memcg: no uncharged pages reach page_cgroup_zoneinfo
[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  * Per memcg event counter is incremented at every pagein/pageout. This counter
78  * is used for trigger some periodic events. This is straightforward and better
79  * than using jiffies etc. to handle periodic memcg event.
80  *
81  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
82  */
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
85
86 /*
87  * Statistics for memory cgroup.
88  */
89 enum mem_cgroup_stat_index {
90         /*
91          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92          */
93         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
94         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
95         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
96         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
97         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
98         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
99         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
100         /* incremented at every  pagein/pageout */
101         MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
102         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
103
104         MEM_CGROUP_STAT_NSTATS,
105 };
106
107 struct mem_cgroup_stat_cpu {
108         s64 count[MEM_CGROUP_STAT_NSTATS];
109 };
110
111 /*
112  * per-zone information in memory controller.
113  */
114 struct mem_cgroup_per_zone {
115         /*
116          * spin_lock to protect the per cgroup LRU
117          */
118         struct list_head        lists[NR_LRU_LISTS];
119         unsigned long           count[NR_LRU_LISTS];
120
121         struct zone_reclaim_stat reclaim_stat;
122         struct rb_node          tree_node;      /* RB tree node */
123         unsigned long long      usage_in_excess;/* Set to the value by which */
124                                                 /* the soft limit is exceeded*/
125         bool                    on_tree;
126         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
127                                                 /* use container_of        */
128 };
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
131
132 struct mem_cgroup_per_node {
133         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
134 };
135
136 struct mem_cgroup_lru_info {
137         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
138 };
139
140 /*
141  * Cgroups above their limits are maintained in a RB-Tree, independent of
142  * their hierarchy representation
143  */
144
145 struct mem_cgroup_tree_per_zone {
146         struct rb_root rb_root;
147         spinlock_t lock;
148 };
149
150 struct mem_cgroup_tree_per_node {
151         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
152 };
153
154 struct mem_cgroup_tree {
155         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
156 };
157
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
159
160 struct mem_cgroup_threshold {
161         struct eventfd_ctx *eventfd;
162         u64 threshold;
163 };
164
165 /* For threshold */
166 struct mem_cgroup_threshold_ary {
167         /* An array index points to threshold just below usage. */
168         int current_threshold;
169         /* Size of entries[] */
170         unsigned int size;
171         /* Array of thresholds */
172         struct mem_cgroup_threshold entries[0];
173 };
174
175 struct mem_cgroup_thresholds {
176         /* Primary thresholds array */
177         struct mem_cgroup_threshold_ary *primary;
178         /*
179          * Spare threshold array.
180          * This is needed to make mem_cgroup_unregister_event() "never fail".
181          * It must be able to store at least primary->size - 1 entries.
182          */
183         struct mem_cgroup_threshold_ary *spare;
184 };
185
186 /* for OOM */
187 struct mem_cgroup_eventfd_list {
188         struct list_head list;
189         struct eventfd_ctx *eventfd;
190 };
191
192 static void mem_cgroup_threshold(struct mem_cgroup *mem);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
194
195 /*
196  * The memory controller data structure. The memory controller controls both
197  * page cache and RSS per cgroup. We would eventually like to provide
198  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199  * to help the administrator determine what knobs to tune.
200  *
201  * TODO: Add a water mark for the memory controller. Reclaim will begin when
202  * we hit the water mark. May be even add a low water mark, such that
203  * no reclaim occurs from a cgroup at it's low water mark, this is
204  * a feature that will be implemented much later in the future.
205  */
206 struct mem_cgroup {
207         struct cgroup_subsys_state css;
208         /*
209          * the counter to account for memory usage
210          */
211         struct res_counter res;
212         /*
213          * the counter to account for mem+swap usage.
214          */
215         struct res_counter memsw;
216         /*
217          * Per cgroup active and inactive list, similar to the
218          * per zone LRU lists.
219          */
220         struct mem_cgroup_lru_info info;
221
222         /*
223           protect against reclaim related member.
224         */
225         spinlock_t reclaim_param_lock;
226
227         /*
228          * While reclaiming in a hierarchy, we cache the last child we
229          * reclaimed from.
230          */
231         int last_scanned_child;
232         /*
233          * Should the accounting and control be hierarchical, per subtree?
234          */
235         bool use_hierarchy;
236         atomic_t        oom_lock;
237         atomic_t        refcnt;
238
239         unsigned int    swappiness;
240         /* OOM-Killer disable */
241         int             oom_kill_disable;
242
243         /* set when res.limit == memsw.limit */
244         bool            memsw_is_minimum;
245
246         /* protect arrays of thresholds */
247         struct mutex thresholds_lock;
248
249         /* thresholds for memory usage. RCU-protected */
250         struct mem_cgroup_thresholds thresholds;
251
252         /* thresholds for mem+swap usage. RCU-protected */
253         struct mem_cgroup_thresholds memsw_thresholds;
254
255         /* For oom notifier event fd */
256         struct list_head oom_notify;
257
258         /*
259          * Should we move charges of a task when a task is moved into this
260          * mem_cgroup ? And what type of charges should we move ?
261          */
262         unsigned long   move_charge_at_immigrate;
263         /*
264          * percpu counter.
265          */
266         struct mem_cgroup_stat_cpu *stat;
267         /*
268          * used when a cpu is offlined or other synchronizations
269          * See mem_cgroup_read_stat().
270          */
271         struct mem_cgroup_stat_cpu nocpu_base;
272         spinlock_t pcp_counter_lock;
273 };
274
275 /* Stuffs for move charges at task migration. */
276 /*
277  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278  * left-shifted bitmap of these types.
279  */
280 enum move_type {
281         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
282         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
283         NR_MOVE_TYPE,
284 };
285
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct {
288         spinlock_t        lock; /* for from, to */
289         struct mem_cgroup *from;
290         struct mem_cgroup *to;
291         unsigned long precharge;
292         unsigned long moved_charge;
293         unsigned long moved_swap;
294         struct task_struct *moving_task;        /* a task moving charges */
295         wait_queue_head_t waitq;                /* a waitq for other context */
296 } mc = {
297         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
298         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
299 };
300
301 static bool move_anon(void)
302 {
303         return test_bit(MOVE_CHARGE_TYPE_ANON,
304                                         &mc.to->move_charge_at_immigrate);
305 }
306
307 static bool move_file(void)
308 {
309         return test_bit(MOVE_CHARGE_TYPE_FILE,
310                                         &mc.to->move_charge_at_immigrate);
311 }
312
313 /*
314  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315  * limit reclaim to prevent infinite loops, if they ever occur.
