m, "file=%lu", file_nr);
4168         for_each_node_state(nid, N_HIGH_MEMORY) {
4169                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4170                                 LRU_ALL_FILE);
4171                 seq_printf(m, " N%d=%lu", nid, node_nr);
4172         }
4173         seq_putc(m, '\n');
4174
4175         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4176         seq_printf(m, "anon=%lu", anon_nr);
4177         for_each_node_state(nid, N_HIGH_MEMORY) {
4178                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4179                                 LRU_ALL_ANON);
4180                 seq_printf(m, " N%d=%lu", nid, node_nr);
4181         }
4182         seq_putc(m, '\n');
4183
4184         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4185         seq_printf(m, "unevictable=%lu", unevictable_nr);
4186         for_each_node_state(nid, N_HIGH_MEMORY) {
4187                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4188                                 BIT(LRU_UNEVICTABLE));
4189                 seq_printf(m, " N%d=%lu", nid, node_nr);
4190         }
4191         seq_putc(m, '\n');
4192         return 0;
4193 }
4194 #endif /* CONFIG_NUMA */
4195
4196 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4197                                  struct cgroup_map_cb *cb)
4198 {
4199         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4200         struct mcs_total_stat mystat;
4201         int i;
4202
4203         memset(&mystat, 0, sizeof(mystat));
4204         mem_cgroup_get_local_stat(mem_cont, &mystat);
4205
4206
4207         for (i = 0; i < NR_MCS_STAT; i++) {
4208                 if (i == MCS_SWAP && !do_swap_account)
4209                         continue;
4210                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4211         }
4212
4213         /* Hierarchical information */
4214         {
4215                 unsigned long long limit, memsw_limit;
4216                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4217                 cb->fill(cb, "hierarchical_memory_limit", limit);
4218                 if (do_swap_account)
4219                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4220         }
4221
4222         memset(&mystat, 0, sizeof(mystat));
4223         mem_cgroup_get_total_stat(mem_cont, &mystat);
4224         for (i = 0; i < NR_MCS_STAT; i++) {
4225                 if (i == MCS_SWAP && !do_swap_account)
4226                         continue;
4227                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4228         }
4229
4230 #ifdef CONFIG_DEBUG_VM
4231         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4232
4233         {
4234                 int nid, zid;
4235                 struct mem_cgroup_per_zone *mz;
4236                 unsigned long recent_rotated[2] = {0, 0};
4237                 unsigned long recent_scanned[2] = {0, 0};
4238
4239                 for_each_online_node(nid)
4240                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4241                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4242
4243                                 recent_rotated[0] +=
4244                                         mz->reclaim_stat.recent_rotated[0];
4245                                 recent_rotated[1] +=
4246                                         mz->reclaim_stat.recent_rotated[1];
4247                                 recent_scanned[0] +=
4248                                         mz->reclaim_stat.recent_scanned[0];
4249                                 recent_scanned[1] +=
4250                                         mz->reclaim_stat.recent_scanned[1];
4251                         }
4252                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4253                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4254                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4255                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4256         }
4257 #endif
4258
4259         return 0;
4260 }
4261
4262 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4263 {
4264         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4265
4266         return mem_cgroup_swappiness(memcg);
4267 }
4268
4269 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4270                                        u64 val)
4271 {
4272         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4273         struct mem_cgroup *parent;
4274
4275         if (val > 100)
4276                 return -EINVAL;
4277
4278         if (cgrp->parent == NULL)
4279                 return -EINVAL;
4280
4281         parent = mem_cgroup_from_cont(cgrp->parent);
4282
4283         cgroup_lock();
4284
4285         /* If under hierarchy, only empty-root can set this value */
4286         if ((parent->use_hierarchy) ||
4287             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4288                 cgroup_unlock();
4289                 return -EINVAL;
4290         }
4291
4292         memcg->swappiness = val;
4293
4294         cgroup_unlock();
4295
4296         return 0;
4297 }
4298
4299 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4300 {
4301         struct mem_cgroup_threshold_ary *t;
4302         u64 usage;
4303         int i;
4304
4305         rcu_read_lock();
4306         if (!swap)
4307                 t = rcu_dereference(memcg->thresholds.primary);
4308         else
4309                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4310
4311         if (!t)
4312                 goto unlock;
4313
4314         usage = mem_cgroup_usage(memcg, swap);
4315
4316         /*
4317          * current_threshold points to threshold just below usage.
4318          * If it's not true, a threshold was crossed after last
4319          * call of __mem_cgroup_threshold().
4320          */
4321         i = t->current_threshold;
4322
4323         /*
4324          * Iterate backward over array of thresholds starting from
4325          * current_threshold and check if a threshold is crossed.
4326          * If none of thresholds below usage is crossed, we read
4327          * only one element of the array here.
4328          */
4329         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4330                 eventfd_signal(t->entries[i].eventfd, 1);
4331
4332         /* i = current_threshold + 1 */
4333         i++;
4334
4335         /*
4336          * Iterate forward over array of thresholds starting from
4337          * current_threshold+1 and check if a threshold is crossed.
4338          * If none of thresholds above usage is crossed, we read
4339          * only one element of the array here.
