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