316  */
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
319
320 enum charge_type {
321         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
322         MEM_CGROUP_CHARGE_TYPE_MAPPED,
323         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
324         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
325         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
326         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
327         NR_CHARGE_TYPE,
328 };
329
330 /* for encoding cft->private value on file */
331 #define _MEM                    (0)
332 #define _MEMSWAP                (1)
333 #define _OOM_TYPE               (2)
334 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
335 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
336 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
337 /* Used for OOM nofiier */
338 #define OOM_CONTROL             (0)
339
340 /*
341  * Reclaim flags for mem_cgroup_hierarchical_reclaim
342  */
343 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
344 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
345 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
346 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
347 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
348 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
349
350 static void mem_cgroup_get(struct mem_cgroup *mem);
351 static void mem_cgroup_put(struct mem_cgroup *mem);
352 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
353 static void drain_all_stock_async(void);
354
355 static struct mem_cgroup_per_zone *
356 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
357 {
358         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
359 }
360
361 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
362 {
363         return &mem->css;
364 }
365
366 static struct mem_cgroup_per_zone *
367 page_cgroup_zoneinfo(struct page_cgroup *pc)
368 {
369         struct mem_cgroup *mem = pc->mem_cgroup;
370         int nid = page_cgroup_nid(pc);
371         int zid = page_cgroup_zid(pc);
372
373         return mem_cgroup_zoneinfo(mem, nid, zid);
374 }
375
376 static struct mem_cgroup_tree_per_zone *
377 soft_limit_tree_node_zone(int nid, int zid)
378 {
379         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
380 }
381
382 static struct mem_cgroup_tree_per_zone *
383 soft_limit_tree_from_page(struct page *page)
384 {
385         int nid = page_to_nid(page);
386         int zid = page_zonenum(page);
387
388         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
389 }
390
391 static void
392 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
393                                 struct mem_cgroup_per_zone *mz,
394                                 struct mem_cgroup_tree_per_zone *mctz,
395                                 unsigned long long new_usage_in_excess)
396 {
397         struct rb_node **p = &mctz->rb_root.rb_node;
398         struct rb_node *parent = NULL;
399         struct mem_cgroup_per_zone *mz_node;
400
401         if (mz->on_tree)
402                 return;
403
404         mz->usage_in_excess = new_usage_in_excess;
405         if (!mz->usage_in_excess)
406                 return;
407         while (*p) {
408                 parent = *p;
409                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
410                                         tree_node);
411                 if (mz->usage_in_excess < mz_node->usage_in_excess)
412                         p = &(*p)->rb_left;
413                 /*
414                  * We can't avoid mem cgroups that are over their soft
415                  * limit by the same amount
416                  */
417                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
418                         p = &(*p)->rb_right;
419         }
420         rb_link_node(&mz->tree_node, parent, p);
421         rb_insert_color(&mz->tree_node, &mctz->rb_root);
422         mz->on_tree = true;
423 }
424
425 static void
426 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
427                                 struct mem_cgroup_per_zone *mz,
428                                 struct mem_cgroup_tree_per_zone *mctz)
429 {
430         if (!mz->on_tree)
431                 return;
432         rb_erase(&mz->tree_node, &mctz->rb_root);
433         mz->on_tree = false;
434 }
435
436 static void
437 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
438                                 struct mem_cgroup_per_zone *mz,
439                                 struct mem_cgroup_tree_per_zone *mctz)
440 {
441         spin_lock(&mctz->lock);
442         __mem_cgroup_remove_exceeded(mem, mz, mctz);
443         spin_unlock(&mctz->lock);
444 }
445
446
447 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
448 {
449         unsigned long long excess;
450         struct mem_cgroup_per_zone *mz;
451         struct mem_cgroup_tree_per_zone *mctz;
452         int nid = page_to_nid(page);
453         int zid = page_zonenum(page);
454         mctz = soft_limit_tree_from_page(page);
455
456         /*
457          * Necessary to update all ancestors when hierarchy is used.
458          * because their event counter is not touched.
459          */
460         for (; mem; mem = parent_mem_cgroup(mem)) {
461                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
462                 excess = res_counter_soft_limit_excess(&mem->res);
463                 /*
464                  * We have to update the tree if mz is on RB-tree or
465                  * mem is over its softlimit.
466                  */
467                 if (excess || mz->on_tree) {
468                         spin_lock(&mctz->lock);
469                         /* if on-tree, remove it */
470                         if (mz->on_tree)
471                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
472                         /*
473                          * Insert again. mz->usage_in_excess will be updated.
474                          * If excess is 0, no tree ops.
475                          */
476                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
477                         spin_unlock(&mctz->lock);
478                 }
479         }
480 }
481
482 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
483 {
484         int node, zone;
485         struct mem_cgroup_per_zone *mz;
486         struct mem_cgroup_tree_per_zone *mctz;
487
488         for_each_node_state(node, N_POSSIBLE) {
489                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
490                         mz = mem_cgroup_zoneinfo(mem, node, zone);
491                         mctz = soft_limit_tree_node_zone(node, zone);
492                         mem_cgroup_remove_exceeded(mem, mz, mctz);
493                 }
494         }
495 }
496
497 static struct mem_cgroup_per_zone *
498 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
499 {
500         struct rb_node *rightmost = NULL;
501         struct mem_cgroup_per_zone *mz;
502
503 retry:
504         mz = NULL;
505         rightmost = rb_last(&mctz->rb_root);
506         if (!rightmost)
507                 goto done;              /* Nothing to reclaim from */
508
509         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
510         /*
511          * Remove the node now but someone else can add it back,
512          * we will to add it back at the end of reclaim to its correct
513          * position in the tree.
514          */
515         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
516         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
517                 !css_tryget(&mz->mem->css))
518                 goto retry;
519 done:
520         return mz;
521 }
522
523 static struct mem_cgroup_per_zone *
524 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
525 {
526         struct mem_cgroup_per_zone *mz;
527
528         spin_lock(&mctz->lock);
529         mz = __mem_cgroup_largest_soft_limit_node(mctz);
530         spin_unlock(&mctz->lock);
531         return mz;
532 }
533
534 /*
535  * Implementation Note: reading percpu statistics for memcg.
536  *
537  * Both of vmstat[] and percpu_counter has threshold and do periodic
538  * synchronization to implement "quick" read. There are trade-off between
539  * reading cost and precision of value. Then, we may have a chance to implement
540  * a periodic synchronizion of counter in memcg's counter.
541  *
542  * But this _read() function is used for user interface now. The user accounts
543  * memory usage by memory cgroup and he _always_ requires exact value because
544  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
545  * have to visit all online cpus and make sum. So, for now, unnecessary
546  * synchronization is not implemented. (just implemented for cpu hotplug)
547  *
548  * If there are kernel internal actions which can make use of some not-exact
549  * value, and reading all cpu value can be performance bottleneck in some
550  * common workload, threashold and synchonization as vmstat[] should be
551  * implemented.