4340          */
4341         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4342                 eventfd_signal(t->entries[i].eventfd, 1);
4343
4344         /* Update current_threshold */
4345         t->current_threshold = i - 1;
4346 unlock:
4347         rcu_read_unlock();
4348 }
4349
4350 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4351 {
4352         while (memcg) {
4353                 __mem_cgroup_threshold(memcg, false);
4354                 if (do_swap_account)
4355                         __mem_cgroup_threshold(memcg, true);
4356
4357                 memcg = parent_mem_cgroup(memcg);
4358         }
4359 }
4360
4361 static int compare_thresholds(const void *a, const void *b)
4362 {
4363         const struct mem_cgroup_threshold *_a = a;
4364         const struct mem_cgroup_threshold *_b = b;
4365
4366         return _a->threshold - _b->threshold;
4367 }
4368
4369 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4370 {
4371         struct mem_cgroup_eventfd_list *ev;
4372
4373         list_for_each_entry(ev, &mem->oom_notify, list)
4374                 eventfd_signal(ev->eventfd, 1);
4375         return 0;
4376 }
4377
4378 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4379 {
4380         struct mem_cgroup *iter;
4381
4382         for_each_mem_cgroup_tree(iter, mem)
4383                 mem_cgroup_oom_notify_cb(iter);
4384 }
4385
4386 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4387         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4388 {
4389         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4390         struct mem_cgroup_thresholds *thresholds;
4391         struct mem_cgroup_threshold_ary *new;
4392         int type = MEMFILE_TYPE(cft->private);
4393         u64 threshold, usage;
4394         int i, size, ret;
4395
4396         ret = res_counter_memparse_write_strategy(args, &threshold);
4397         if (ret)
4398                 return ret;
4399
4400         mutex_lock(&memcg->thresholds_lock);
4401
4402         if (type == _MEM)
4403                 thresholds = &memcg->thresholds;
4404         else if (type == _MEMSWAP)
4405                 thresholds = &memcg->memsw_thresholds;
4406         else
4407                 BUG();
4408
4409         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4410
4411         /* Check if a threshold crossed before adding a new one */
4412         if (thresholds->primary)
4413                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4414
4415         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4416
4417         /* Allocate memory for new array of thresholds */
4418         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4419                         GFP_KERNEL);
4420         if (!new) {
4421                 ret = -ENOMEM;
4422                 goto unlock;
4423         }
4424         new->size = size;
4425
4426         /* Copy thresholds (if any) to new array */
4427         if (thresholds->primary) {
4428                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4429                                 sizeof(struct mem_cgroup_threshold));
4430         }
4431
4432         /* Add new threshold */
4433         new->entries[size - 1].eventfd = eventfd;
4434         new->entries[size - 1].threshold = threshold;
4435
4436         /* Sort thresholds. Registering of new threshold isn't time-critical */
4437         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4438                         compare_thresholds, NULL);
4439
4440         /* Find current threshold */
4441         new->current_threshold = -1;
4442         for (i = 0; i < size; i++) {
4443                 if (new->entries[i].threshold < usage) {
4444                         /*
4445                          * new->current_threshold will not be used until
4446                          * rcu_assign_pointer(), so it's safe to increment
4447                          * it here.
4448                          */
4449                         ++new->current_threshold;
4450                 }
4451         }
4452
4453         /* Free old spare buffer and save old primary buffer as spare */
4454         kfree(thresholds->spare);
4455         thresholds->spare = thresholds->primary;
4456
4457         rcu_assign_pointer(thresholds->primary, new);
4458
4459         /* To be sure that nobody uses thresholds */
4460         synchronize_rcu();
4461
4462 unlock:
4463         mutex_unlock(&memcg->thresholds_lock);
4464
4465         return ret;
4466 }
4467
4468 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4469         struct cftype *cft, struct eventfd_ctx *eventfd)
4470 {
4471         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4472         struct mem_cgroup_thresholds *thresholds;
4473         struct mem_cgroup_threshold_ary *new;
4474         int type = MEMFILE_TYPE(cft->private);
4475         u64 usage;
4476         int i, j, size;
4477
4478         mutex_lock(&memcg->thresholds_lock);
4479         if (type == _MEM)
4480                 thresholds = &memcg->thresholds;
4481         else if (type == _MEMSWAP)
4482                 thresholds = &memcg->memsw_thresholds;
4483         else
4484                 BUG();
4485
4486         /*
4487          * Something went wrong if we trying to unregister a threshold
4488          * if we don't have thresholds
4489          */
4490         BUG_ON(!thresholds);
4491
4492         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4493
4494         /* Check if a threshold crossed before removing */
4495         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4496
4497         /* Calculate new number of threshold */
4498         size = 0;
4499         for (i = 0; i < thresholds->primary->size; i++) {
4500                 if (thresholds->primary->entries[i].eventfd != eventfd)
4501                         size++;
4502         }
4503
4504         new = thresholds->spare;
4505
4506         /* Set thresholds array to NULL if we don't have thresholds */
4507         if (!size) {
4508                 kfree(new);
4509                 new = NULL;
4510                 goto swap_buffers;
4511         }
4512
4513         new->size = size;
4514
4515         /* Copy thresholds and find current threshold */
4516         new->current_threshold = -1;
4517         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4518                 if (thresholds->primary->entries[i].eventfd == eventfd)
4519                         continue;
4520
4521                 new->entries[j] = thresholds->primary->entries[i];
4522                 if (new->entries[j].threshold < usage) {
4523                         /*
4524                          * new->current_threshold will not be used
4525                          * until rcu_assign_pointer(), so it's safe to increment
4526                          * it here.
