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