552  */
553 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
554                 enum mem_cgroup_stat_index idx)
555 {
556         int cpu;
557         s64 val = 0;
558
559         get_online_cpus();
560         for_each_online_cpu(cpu)
561                 val += per_cpu(mem->stat->count[idx], cpu);
562 #ifdef CONFIG_HOTPLUG_CPU
563         spin_lock(&mem->pcp_counter_lock);
564         val += mem->nocpu_base.count[idx];
565         spin_unlock(&mem->pcp_counter_lock);
566 #endif
567         put_online_cpus();
568         return val;
569 }
570
571 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
572 {
573         s64 ret;
574
575         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
576         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
577         return ret;
578 }
579
580 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
581                                          bool charge)
582 {
583         int val = (charge) ? 1 : -1;
584         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
585 }
586
587 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
588                                          bool file, int nr_pages)
589 {
590         preempt_disable();
591
592         if (file)
593                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
594         else
595                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
596
597         /* pagein of a big page is an event. So, ignore page size */
598         if (nr_pages > 0)
599                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
600         else {
601                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
602                 nr_pages = -nr_pages; /* for event */
603         }
604
605         __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages);
606
607         preempt_enable();
608 }
609
610 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
611                                         enum lru_list idx)
612 {
613         int nid, zid;
614         struct mem_cgroup_per_zone *mz;
615         u64 total = 0;
616
617         for_each_online_node(nid)
618                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
619                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
620                         total += MEM_CGROUP_ZSTAT(mz, idx);
621                 }
622         return total;
623 }
624
625 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
626 {
627         s64 val;
628
629         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
630
631         return !(val & ((1 << event_mask_shift) - 1));
632 }
633
634 /*
635  * Check events in order.
636  *
637  */
638 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
639 {
640         /* threshold event is triggered in finer grain than soft limit */
641         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
642                 mem_cgroup_threshold(mem);
643                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
644                         mem_cgroup_update_tree(mem, page);
645         }
646 }
647
648 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
649 {
650         return container_of(cgroup_subsys_state(cont,
651                                 mem_cgroup_subsys_id), struct mem_cgroup,
652                                 css);
653 }
654
655 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
656 {
657         /*
658          * mm_update_next_owner() may clear mm->owner to NULL
659          * if it races with swapoff, page migration, etc.
660          * So this can be called with p == NULL.
661          */
662         if (unlikely(!p))
663                 return NULL;
664
665         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
666                                 struct mem_cgroup, css);
667 }
668
669 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
670 {
671         struct mem_cgroup *mem = NULL;
672
673         if (!mm)
674                 return NULL;
675         /*
676          * Because we have no locks, mm->owner's may be being moved to other
677          * cgroup. We use css_tryget() here even if this looks
678          * pessimistic (rather than adding locks here).
679          */
680         rcu_read_lock();
681         do {
682                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
683                 if (unlikely(!mem))
684                         break;
685         } while (!css_tryget(&mem->css));
686         rcu_read_unlock();
687         return mem;
688 }
689
690 /* The caller has to guarantee "mem" exists before calling this */
691 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
692 {
693         struct cgroup_subsys_state *css;
694         int found;
695
696         if (!mem) /* ROOT cgroup has the smallest ID */
697                 return root_mem_cgroup; /*css_put/get against root is ignored*/
698         if (!mem->use_hierarchy) {
699                 if (css_tryget(&mem->css))
700                         return mem;
701                 return NULL;
702         }
703         rcu_read_lock();
704         /*
705          * searching a memory cgroup which has the smallest ID under given
706          * ROOT cgroup. (ID >= 1)
707          */
708         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
709         if (css && css_tryget(css))
710                 mem = container_of(css, struct mem_cgroup, css);
711         else
712                 mem = NULL;
713         rcu_read_unlock();
714         return mem;
715 }
716
717 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
718                                         struct mem_cgroup *root,
719                                         bool cond)
720 {
721         int nextid = css_id(&iter->css) + 1;
722         int found;
723         int hierarchy_used;
724         struct cgroup_subsys_state *css;
725
726         hierarchy_used = iter->use_hierarchy;
727
728         css_put(&iter->css);
729         /* If no ROOT, walk all, ignore hierarchy */
730         if (!cond || (root && !hierarchy_used))
731                 return NULL;
732
733         if (!root)
734                 root = root_mem_cgroup;
735
736         do {
737                 iter = NULL;
738                 rcu_read_lock();
739
740                 css = css_get_next(&mem_cgroup_subsys, nextid,
741                                 &root->css, &found);
742                 if (css && css_tryget(css))
743                         iter = container_of(css, struct mem_cgroup, css);
744                 rcu_read_unlock();
745                 /* If css is NULL, no more cgroups will be found */
746                 nextid = found + 1;
747         } while (css && !iter);
748
749         return iter;
750 }
751 /*
752  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
753  * be careful that "break" loop is not allowed. We have reference count.
754  * Instead of that modify "cond" to be false and "continue" to exit the loop.
755  */
756 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
757         for (iter = mem_cgroup_start_loop(root);\
758              iter != NULL;\
759              iter = mem_cgroup_get_next(iter, root, cond))
760
761 #define for_each_mem_cgroup_tree(iter, root) \
762         for_each_mem_cgroup_tree_cond(iter, root, true)
763
764 #define for_each_mem_cgroup_all(iter) \
765         for_each_mem_cgroup_tree_cond(iter, NULL, true)
766
767
768 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
769 {
770         return (mem == root_mem_cgroup);
771 }
772
773 /*
774  * Following LRU functions are allowed to be used without PCG_LOCK.
775  * Operations are called by routine of global LRU independently from memcg.
776  * What we have to take care of here is validness of pc->mem_cgroup.
777  *
778  * Changes to pc->mem_cgroup happens when
779  * 1. charge
780  * 2. moving account
781  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
782  * It is added to LRU before charge.
783  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
784  * When moving account, the page is not on LRU. It's isolated.
785  */
786
787 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
788 {
789         struct page_cgroup *pc;
790         struct mem_cgroup_per_zone *mz;
791
792         if (mem_cgroup_disabled())
793                 return;
794         pc = lookup_page_cgroup(page);
795         /* can happen while we handle swapcache. */
796         if (!TestClearPageCgroupAcctLRU(pc))
797                 return;
798         VM_BUG_ON(!pc->mem_cgroup);
799         /*
800          * We don't check PCG_USED bit. It's cleared when the "page" is finally
801          * removed from global LRU.
802          */
803         mz = page_cgroup_zoneinfo(pc);
804         /* huge page split is done under lru_lock. so, we have no races. */
805         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
806         if (mem_cgroup_is_root(pc->mem_cgroup))
807                 return;
808         VM_BUG_ON(list_empty(&pc->lru));
809         list_del_init(&pc->lru);
810 }
811
812 void mem_cgroup_del_lru(struct page *page)
813 {
814         mem_cgroup_del_lru_list(page, page_lru(page));
815 }
816
817 /*
818  * Writeback is about to end against a page which has been marked for immediate
819  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
820  * inactive list.