4527                          */
4528                         ++new->current_threshold;
4529                 }
4530                 j++;
4531         }
4532
4533 swap_buffers:
4534         /* Swap primary and spare array */
4535         thresholds->spare = thresholds->primary;
4536         rcu_assign_pointer(thresholds->primary, new);
4537
4538         /* To be sure that nobody uses thresholds */
4539         synchronize_rcu();
4540
4541         mutex_unlock(&memcg->thresholds_lock);
4542 }
4543
4544 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4545         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4546 {
4547         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4548         struct mem_cgroup_eventfd_list *event;
4549         int type = MEMFILE_TYPE(cft->private);
4550
4551         BUG_ON(type != _OOM_TYPE);
4552         event = kmalloc(sizeof(*event), GFP_KERNEL);
4553         if (!event)
4554                 return -ENOMEM;
4555
4556         spin_lock(&memcg_oom_lock);
4557
4558         event->eventfd = eventfd;
4559         list_add(&event->list, &memcg->oom_notify);
4560
4561         /* already in OOM ? */
4562         if (atomic_read(&memcg->under_oom))
4563                 eventfd_signal(eventfd, 1);
4564         spin_unlock(&memcg_oom_lock);
4565
4566         return 0;
4567 }
4568
4569 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4570         struct cftype *cft, struct eventfd_ctx *eventfd)
4571 {
4572         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4573         struct mem_cgroup_eventfd_list *ev, *tmp;
4574         int type = MEMFILE_TYPE(cft->private);
4575
4576         BUG_ON(type != _OOM_TYPE);
4577
4578         spin_lock(&memcg_oom_lock);
4579
4580         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4581                 if (ev->eventfd == eventfd) {
4582                         list_del(&ev->list);
4583                         kfree(ev);
4584                 }
4585         }
4586
4587         spin_unlock(&memcg_oom_lock);
4588 }
4589
4590 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4591         struct cftype *cft,  struct cgroup_map_cb *cb)
4592 {
4593         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4594
4595         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4596
4597         if (atomic_read(&mem->under_oom))
4598                 cb->fill(cb, "under_oom", 1);
4599         else
4600                 cb->fill(cb, "under_oom", 0);
4601         return 0;
4602 }
4603
4604 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4605         struct cftype *cft, u64 val)
4606 {
4607         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4608         struct mem_cgroup *parent;
4609
4610         /* cannot set to root cgroup and only 0 and 1 are allowed */
4611         if (!cgrp->parent || !((val == 0) || (val == 1)))
4612                 return -EINVAL;
4613
4614         parent = mem_cgroup_from_cont(cgrp->parent);
4615
4616         cgroup_lock();
4617         /* oom-kill-disable is a flag for subhierarchy. */
4618         if ((parent->use_hierarchy) ||
4619             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4620                 cgroup_unlock();
4621                 return -EINVAL;
4622         }
4623         mem->oom_kill_disable = val;
4624         if (!val)
4625                 memcg_oom_recover(mem);
4626         cgroup_unlock();
4627         return 0;
4628 }
4629
4630 #ifdef CONFIG_NUMA
4631 static const struct file_operations mem_control_numa_stat_file_operations = {
4632         .read = seq_read,
4633         .llseek = seq_lseek,
4634         .release = single_release,
4635 };
4636
4637 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4638 {
4639         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4640
4641         file->f_op = &mem_control_numa_stat_file_operations;
4642         return single_open(file, mem_control_numa_stat_show, cont);
4643 }
4644 #endif /* CONFIG_NUMA */
4645
4646 static struct cftype mem_cgroup_files[] = {
4647         {
4648                 .name = "usage_in_bytes",
4649                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4650                 .read_u64 = mem_cgroup_read,
4651                 .register_event = mem_cgroup_usage_register_event,
4652                 .unregister_event = mem_cgroup_usage_unregister_event,
4653         },
4654         {
4655                 .name = "max_usage_in_bytes",
4656                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4657                 .trigger = mem_cgroup_reset,
4658                 .read_u64 = mem_cgroup_read,
4659         },
4660         {
4661                 .name = "limit_in_bytes",
4662                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4663                 .write_string = mem_cgroup_write,
4664                 .read_u64 = mem_cgroup_read,
4665         },
4666         {
4667                 .name = "soft_limit_in_bytes",
4668                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4669                 .write_string = mem_cgroup_write,
4670                 .read_u64 = mem_cgroup_read,
4671         },
4672         {
4673                 .name = "failcnt",
4674                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4675                 .trigger = mem_cgroup_reset,
4676                 .read_u64 = mem_cgroup_read,
4677         },
4678         {
4679                 .name = "stat",
4680                 .read_map = mem_control_stat_show,
4681         },
4682         {
4683                 .name = "force_empty",