821  */
822 void mem_cgroup_rotate_reclaimable_page(struct page *page)
823 {
824         struct mem_cgroup_per_zone *mz;
825         struct page_cgroup *pc;
826         enum lru_list lru = page_lru(page);
827
828         if (mem_cgroup_disabled())
829                 return;
830
831         pc = lookup_page_cgroup(page);
832         /* unused or root page is not rotated. */
833         if (!PageCgroupUsed(pc))
834                 return;
835         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
836         smp_rmb();
837         if (mem_cgroup_is_root(pc->mem_cgroup))
838                 return;
839         mz = page_cgroup_zoneinfo(pc);
840         list_move_tail(&pc->lru, &mz->lists[lru]);
841 }
842
843 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
844 {
845         struct mem_cgroup_per_zone *mz;
846         struct page_cgroup *pc;
847
848         if (mem_cgroup_disabled())
849                 return;
850
851         pc = lookup_page_cgroup(page);
852         /* unused or root page is not rotated. */
853         if (!PageCgroupUsed(pc))
854                 return;
855         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
856         smp_rmb();
857         if (mem_cgroup_is_root(pc->mem_cgroup))
858                 return;
859         mz = page_cgroup_zoneinfo(pc);
860         list_move(&pc->lru, &mz->lists[lru]);
861 }
862
863 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
864 {
865         struct page_cgroup *pc;
866         struct mem_cgroup_per_zone *mz;
867
868         if (mem_cgroup_disabled())
869                 return;
870         pc = lookup_page_cgroup(page);
871         VM_BUG_ON(PageCgroupAcctLRU(pc));
872         if (!PageCgroupUsed(pc))
873                 return;
874         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
875         smp_rmb();
876         mz = page_cgroup_zoneinfo(pc);
877         /* huge page split is done under lru_lock. so, we have no races. */
878         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
879         SetPageCgroupAcctLRU(pc);
880         if (mem_cgroup_is_root(pc->mem_cgroup))
881                 return;
882         list_add(&pc->lru, &mz->lists[lru]);
883 }
884
885 /*
886  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
887  * lru because the page may.be reused after it's fully uncharged (because of
888  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
889  * it again. This function is only used to charge SwapCache. It's done under
890  * lock_page and expected that zone->lru_lock is never held.
891  */
892 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
893 {
894         unsigned long flags;
895         struct zone *zone = page_zone(page);
896         struct page_cgroup *pc = lookup_page_cgroup(page);
897
898         spin_lock_irqsave(&zone->lru_lock, flags);
899         /*
900          * Forget old LRU when this page_cgroup is *not* used. This Used bit
901          * is guarded by lock_page() because the page is SwapCache.
902          */
903         if (!PageCgroupUsed(pc))
904                 mem_cgroup_del_lru_list(page, page_lru(page));
905         spin_unlock_irqrestore(&zone->lru_lock, flags);
906 }
907
908 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
909 {
910         unsigned long flags;
911         struct zone *zone = page_zone(page);
912         struct page_cgroup *pc = lookup_page_cgroup(page);
913
914         spin_lock_irqsave(&zone->lru_lock, flags);
915         /* link when the page is linked to LRU but page_cgroup isn't */
916         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
917                 mem_cgroup_add_lru_list(page, page_lru(page));
918         spin_unlock_irqrestore(&zone->lru_lock, flags);
919 }
920
921
922 void mem_cgroup_move_lists(struct page *page,
923                            enum lru_list from, enum lru_list to)
924 {
925         if (mem_cgroup_disabled())
926                 return;
927         mem_cgroup_del_lru_list(page, from);
928         mem_cgroup_add_lru_list(page, to);
929 }
930
931 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
932 {
933         int ret;
934         struct mem_cgroup *curr = NULL;
935         struct task_struct *p;
936
937         p = find_lock_task_mm(task);
938         if (!p)
939                 return 0;
940         curr = try_get_mem_cgroup_from_mm(p->mm);
941         task_unlock(p);
942         if (!curr)
943                 return 0;
944         /*
945          * We should check use_hierarchy of "mem" not "curr". Because checking
946          * use_hierarchy of "curr" here make this function true if hierarchy is
947          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
948          * hierarchy(even if use_hierarchy is disabled in "mem").
949          */
950         if (mem->use_hierarchy)
951                 ret = css_is_ancestor(&curr->css, &mem->css);
952         else
953                 ret = (curr == mem);
954         css_put(&curr->css);
955         return ret;
956 }
957
958 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
959 {
960         unsigned long active;
961         unsigned long inactive;
962         unsigned long gb;
963         unsigned long inactive_ratio;
964
965         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
966         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
967
968         gb = (inactive + active) >> (30 - PAGE_SHIFT);
969         if (gb)
970                 inactive_ratio = int_sqrt(10 * gb);
971         else
972                 inactive_ratio = 1;
973
974         if (present_pages) {
975                 present_pages[0] = inactive;
976                 present_pages[1] = active;
977         }
978
979         return inactive_ratio;
980 }
981
982 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
983 {
984         unsigned long active;
985         unsigned long inactive;
986         unsigned long present_pages[2];
987         unsigned long inactive_ratio;
988
989         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
990
991         inactive = present_pages[0];
992         active = present_pages[1];
993
994         if (inactive * inactive_ratio < active)
995                 return 1;
996
997         return 0;
998 }
999
1000 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1001 {
1002         unsigned long active;
1003         unsigned long inactive;
1004
1005         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1006         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1007
1008         return (active > inactive);
1009 }
1010
1011 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
1012                                        struct zone *zone,
1013                                        enum lru_list lru)
1014 {
1015         int nid = zone_to_nid(zone);
1016         int zid = zone_idx(zone);
1017         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1018
1019         return MEM_CGROUP_ZSTAT(mz, lru);
1020 }
1021
1022 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1023                                                       struct zone *zone)
1024 {
1025         int nid = zone_to_nid(zone);
1026         int zid = zone_idx(zone);
1027         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1028
1029         return &mz->reclaim_stat;
1030 }
1031
1032 struct zone_reclaim_stat *
1033 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1034 {
1035         struct page_cgroup *pc;
1036         struct mem_cgroup_per_zone *mz;
1037
1038         if (mem_cgroup_disabled())
1039                 return NULL;
1040
1041         pc = lookup_page_cgroup(page);
1042         if (!PageCgroupUsed(pc))
1043                 return NULL;
1044         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1045         smp_rmb();
1046         mz = page_cgroup_zoneinfo(pc);
1047         if (!mz)
1048                 return NULL;
1049
1050         return &mz->reclaim_stat;
1051 }
1052
1053 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1054                                         struct list_head *dst,
1055                                         unsigned long *scanned, int order,
1056                                         int mode, struct zone *z,
1057                                         struct mem_cgroup *mem_cont,
1058                                         int active, int file)
1059 {
1060         unsigned long nr_taken = 0;
1061         struct page *page;
1062         unsigned long scan;
1063         LIST_HEAD(pc_list);
1064         struct list_head *src;
1065         struct page_cgroup *pc, *tmp;
1066         int nid = zone_to_nid(z);
1067         int zid = zone_idx(z);
1068         struct mem_cgroup_per_zone *mz;
1069         int lru = LRU_FILE * file + active;
1070         int ret;
1071
1072         BUG_ON(!mem_cont);
1073         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1074         src = &mz->lists[lru];
1075
1076         scan = 0;
1077         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1078                 if (scan >= nr_to_scan)
1079                         break;
1080
1081                 page = pc->page;
1082                 if (unlikely(!PageCgroupUsed(pc)))
1083                         continue;
1084                 if (unlikely(!PageLRU(page)))
1085                         continue;
1086
1087                 scan++;
1088                 ret = __isolate_lru_page(page, mode, file);
1089                 switch (ret) {
1090                 case 0:
1091                         list_move(&page->lru, dst);
1092                         mem_cgroup_del_lru(page);
1093                         nr_taken += hpage_nr_pages(page);
1094                         break;
1095                 case -EBUSY:
1096                         /* we don't affect global LRU but rotate in our LRU */
1097                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1098                         break;
1099                 default:
1100                         break;
1101                 }
1102         }
1103
1104         *scanned = scan;
1105
1106         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1107                                       0, 0, 0, mode);
1108
1109         return nr_taken;
1110 }
1111
1112 #define mem_cgroup_from_res_counter(counter, member)    \
1113         container_of(counter, struct mem_cgroup, member)
1114
1115 /**
1116  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1117  * @mem: the memory cgroup
1118  *
1119  * Returns the maximum amount of memory @mem can be charged with, in
1120  * bytes.