4684                 .trigger = mem_cgroup_force_empty_write,
4685         },
4686         {
4687                 .name = "use_hierarchy",
4688                 .write_u64 = mem_cgroup_hierarchy_write,
4689                 .read_u64 = mem_cgroup_hierarchy_read,
4690         },
4691         {
4692                 .name = "swappiness",
4693                 .read_u64 = mem_cgroup_swappiness_read,
4694                 .write_u64 = mem_cgroup_swappiness_write,
4695         },
4696         {
4697                 .name = "move_charge_at_immigrate",
4698                 .read_u64 = mem_cgroup_move_charge_read,
4699                 .write_u64 = mem_cgroup_move_charge_write,
4700         },
4701         {
4702                 .name = "oom_control",
4703                 .read_map = mem_cgroup_oom_control_read,
4704                 .write_u64 = mem_cgroup_oom_control_write,
4705                 .register_event = mem_cgroup_oom_register_event,
4706                 .unregister_event = mem_cgroup_oom_unregister_event,
4707                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4708         },
4709 #ifdef CONFIG_NUMA
4710         {
4711                 .name = "numa_stat",
4712                 .open = mem_control_numa_stat_open,
4713                 .mode = S_IRUGO,
4714         },
4715 #endif
4716 };
4717
4718 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4719 static struct cftype memsw_cgroup_files[] = {
4720         {
4721                 .name = "memsw.usage_in_bytes",
4722                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4723                 .read_u64 = mem_cgroup_read,
4724                 .register_event = mem_cgroup_usage_register_event,
4725                 .unregister_event = mem_cgroup_usage_unregister_event,
4726         },
4727         {
4728                 .name = "memsw.max_usage_in_bytes",
4729                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4730                 .trigger = mem_cgroup_reset,
4731                 .read_u64 = mem_cgroup_read,
4732         },
4733         {
4734                 .name = "memsw.limit_in_bytes",
4735                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4736                 .write_string = mem_cgroup_write,
4737                 .read_u64 = mem_cgroup_read,
4738         },
4739         {
4740                 .name = "memsw.failcnt",
4741                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4742                 .trigger = mem_cgroup_reset,
4743                 .read_u64 = mem_cgroup_read,
4744         },
4745 };
4746
4747 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4748 {
4749         if (!do_swap_account)
4750                 return 0;
4751         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4752                                 ARRAY_SIZE(memsw_cgroup_files));
4753 };
4754 #else
4755 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4756 {
4757         return 0;
4758 }
4759 #endif
4760
4761 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4762 {
4763         struct mem_cgroup_per_node *pn;
4764         struct mem_cgroup_per_zone *mz;
4765         enum lru_list l;
4766         int zone, tmp = node;
4767         /*
4768          * This routine is called against possible nodes.
4769          * But it's BUG to call kmalloc() against offline node.
4770          *
4771          * TODO: this routine can waste much memory for nodes which will
4772          *       never be onlined. It's better to use memory hotplug callback
4773          *       function.
4774          */
4775         if (!node_state(node, N_NORMAL_MEMORY))
4776                 tmp = -1;
4777         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4778         if (!pn)
4779                 return 1;
4780
4781         mem->info.nodeinfo[node] = pn;
4782         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4783                 mz = &pn->zoneinfo[zone];
4784                 for_each_lru(l)
4785                         INIT_LIST_HEAD(&mz->lists[l]);
4786                 mz->usage_in_excess = 0;
4787                 mz->on_tree = false;
4788                 mz->mem = mem;
4789         }
4790         return 0;
4791 }
4792
4793 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4794 {
4795         kfree(mem->info.nodeinfo[node]);
4796 }
4797
4798 static struct mem_cgroup *mem_cgroup_alloc(void)
4799 {
4800         struct mem_cgroup *mem;
4801         int size = sizeof(struct mem_cgroup);
4802
4803         /* Can be very big if MAX_NUMNODES is very big */
4804         if (size < PAGE_SIZE)
4805                 mem = kzalloc(size, GFP_KERNEL);
4806         else
4807                 mem = vzalloc(size);
4808
4809         if (!mem)
4810                 return NULL;
4811
4812         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4813         if (!mem->stat)
4814                 goto out_free;
4815         spin_lock_init(&mem->pcp_counter_lock);
4816         return mem;
4817
4818 out_free:
4819         if (size < PAGE_SIZE)
4820                 kfree(mem);
4821         else
4822                 vfree(mem);
4823         return NULL;
4824 }
4825
4826 /*
4827  * At destroying mem_cgroup, references from swap_cgroup can remain.
4828  * (scanning all at force_empty is too costly...)
4829  *
4830  * Instead of clearing all references at force_empty, we remember
4831  * the number of reference from swap_cgroup and free mem_cgroup when
4832  * it goes down to 0.
4833  *
4834  * Removal of cgroup itself succeeds regardless of refs from swap.