1121  */
1122 static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem)
1123 {
1124         unsigned long long margin;
1125
1126         margin = res_counter_margin(&mem->res);
1127         if (do_swap_account)
1128                 margin = min(margin, res_counter_margin(&mem->memsw));
1129         return margin;
1130 }
1131
1132 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1133 {
1134         struct cgroup *cgrp = memcg->css.cgroup;
1135         unsigned int swappiness;
1136
1137         /* root ? */
1138         if (cgrp->parent == NULL)
1139                 return vm_swappiness;
1140
1141         spin_lock(&memcg->reclaim_param_lock);
1142         swappiness = memcg->swappiness;
1143         spin_unlock(&memcg->reclaim_param_lock);
1144
1145         return swappiness;
1146 }
1147
1148 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1149 {
1150         int cpu;
1151
1152         get_online_cpus();
1153         spin_lock(&mem->pcp_counter_lock);
1154         for_each_online_cpu(cpu)
1155                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1156         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1157         spin_unlock(&mem->pcp_counter_lock);
1158         put_online_cpus();
1159
1160         synchronize_rcu();
1161 }
1162
1163 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1164 {
1165         int cpu;
1166
1167         if (!mem)
1168                 return;
1169         get_online_cpus();
1170         spin_lock(&mem->pcp_counter_lock);
1171         for_each_online_cpu(cpu)
1172                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1173         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1174         spin_unlock(&mem->pcp_counter_lock);
1175         put_online_cpus();
1176 }
1177 /*
1178  * 2 routines for checking "mem" is under move_account() or not.
1179  *
1180  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1181  *                        for avoiding race in accounting. If true,
1182  *                        pc->mem_cgroup may be overwritten.
1183  *
1184  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1185  *                        under hierarchy of moving cgroups. This is for
1186  *                        waiting at hith-memory prressure caused by "move".
1187  */
1188
1189 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1190 {
1191         VM_BUG_ON(!rcu_read_lock_held());
1192         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1193 }
1194
1195 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1196 {
1197         struct mem_cgroup *from;
1198         struct mem_cgroup *to;
1199         bool ret = false;
1200         /*
1201          * Unlike task_move routines, we access mc.to, mc.from not under
1202          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1203          */
1204         spin_lock(&mc.lock);
1205         from = mc.from;
1206         to = mc.to;
1207         if (!from)
1208                 goto unlock;
1209         if (from == mem || to == mem
1210             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1211             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1212                 ret = true;
1213 unlock:
1214         spin_unlock(&mc.lock);
1215         return ret;
1216 }
1217
1218 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1219 {
1220         if (mc.moving_task && current != mc.moving_task) {
1221                 if (mem_cgroup_under_move(mem)) {
1222                         DEFINE_WAIT(wait);
1223                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1224                         /* moving charge context might have finished. */
1225                         if (mc.moving_task)
1226                                 schedule();
1227                         finish_wait(&mc.waitq, &wait);
1228                         return true;
1229                 }
1230         }
1231         return false;
1232 }
1233
1234 /**
1235  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1236  * @memcg: The memory cgroup that went over limit
1237  * @p: Task that is going to be killed
1238  *
1239  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1240  * enabled
1241  */
1242 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1243 {
1244         struct cgroup *task_cgrp;
1245         struct cgroup *mem_cgrp;
1246         /*
1247          * Need a buffer in BSS, can't rely on allocations. The code relies
1248          * on the assumption that OOM is serialized for memory controller.
1249          * If this assumption is broken, revisit this code.
1250          */
1251         static char memcg_name[PATH_MAX];
1252         int ret;
1253
1254         if (!memcg || !p)
1255                 return;
1256
1257
1258         rcu_read_lock();
1259
1260         mem_cgrp = memcg->css.cgroup;
1261         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1262
1263         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1264         if (ret < 0) {
1265                 /*
1266                  * Unfortunately, we are unable to convert to a useful name
1267                  * But we'll still print out the usage information
1268                  */
1269                 rcu_read_unlock();
1270                 goto done;
1271         }
1272         rcu_read_unlock();
1273
1274         printk(KERN_INFO "Task in %s killed", memcg_name);
1275
1276         rcu_read_lock();
1277         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1278         if (ret < 0) {
1279                 rcu_read_unlock();
1280                 goto done;
1281         }
1282         rcu_read_unlock();
1283
1284         /*
1285          * Continues from above, so we don't need an KERN_ level
1286          */
1287         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1288 done:
1289
1290         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1291                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1292                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1293                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1294         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1295                 "failcnt %llu\n",
1296                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1297                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1298                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1299 }
1300
1301 /*
1302  * This function returns the number of memcg under hierarchy tree. Returns
1303  * 1(self count) if no children.
1304  */
1305 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1306 {
1307         int num = 0;
1308         struct mem_cgroup *iter;
1309
1310         for_each_mem_cgroup_tree(iter, mem)
1311                 num++;
1312         return num;
1313 }
1314
1315 /*
1316  * Return the memory (and swap, if configured) limit for a memcg.
1317  */
1318 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1319 {
1320         u64 limit;
1321         u64 memsw;
1322
1323         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1324         limit += total_swap_pages << PAGE_SHIFT;
1325
1326         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1327         /*
1328          * If memsw is finite and limits the amount of swap space available
1329          * to this memcg, return that limit.
1330          */
1331         return min(limit, memsw);
1332 }
1333
1334 /*
1335  * Visit the first child (need not be the first child as per the ordering
1336  * of the cgroup list, since we track last_scanned_child) of @mem and use
1337  * that to reclaim free pages from.
1338  */
1339 static struct mem_cgroup *
1340 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1341 {
1342         struct mem_cgroup *ret = NULL;
1343         struct cgroup_subsys_state *css;
1344         int nextid, found;
1345
1346         if (!root_mem->use_hierarchy) {
1347                 css_get(&root_mem->css);
1348                 ret = root_mem;
1349         }
1350
1351         while (!ret) {
1352                 rcu_read_lock();
1353                 nextid = root_mem->last_scanned_child + 1;
1354                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1355                                    &found);
1356                 if (css && css_tryget(css))
1357                         ret = container_of(css, struct mem_cgroup, css);
1358
1359                 rcu_read_unlock();
1360                 /* Updates scanning parameter */
1361                 spin_lock(&root_mem->reclaim_param_lock);
1362                 if (!css) {
1363                         /* this means start scan from ID:1 */
1364                         root_mem->last_scanned_child = 0;
1365                 } else
1366                         root_mem->last_scanned_child = found;
1367                 spin_unlock(&root_mem->reclaim_param_lock);
1368         }
1369
1370         return ret;
1371 }
1372
1373 /*
1374  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1375  * we reclaimed from, so that we don't end up penalizing one child extensively
1376  * based on its position in the children list.
1377  *
1378  * root_mem is the original ancestor that we've been reclaim from.
1379  *
1380  * We give up and return to the caller when we visit root_mem twice.
1381  * (other groups can be removed while we're walking....)
1382  *
1383  * If shrink==true, for avoiding to free too much, this returns immedieately.
1384  */
1385 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1386                                                 struct zone *zone,
1387                                                 gfp_t gfp_mask,
1388                                                 unsigned long reclaim_options)
1389 {
1390         struct mem_cgroup *victim;
1391         int ret, total = 0;
1392         int loop = 0;
1393         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1394         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1395         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1396         unsigned long excess;
1397
1398         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1399
1400         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1401         if (root_mem->memsw_is_minimum)
1402                 noswap = true;
1403
1404         while (1) {
1405                 victim = mem_cgroup_select_victim(root_mem);
1406                 if (victim == root_mem) {
1407                         loop++;
1408                         if (loop >= 1)
1409                                 drain_all_stock_async();
1410                         if (loop >= 2) {
1411                                 /*
1412                                  * If we have not been able to reclaim
1413                                  * anything, it might because there are
1414                                  * no reclaimable pages under this hierarchy
1415                                  */
1416                                 if (!check_soft || !total) {
1417                                         css_put(&victim->css);
1418                                         break;
1419                                 }
1420                                 /*
1421                                  * We want to do more targetted reclaim.
1422                                  * excess >> 2 is not to excessive so as to
1423                                  * reclaim too much, nor too less that we keep
1424                                  * coming back to reclaim from this cgroup
1425                                  */
1426                                 if (total >= (excess >> 2) ||
1427                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1428                                         css_put(&victim->css);
1429                                         break;
1430                                 }
1431                         }
1432                 }
1433                 if (!mem_cgroup_local_usage(victim)) {
1434                         /* this cgroup's local usage == 0 */
1435                         css_put(&victim->css);
1436                         continue;
1437                 }
1438                 /* we use swappiness of local cgroup */
1439                 if (check_soft)
1440                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1441                                 noswap, get_swappiness(victim), zone);
1442                 else
1443                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1444                                                 noswap, get_swappiness(victim));
1445                 css_put(&victim->css);
1446                 /*
1447                  * At shrinking usage, we can't check we should stop here or
1448                  * reclaim more. It's depends on callers. last_scanned_child
1449                  * will work enough for keeping fairness under tree.
1450                  */
1451                 if (shrink)
1452                         return ret;
1453                 total += ret;
1454                 if (check_soft) {
1455                         if (!res_counter_soft_limit_excess(&root_mem->res))
1456                                 return total;
1457                 } else if (mem_cgroup_margin(root_mem))
1458                         return 1 + total;
1459         }
1460         return total;
1461 }
1462
1463 /*
1464  * Check OOM-Killer is already running under our hierarchy.
1465  * If someone is running, return false.
1466  */
1467 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1468 {
1469         int x, lock_count = 0;
1470         struct mem_cgroup *iter;
1471
1472         for_each_mem_cgroup_tree(iter, mem) {
1473                 x = atomic_inc_return(&iter->oom_lock);
1474                 lock_count = max(x, lock_count);
1475         }
1476
1477         if (lock_count == 1)
1478                 return true;
1479         return false;
1480 }
1481
1482 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1483 {
1484         struct mem_cgroup *iter;
1485
1486         /*
1487          * When a new child is created while the hierarchy is under oom,
1488          * mem_cgroup_oom_lock() may not be called. We have to use
1489          * atomic_add_unless() here.
1490          */
1491         for_each_mem_cgroup_tree(iter, mem)
1492                 atomic_add_unless(&iter->oom_lock, -1, 0);
1493         return 0;
1494 }
1495
1496
1497 static DEFINE_MUTEX(memcg_oom_mutex);
1498 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1499
1500 struct oom_wait_info {
1501         struct mem_cgroup *mem;
1502         wait_queue_t    wait;
1503 };
1504
1505 static int memcg_oom_wake_function(wait_queue_t *wait,
1506         unsigned mode, int sync, void *arg)
1507 {
1508         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1509         struct oom_wait_info *oom_wait_info;
1510
1511         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1512
1513         if (oom_wait_info->mem == wake_mem)
1514                 goto wakeup;
1515         /* if no hierarchy, no match */
1516         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1517                 return 0;
1518         /*
1519          * Both of oom_wait_info->mem and wake_mem are stable under us.
1520          * Then we can use css_is_ancestor without taking care of RCU.
1521          */
1522         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1523             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1524                 return 0;
1525
1526 wakeup:
1527         return autoremove_wake_function(wait, mode, sync, arg);
1528 }
1529
1530 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1531 {
1532         /* for filtering, pass "mem" as argument. */
1533         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1534 }
1535
1536 static void memcg_oom_recover(struct mem_cgroup *mem)
1537 {
1538         if (mem && atomic_read(&mem->oom_lock))
1539                 memcg_wakeup_oom(mem);
1540 }
1541
1542 /*
1543  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1544  */
1545 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1546 {
1547         struct oom_wait_info owait;
1548         bool locked, need_to_kill;
1549
1550         owait.mem = mem;
1551         owait.wait.flags = 0;
1552         owait.wait.func = memcg_oom_wake_function;
1553         owait.wait.private = current;
1554         INIT_LIST_HEAD(&owait.wait.task_list);
1555         need_to_kill = true;
1556         /* At first, try to OOM lock hierarchy under mem.*/
1557         mutex_lock(&memcg_oom_mutex);
1558         locked = mem_cgroup_oom_lock(mem);
1559         /*
1560          * Even if signal_pending(), we can't quit charge() loop without
1561          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1562          * under OOM is always welcomed, use TASK_KILLABLE here.