4835  */
4836
4837 static void __mem_cgroup_free(struct mem_cgroup *mem)
4838 {
4839         int node;
4840
4841         mem_cgroup_remove_from_trees(mem);
4842         free_css_id(&mem_cgroup_subsys, &mem->css);
4843
4844         for_each_node_state(node, N_POSSIBLE)
4845                 free_mem_cgroup_per_zone_info(mem, node);
4846
4847         free_percpu(mem->stat);
4848         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4849                 kfree(mem);
4850         else
4851                 vfree(mem);
4852 }
4853
4854 static void mem_cgroup_get(struct mem_cgroup *mem)
4855 {
4856         atomic_inc(&mem->refcnt);
4857 }
4858
4859 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4860 {
4861         if (atomic_sub_and_test(count, &mem->refcnt)) {
4862                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4863                 __mem_cgroup_free(mem);
4864                 if (parent)
4865                         mem_cgroup_put(parent);
4866         }
4867 }
4868
4869 static void mem_cgroup_put(struct mem_cgroup *mem)
4870 {
4871         __mem_cgroup_put(mem, 1);
4872 }
4873
4874 /*
4875  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4876  */
4877 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4878 {
4879         if (!mem->res.parent)
4880                 return NULL;
4881         return mem_cgroup_from_res_counter(mem->res.parent, res);
4882 }
4883
4884 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4885 static void __init enable_swap_cgroup(void)
4886 {
4887         if (!mem_cgroup_disabled() && really_do_swap_account)
4888                 do_swap_account = 1;
4889 }
4890 #else
4891 static void __init enable_swap_cgroup(void)
4892 {
4893 }
4894 #endif
4895
4896 static int mem_cgroup_soft_limit_tree_init(void)
4897 {
4898         struct mem_cgroup_tree_per_node *rtpn;
4899         struct mem_cgroup_tree_per_zone *rtpz;
4900         int tmp, node, zone;
4901
4902         for_each_node_state(node, N_POSSIBLE) {
4903                 tmp = node;
4904                 if (!node_state(node, N_NORMAL_MEMORY))
4905                         tmp = -1;
4906                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4907                 if (!rtpn)
4908                         return 1;
4909
4910                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4911
4912                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4913                         rtpz = &rtpn->rb_tree_per_zone[zone];
4914                         rtpz->rb_root = RB_ROOT;
4915                         spin_lock_init(&rtpz->lock);
4916                 }
4917         }
4918         return 0;
4919 }
4920
4921 static struct cgroup_subsys_state * __ref
4922 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4923 {
4924         struct mem_cgroup *mem, *parent;
4925         long error = -ENOMEM;
4926         int node;
4927
4928         mem = mem_cgroup_alloc();
4929         if (!mem)
4930                 return ERR_PTR(error);
4931
4932         for_each_node_state(node, N_POSSIBLE)
4933                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4934                         goto free_out;
4935
4936         /* root ? */
4937         if (cont->parent == NULL) {
4938                 int cpu;
4939                 enable_swap_cgroup();
4940                 parent = NULL;
4941                 root_mem_cgroup = mem;
4942                 if (mem_cgroup_soft_limit_tree_init())
4943                         goto free_out;
4944                 for_each_possible_cpu(cpu) {
4945                         struct memcg_stock_pcp *stock =
4946                                                 &per_cpu(memcg_stock, cpu);
4947                         INIT_WORK(&stock->work, drain_local_stock);
4948                 }
4949                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4950         } else {
4951                 parent = mem_cgroup_from_cont(cont->parent);
4952                 mem->use_hierarchy = parent->use_hierarchy;
4953                 mem->oom_kill_disable = parent->oom_kill_disable;
4954         }
4955
4956         if (parent && parent->use_hierarchy) {
4957                 res_counter_init(&mem->res, &parent->res);
4958                 res_counter_init(&mem->memsw, &parent->memsw);
4959                 /*
4960                  * We increment refcnt of the parent to ensure that we can
4961                  * safely access it on res_counter_charge/uncharge.
4962                  * This refcnt will be decremented when freeing this
4963                  * mem_cgroup(see mem_cgroup_put).
4964                  */
4965                 mem_cgroup_get(parent);
4966         } else {
4967                 res_counter_init(&mem->res, NULL);
4968                 res_counter_init(&mem->memsw, NULL);
4969         }
4970         mem->last_scanned_child = 0;
4971         mem->last_scanned_node = MAX_NUMNODES;
4972         INIT_LIST_HEAD(&mem->oom_notify);
4973
4974         if (parent)
4975                 mem->swappiness = mem_cgroup_swappiness(parent);
4976         atomic_set(&mem->refcnt, 1);
4977         mem->move_charge_at_immigrate = 0;
4978         mutex_init(&mem->thresholds_lock);
4979         return &mem->css;
4980 free_out:
4981         __mem_cgroup_free(mem);
4982         root_mem_cgroup = NULL;
4983         return ERR_PTR(error);
4984 }
4985
4986 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4987                                         struct cgroup *cont)
4988 {
4989         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4990
4991         return mem_cgroup_force_empty(mem, false);
4992 }
4993
4994 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4995                                 struct cgroup *cont)
4996 {
4997         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4998
4999         mem_cgroup_put(mem);
5000 }
5001
5002 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5003                                 struct cgroup *cont)
5004 {
5005         int ret;
5006
5007         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5008                                 ARRAY_SIZE(mem_cgroup_files));
5009
5010         if (!ret)
5011                 ret = register_memsw_files(cont, ss);
5012         return ret;
5013 }
5014
5015 #ifdef CONFIG_MMU
5016 /* Handlers for move charge at task migration. */
5017 #define PRECHARGE_COUNT_AT_ONCE 256
5018 static int mem_cgroup_do_precharge(unsigned long count)
5019 {
5020         int ret = 0;
5021         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5022         struct mem_cgroup *mem = mc.to;
5023
5024         if (mem_cgroup_is_root(mem)) {
5025                 mc.precharge += count;
5026                 /* we don't need css_get for root */
5027                 return ret;
5028         }
5029         /* try to charge at once */
5030         if (count > 1) {
5031                 struct res_counter *dummy;
5032                 /*
5033                  * "mem" cannot be under rmdir() because we've already checked
5034                  * by cgroup_lock_live_cgroup() that it is not removed and we
5035                  * are still under the same cgroup_mutex. So we can postpone
5036                  * css_get().
5037                  */
5038                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5039                         goto one_by_one;
5040                 if (do_swap_account && res_counter_charge(&mem->memsw,
5041                                                 PAGE_SIZE * count, &dummy)) {
5042                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5043                         goto one_by_one;
5044                 }
5045                 mc.precharge += count;
5046                 return ret;
5047         }
5048 one_by_one:
5049         /* fall back to one by one charge */
5050         while (count--) {
5051                 if (signal_pending(current)) {
5052                         ret = -EINTR;
5053                         break;
5054                 }
5055                 if (!batch_count--) {
5056                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5057                         cond_resched();
5058                 }
5059                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5060                 if (ret || !mem)
5061                         /* mem_cgroup_clear_mc() will do uncharge later */
5062                         return -ENOMEM;
5063                 mc.precharge++;
5064         }
5065         return ret;
5066 }
5067
5068 /**
5069  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5070  * @vma: the vma the pte to be checked belongs
5071  * @addr: the address corresponding to the pte to be checked
5072  * @ptent: the pte to be checked
5073  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5074  *
5075  * Returns
5076  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5077  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5078  *     move charge. if @target is not NULL, the page is stored in target->page
5079  *     with extra refcnt got(Callers should handle it).
5080  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5081  *     target for charge migration. if @target is not NULL, the entry is stored
5082  *     in target->ent.
5083  *
5084  * Called with pte lock held.
5085  */
5086 union mc_target {
5087         struct page     *page;
5088         swp_entry_t     ent;
5089 };
5090
5091 enum mc_target_type {
5092         MC_TARGET_NONE, /* not used */
5093         MC_TARGET_PAGE,
5094         MC_TARGET_SWAP,
5095 };
5096
5097 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5098                                                 unsigned long addr, pte_t ptent)
5099 {
5100         struct page *page = vm_normal_page(vma, addr, ptent);
5101
5102         if (!page || !page_mapped(page))
5103                 return NULL;
5104         if (PageAnon(page)) {
5105                 /* we don't move shared anon */
5106                 if (!move_anon() || page_mapcount(page) > 2)
5107                         return NULL;
5108         } else if (!move_file())
5109                 /* we ignore mapcount for file pages */
5110                 return NULL;
5111         if (!get_page_unless_zero(page))
5112                 return NULL;
5113
5114         return page;
5115 }
5116
5117 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5118                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5119 {
5120         int usage_count;
5121         struct page *page = NULL;
5122         swp_entry_t ent = pte_to_swp_entry(ptent);
5123
5124         if (!move_anon() || non_swap_entry(ent))
5125                 return NULL;
5126         usage_count = mem_cgroup_count_swap_user(ent, &page);
5127         if (usage_count > 1) { /* we don't move shared anon */
5128                 if (page)
5129                         put_page(page);
5130                 return NULL;
5131         }
5132         if (do_swap_account)
5133                 entry->val = ent.val;
5134
5135         return page;
5136 }
5137
5138 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5139                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5140 {
5141         struct page *page = NULL;
5142         struct inode *inode;
5143         struct address_space *mapping;
5144         pgoff_t pgoff;
5145
5146         if (!vma->vm_file) /* anonymous vma */
5147                 return NULL;
5148         if (!move_file())
5149                 return NULL;
5150
5151         inode = vma->vm_file->f_path.dentry->d_inode;
5152         mapping = vma->vm_file->f_mapping;
5153         if (pte_none(ptent))
5154                 pgoff = linear_page_index(vma, addr);
5155         else /* pte_file(ptent) is true */
5156                 pgoff = pte_to_pgoff(ptent);
5157
5158         /* page is moved even if it's not RSS of this task(page-faulted). */
5159         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5160                 page = find_get_page(mapping, pgoff);
5161         } else { /* shmem/tmpfs file. we should take account of swap too. */
5162                 swp_entry_t ent;
5163                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5164                 if (do_swap_account)
5165                         entry->val = ent.val;
5166         }
5167
5168         return page;
5169 }
5170
5171 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5172                 unsigned long addr, pte_t ptent, union mc_target *target)
5173 {
5174         struct page *page = NULL;
5175         struct page_cgroup *pc;
5176         int ret = 0;
5177         swp_entry_t ent = { .val = 0 };
5178
5179         if (pte_present(ptent))
5180                 page = mc_handle_present_pte(vma, addr, ptent);
5181         else if (is_swap_pte(ptent))
5182                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5183         else if (pte_none(ptent) || pte_file(ptent))
5184                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5185
5186         if (!page && !ent.val)
5187                 return 0;
5188         if (page) {
5189                 pc = lookup_page_cgroup(page);
5190                 /*
5191                  * Do only loose check w/o page_cgroup lock.
5192                  * mem_cgroup_move_account() checks the pc is valid or not under
5193                  * the lock.