1563          */
1564         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1565         if (!locked || mem->oom_kill_disable)
1566                 need_to_kill = false;
1567         if (locked)
1568                 mem_cgroup_oom_notify(mem);
1569         mutex_unlock(&memcg_oom_mutex);
1570
1571         if (need_to_kill) {
1572                 finish_wait(&memcg_oom_waitq, &owait.wait);
1573                 mem_cgroup_out_of_memory(mem, mask);
1574         } else {
1575                 schedule();
1576                 finish_wait(&memcg_oom_waitq, &owait.wait);
1577         }
1578         mutex_lock(&memcg_oom_mutex);
1579         mem_cgroup_oom_unlock(mem);
1580         memcg_wakeup_oom(mem);
1581         mutex_unlock(&memcg_oom_mutex);
1582
1583         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1584                 return false;
1585         /* Give chance to dying process */
1586         schedule_timeout(1);
1587         return true;
1588 }
1589
1590 /*
1591  * Currently used to update mapped file statistics, but the routine can be
1592  * generalized to update other statistics as well.
1593  *
1594  * Notes: Race condition
1595  *
1596  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1597  * it tends to be costly. But considering some conditions, we doesn't need
1598  * to do so _always_.
1599  *
1600  * Considering "charge", lock_page_cgroup() is not required because all
1601  * file-stat operations happen after a page is attached to radix-tree. There
1602  * are no race with "charge".
1603  *
1604  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1605  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1606  * if there are race with "uncharge". Statistics itself is properly handled
1607  * by flags.
1608  *
1609  * Considering "move", this is an only case we see a race. To make the race
1610  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1611  * possibility of race condition. If there is, we take a lock.
1612  */
1613
1614 void mem_cgroup_update_page_stat(struct page *page,
1615                                  enum mem_cgroup_page_stat_item idx, int val)
1616 {
1617         struct mem_cgroup *mem;
1618         struct page_cgroup *pc = lookup_page_cgroup(page);
1619         bool need_unlock = false;
1620         unsigned long uninitialized_var(flags);
1621
1622         if (unlikely(!pc))
1623                 return;
1624
1625         rcu_read_lock();
1626         mem = pc->mem_cgroup;
1627         if (unlikely(!mem || !PageCgroupUsed(pc)))
1628                 goto out;
1629         /* pc->mem_cgroup is unstable ? */
1630         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1631                 /* take a lock against to access pc->mem_cgroup */
1632                 move_lock_page_cgroup(pc, &flags);
1633                 need_unlock = true;
1634                 mem = pc->mem_cgroup;
1635                 if (!mem || !PageCgroupUsed(pc))
1636                         goto out;
1637         }
1638
1639         switch (idx) {
1640         case MEMCG_NR_FILE_MAPPED:
1641                 if (val > 0)
1642                         SetPageCgroupFileMapped(pc);
1643                 else if (!page_mapped(page))
1644                         ClearPageCgroupFileMapped(pc);
1645                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1646                 break;
1647         default:
1648                 BUG();
1649         }
1650
1651         this_cpu_add(mem->stat->count[idx], val);
1652
1653 out:
1654         if (unlikely(need_unlock))
1655                 move_unlock_page_cgroup(pc, &flags);
1656         rcu_read_unlock();
1657         return;
1658 }
1659 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1660
1661 /*
1662  * size of first charge trial. "32" comes from vmscan.c's magic value.
1663  * TODO: maybe necessary to use big numbers in big irons.
1664  */
1665 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1666 struct memcg_stock_pcp {
1667         struct mem_cgroup *cached; /* this never be root cgroup */
1668         int charge;
1669         struct work_struct work;
1670 };
1671 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1672 static atomic_t memcg_drain_count;
1673
1674 /*
1675  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1676  * from local stock and true is returned. If the stock is 0 or charges from a
1677  * cgroup which is not current target, returns false. This stock will be
1678  * refilled.
1679  */
1680 static bool consume_stock(struct mem_cgroup *mem)
1681 {
1682         struct memcg_stock_pcp *stock;
1683         bool ret = true;
1684
1685         stock = &get_cpu_var(memcg_stock);
1686         if (mem == stock->cached && stock->charge)
1687                 stock->charge -= PAGE_SIZE;
1688         else /* need to call res_counter_charge */
1689                 ret = false;
1690         put_cpu_var(memcg_stock);
1691         return ret;
1692 }
1693
1694 /*
1695  * Returns stocks cached in percpu to res_counter and reset cached information.
1696  */
1697 static void drain_stock(struct memcg_stock_pcp *stock)
1698 {
1699         struct mem_cgroup *old = stock->cached;
1700
1701         if (stock->charge) {
1702                 res_counter_uncharge(&old->res, stock->charge);
1703                 if (do_swap_account)
1704                         res_counter_uncharge(&old->memsw, stock->charge);
1705         }
1706         stock->cached = NULL;
1707         stock->charge = 0;
1708 }
1709
1710 /*
1711  * This must be called under preempt disabled or must be called by
1712  * a thread which is pinned to local cpu.
1713  */
1714 static void drain_local_stock(struct work_struct *dummy)
1715 {
1716         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1717         drain_stock(stock);
1718 }
1719
1720 /*
1721  * Cache charges(val) which is from res_counter, to local per_cpu area.
1722  * This will be consumed by consume_stock() function, later.
1723  */
1724 static void refill_stock(struct mem_cgroup *mem, int val)
1725 {
1726         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1727
1728         if (stock->cached != mem) { /* reset if necessary */
1729                 drain_stock(stock);
1730                 stock->cached = mem;
1731         }
1732         stock->charge += val;
1733         put_cpu_var(memcg_stock);
1734 }
1735
1736 /*
1737  * Tries to drain stocked charges in other cpus. This function is asynchronous
1738  * and just put a work per cpu for draining localy on each cpu. Caller can
1739  * expects some charges will be back to res_counter later but cannot wait for
1740  * it.
1741  */
1742 static void drain_all_stock_async(void)
1743 {
1744         int cpu;
1745         /* This function is for scheduling "drain" in asynchronous way.
1746          * The result of "drain" is not directly handled by callers. Then,
1747          * if someone is calling drain, we don't have to call drain more.
1748          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1749          * there is a race. We just do loose check here.
1750          */
1751         if (atomic_read(&memcg_drain_count))
1752                 return;
1753         /* Notify other cpus that system-wide "drain" is running */
1754         atomic_inc(&memcg_drain_count);
1755         get_online_cpus();
1756         for_each_online_cpu(cpu) {
1757                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1758                 schedule_work_on(cpu, &stock->work);
1759         }
1760         put_online_cpus();
1761         atomic_dec(&memcg_drain_count);
1762         /* We don't wait for flush_work */
1763 }
1764
1765 /* This is a synchronous drain interface. */
1766 static void drain_all_stock_sync(void)
1767 {
1768         /* called when force_empty is called */
1769         atomic_inc(&memcg_drain_count);
1770         schedule_on_each_cpu(drain_local_stock);
1771         atomic_dec(&memcg_drain_count);
1772 }
1773
1774 /*
1775  * This function drains percpu counter value from DEAD cpu and
1776  * move it to local cpu. Note that this function can be preempted.