5194                  */
5195                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5196                         ret = MC_TARGET_PAGE;
5197                         if (target)
5198                                 target->page = page;
5199                 }
5200                 if (!ret || !target)
5201                         put_page(page);
5202         }
5203         /* There is a swap entry and a page doesn't exist or isn't charged */
5204         if (ent.val && !ret &&
5205                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5206                 ret = MC_TARGET_SWAP;
5207                 if (target)
5208                         target->ent = ent;
5209         }
5210         return ret;
5211 }
5212
5213 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5214                                         unsigned long addr, unsigned long end,
5215                                         struct mm_walk *walk)
5216 {
5217         struct vm_area_struct *vma = walk->private;
5218         pte_t *pte;
5219         spinlock_t *ptl;
5220
5221         split_huge_page_pmd(walk->mm, pmd);
5222
5223         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5224         for (; addr != end; pte++, addr += PAGE_SIZE)
5225                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5226                         mc.precharge++; /* increment precharge temporarily */
5227         pte_unmap_unlock(pte - 1, ptl);
5228         cond_resched();
5229
5230         return 0;
5231 }
5232
5233 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5234 {
5235         unsigned long precharge;
5236         struct vm_area_struct *vma;
5237
5238         down_read(&mm->mmap_sem);
5239         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5240                 struct mm_walk mem_cgroup_count_precharge_walk = {
5241                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5242                         .mm = mm,
5243                         .private = vma,
5244                 };
5245                 if (is_vm_hugetlb_page(vma))
5246                         continue;
5247                 walk_page_range(vma->vm_start, vma->vm_end,
5248                                         &mem_cgroup_count_precharge_walk);
5249         }
5250         up_read(&mm->mmap_sem);
5251
5252         precharge = mc.precharge;
5253         mc.precharge = 0;
5254
5255         return precharge;
5256 }
5257
5258 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5259 {
5260         unsigned long precharge = mem_cgroup_count_precharge(mm);
5261
5262         VM_BUG_ON(mc.moving_task);
5263         mc.moving_task = current;
5264         return mem_cgroup_do_precharge(precharge);
5265 }
5266
5267 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5268 static void __mem_cgroup_clear_mc(void)
5269 {
5270         struct mem_cgroup *from = mc.from;
5271         struct mem_cgroup *to = mc.to;
5272
5273         /* we must uncharge all the leftover precharges from mc.to */
5274         if (mc.precharge) {
5275                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5276                 mc.precharge = 0;
5277         }
5278         /*
5279          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5280          * we must uncharge here.
5281          */
5282         if (mc.moved_charge) {
5283                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5284                 mc.moved_charge = 0;
5285         }
5286         /* we must fixup refcnts and charges */
5287         if (mc.moved_swap) {
5288                 /* uncharge swap account from the old cgroup */
5289                 if (!mem_cgroup_is_root(mc.from))
5290                         res_counter_uncharge(&mc.from->memsw,
5291                                                 PAGE_SIZE * mc.moved_swap);
5292                 __mem_cgroup_put(mc.from, mc.moved_swap);
5293
5294                 if (!mem_cgroup_is_root(mc.to)) {
5295                         /*
5296                          * we charged both to->res and to->memsw, so we should
5297                          * uncharge to->res.
5298                          */
5299                         res_counter_uncharge(&mc.to->res,
5300                                                 PAGE_SIZE * mc.moved_swap);
5301                 }
5302                 /* we've already done mem_cgroup_get(mc.to) */
5303                 mc.moved_swap = 0;
5304         }
5305         memcg_oom_recover(from);
5306         memcg_oom_recover(to);
5307         wake_up_all(&mc.waitq);
5308 }
5309
5310 static void mem_cgroup_clear_mc(void)
5311 {
5312         struct mem_cgroup *from = mc.from;
5313
5314         /*
5315          * we must clear moving_task before waking up waiters at the end of
5316          * task migration.
5317          */
5318         mc.moving_task = NULL;
5319         __mem_cgroup_clear_mc();
5320         spin_lock(&mc.lock);
5321         mc.from = NULL;
5322         mc.to = NULL;
5323         spin_unlock(&mc.lock);
5324         mem_cgroup_end_move(from);
5325 }
5326
5327 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5328                                 struct cgroup *cgroup,
5329                                 struct task_struct *p)
5330 {
5331         int ret = 0;
5332         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5333
5334         if (mem->move_charge_at_immigrate) {
5335                 struct mm_struct *mm;
5336                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5337
5338                 VM_BUG_ON(from == mem);
5339
5340                 mm = get_task_mm(p);
5341                 if (!mm)
5342                         return 0;
5343                 /* We move charges only when we move a owner of the mm */
5344                 if (mm->owner == p) {
5345                         VM_BUG_ON(mc.from);
5346                         VM_BUG_ON(mc.to);
5347                         VM_BUG_ON(mc.precharge);
5348                         VM_BUG_ON(mc.moved_charge);
5349                         VM_BUG_ON(mc.moved_swap);
5350                         mem_cgroup_start_move(from);
5351                         spin_lock(&mc.lock);
5352                         mc.from = from;
5353                         mc.to = mem;
5354                         spin_unlock(&mc.lock);
5355                         /* We set mc.moving_task later */
5356
5357                         ret = mem_cgroup_precharge_mc(mm);
5358                         if (ret)
5359                                 mem_cgroup_clear_mc();
5360                 }
5361                 mmput(mm);
5362         }
5363         return ret;
5364 }
5365
5366 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5367                                 struct cgroup *cgroup,
5368                                 struct task_struct *p)
5369 {
5370         mem_cgroup_clear_mc();
5371 }
5372
5373 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5374                                 unsigned long addr, unsigned long end,
5375                                 struct mm_walk *walk)
5376 {
5377         int ret = 0;
5378         struct vm_area_struct *vma = walk->private;
5379         pte_t *pte;
5380         spinlock_t *ptl;
5381
5382         split_huge_page_pmd(walk->mm, pmd);
5383 retry:
5384         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5385         for (; addr != end; addr += PAGE_SIZE) {
5386                 pte_t ptent = *(pte++);
5387                 union mc_target target;
5388                 int type;
5389                 struct page *page;
5390                 struct page_cgroup *pc;
5391                 swp_entry_t ent;
5392
5393                 if (!mc.precharge)
5394                         break;
5395
5396                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5397                 switch (type) {
5398                 case MC_TARGET_PAGE:
5399                         page = target.page;
5400                         if (isolate_lru_page(page))
5401                                 goto put;
5402                         pc = lookup_page_cgroup(page);
5403                         if (!mem_cgroup_move_account(page, 1, pc,
5404                                                      mc.from, mc.to, false)) {
5405                                 mc.precharge--;
5406                                 /* we uncharge from mc.from later. */
5407                                 mc.moved_charge++;
5408                         }
5409                         putback_lru_page(page);
5410 put:                    /* is_target_pte_for_mc() gets the page */
5411                         put_page(page);
5412                         break;
5413                 case MC_TARGET_SWAP:
5414                         ent = target.ent;
5415                         if (!mem_cgroup_move_swap_account(ent,
5416                                                 mc.from, mc.to, false)) {
5417                                 mc.precharge--;
5418                                 /* we fixup refcnts and charges later. */
5419                                 mc.moved_swap++;
5420                         }
5421                         break;
5422                 default:
5423                         break;
5424                 }
5425         }
5426         pte_unmap_unlock(pte - 1, ptl);
5427         cond_resched();
5428
5429         if (addr != end) {
5430                 /*
5431                  * We have consumed all precharges we got in can_attach().
5432                  * We try charge one by one, but don't do any additional
5433                  * charges to mc.to if we have failed in charge once in attach()
5434                  * phase.
5435                  */
5436                 ret = mem_cgroup_do_precharge(1);
5437                 if (!ret)
5438                         goto retry;
5439         }
5440
5441         return ret;
5442 }
5443
5444 static void mem_cgroup_move_charge(struct mm_struct *mm)
5445 {
5446         struct vm_area_struct *vma;
5447
5448         lru_add_drain_all();
5449 retry:
5450         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5451                 /*
5452                  * Someone who are holding the mmap_sem might be waiting in
5453                  * waitq. So we cancel all extra charges, wake up all waiters,
5454                  * and retry. Because we cancel precharges, we might not be able
5455                  * to move enough charges, but moving charge is a best-effort
5456                  * feature anyway, so it wouldn't be a big problem.
5457                  */
5458                 __mem_cgroup_clear_mc();
5459                 cond_resched();
5460                 goto retry;
5461         }
5462         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5463                 int ret;
5464                 struct mm_walk mem_cgroup_move_charge_walk = {
5465                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5466                         .mm = mm,
5467                         .private = vma,
5468                 };
5469                 if (is_vm_hugetlb_page(vma))
5470                         continue;
5471                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5472                                                 &mem_cgroup_move_charge_walk);
5473                 if (ret)
5474                         /*
5475                          * means we have consumed all precharges and failed in
5476                          * doing additional charge. Just abandon here.
5477                          */
5478                         break;
5479         }
5480         up_read(&mm->mmap_sem);
5481 }
5482
5483 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5484                                 struct cgroup *cont,
5485                                 struct cgroup *old_cont,
5486                                 struct task_struct *p)
5487 {
5488         struct mm_struct *mm = get_task_mm(p);
5489
5490         if (mm) {
5491                 if (mc.to)
5492                         mem_cgroup_move_charge(mm);
5493                 put_swap_token(mm);
5494                 mmput(mm);
5495         }
5496         if (mc.to)
5497                 mem_cgroup_clear_mc();
5498 }
5499 #else   /* !CONFIG_MMU */
5500 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5501                                 struct cgroup *cgroup,
5502                                 struct task_struct *p)
5503 {
5504         return 0;
5505 }
5506 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5507                                 struct cgroup *cgroup,
5508                                 struct task_struct *p)
5509 {
5510 }
5511 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5512                                 struct cgroup *cont,
5513                                 struct cgroup *old_cont,
5514                                 struct task_struct *p)
5515 {
5516 }
5517 #endif
5518
5519 struct cgroup_subsys mem_cgroup_subsys = {
5520         .name = "memory",
5521         .subsys_id = mem_cgroup_subsys_id,
5522         .create = mem_cgroup_create,
5523         .pre_destroy = mem_cgroup_pre_destroy,
5524         .destroy = mem_cgroup_destroy,
5525         .populate = mem_cgroup_populate,
5526         .can_attach = mem_cgroup_can_attach,
5527         .cancel_attach = mem_cgroup_cancel_attach,
5528         .attach = mem_cgroup_move_task,
5529         .early_init = 0,
5530         .use_id = 1,
5531 };
5532
5533 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5534 static int __init enable_swap_account(char *s)
5535 {
5536         /* consider enabled if no parameter or 1 is given */
5537         if (!strcmp(s, "1"))
5538                 really_do_swap_account = 1;
5539         else if (!strcmp(s, "0"))
5540                 really_do_swap_account = 0;
5541         return 1;
5542 }
5543 __setup("swapaccount=", enable_swap_account);
5544
5545 #endif