1777  */
1778 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1779 {
1780         int i;
1781
1782         spin_lock(&mem->pcp_counter_lock);
1783         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1784                 s64 x = per_cpu(mem->stat->count[i], cpu);
1785
1786                 per_cpu(mem->stat->count[i], cpu) = 0;
1787                 mem->nocpu_base.count[i] += x;
1788         }
1789         /* need to clear ON_MOVE value, works as a kind of lock. */
1790         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1791         spin_unlock(&mem->pcp_counter_lock);
1792 }
1793
1794 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1795 {
1796         int idx = MEM_CGROUP_ON_MOVE;
1797
1798         spin_lock(&mem->pcp_counter_lock);
1799         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1800         spin_unlock(&mem->pcp_counter_lock);
1801 }
1802
1803 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1804                                         unsigned long action,
1805                                         void *hcpu)
1806 {
1807         int cpu = (unsigned long)hcpu;
1808         struct memcg_stock_pcp *stock;
1809         struct mem_cgroup *iter;
1810
1811         if ((action == CPU_ONLINE)) {
1812                 for_each_mem_cgroup_all(iter)
1813                         synchronize_mem_cgroup_on_move(iter, cpu);
1814                 return NOTIFY_OK;
1815         }
1816
1817         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1818                 return NOTIFY_OK;
1819
1820         for_each_mem_cgroup_all(iter)
1821                 mem_cgroup_drain_pcp_counter(iter, cpu);
1822
1823         stock = &per_cpu(memcg_stock, cpu);
1824         drain_stock(stock);
1825         return NOTIFY_OK;
1826 }
1827
1828
1829 /* See __mem_cgroup_try_charge() for details */
1830 enum {
1831         CHARGE_OK,              /* success */
1832         CHARGE_RETRY,           /* need to retry but retry is not bad */
1833         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1834         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1835         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1836 };
1837
1838 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1839                                 int csize, bool oom_check)
1840 {
1841         struct mem_cgroup *mem_over_limit;
1842         struct res_counter *fail_res;
1843         unsigned long flags = 0;
1844         int ret;
1845
1846         ret = res_counter_charge(&mem->res, csize, &fail_res);
1847
1848         if (likely(!ret)) {
1849                 if (!do_swap_account)
1850                         return CHARGE_OK;
1851                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1852                 if (likely(!ret))
1853                         return CHARGE_OK;
1854
1855                 res_counter_uncharge(&mem->res, csize);
1856                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1857                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1858         } else
1859                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1860         /*
1861          * csize can be either a huge page (HPAGE_SIZE), a batch of
1862          * regular pages (CHARGE_SIZE), or a single regular page
1863          * (PAGE_SIZE).
1864          *
1865          * Never reclaim on behalf of optional batching, retry with a
1866          * single page instead.
1867          */
1868         if (csize == CHARGE_SIZE)
1869                 return CHARGE_RETRY;
1870
1871         if (!(gfp_mask & __GFP_WAIT))
1872                 return CHARGE_WOULDBLOCK;
1873
1874         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1875                                               gfp_mask, flags);
1876         if (mem_cgroup_margin(mem_over_limit) >= csize)
1877                 return CHARGE_RETRY;
1878         /*
1879          * Even though the limit is exceeded at this point, reclaim
1880          * may have been able to free some pages.  Retry the charge
1881          * before killing the task.
1882          *
1883          * Only for regular pages, though: huge pages are rather
1884          * unlikely to succeed so close to the limit, and we fall back
1885          * to regular pages anyway in case of failure.
1886          */
1887         if (csize == PAGE_SIZE && ret)
1888                 return CHARGE_RETRY;
1889
1890         /*
1891          * At task move, charge accounts can be doubly counted. So, it's
1892          * better to wait until the end of task_move if something is going on.
1893          */
1894         if (mem_cgroup_wait_acct_move(mem_over_limit))
1895                 return CHARGE_RETRY;
1896
1897         /* If we don't need to call oom-killer at el, return immediately */
1898         if (!oom_check)
1899                 return CHARGE_NOMEM;
1900         /* check OOM */
1901         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1902                 return CHARGE_OOM_DIE;
1903
1904         return CHARGE_RETRY;
1905 }
1906
1907 /*
1908  * Unlike exported interface, "oom" parameter is added. if oom==true,
1909  * oom-killer can be invoked.
1910  */
1911 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1912                                    gfp_t gfp_mask,
1913                                    struct mem_cgroup **memcg, bool oom,
1914                                    int page_size)
1915 {
1916         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1917         struct mem_cgroup *mem = NULL;
1918         int ret;
1919         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1920
1921         /*
1922          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1923          * in system level. So, allow to go ahead dying process in addition to
1924          * MEMDIE process.
1925          */
1926         if (unlikely(test_thread_flag(TIF_MEMDIE)
1927                      || fatal_signal_pending(current)))
1928                 goto bypass;
1929
1930         /*
1931          * We always charge the cgroup the mm_struct belongs to.
1932          * The mm_struct's mem_cgroup changes on task migration if the
1933          * thread group leader migrates. It's possible that mm is not
1934          * set, if so charge the init_mm (happens for pagecache usage).
1935          */
1936         if (!*memcg && !mm)
1937                 goto bypass;
1938 again:
1939         if (*memcg) { /* css should be a valid one */
1940                 mem = *memcg;
1941                 VM_BUG_ON(css_is_removed(&mem->css));
1942                 if (mem_cgroup_is_root(mem))
1943                         goto done;
1944                 if (page_size == PAGE_SIZE && consume_stock(mem))
1945                         goto done;
1946                 css_get(&mem->css);
1947         } else {
1948                 struct task_struct *p;
1949
1950                 rcu_read_lock();
1951                 p = rcu_dereference(mm->owner);
1952                 /*
1953                  * Because we don't have task_lock(), "p" can exit.
1954                  * In that case, "mem" can point to root or p can be NULL with
1955                  * race with swapoff. Then, we have small risk of mis-accouning.
1956                  * But such kind of mis-account by race always happens because
1957                  * we don't have cgroup_mutex(). It's overkill and we allo that
1958                  * small race, here.
1959                  * (*) swapoff at el will charge against mm-struct not against
1960                  * task-struct. So, mm->owner can be NULL.
1961                  */
1962                 mem = mem_cgroup_from_task(p);
1963                 if (!mem || mem_cgroup_is_root(mem)) {
1964                         rcu_read_unlock();
1965                         goto done;
1966                 }
1967                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1968                         /*
1969                          * It seems dagerous to access memcg without css_get().
1970                          * But considering how consume_stok works, it's not
1971                          * necessary. If consume_stock success, some charges
1972                          * from this memcg are cached on this cpu. So, we
1973                          * don't need to call css_get()/css_tryget() before
1974                          * calling consume_stock().
1975                          */
1976                         rcu_read_unlock();
1977                         goto done;
1978                 }
1979                 /* after here, we may be blocked. we need to get refcnt */
1980                 if (!css_tryget(&mem->css)) {
1981                         rcu_read_unlock();
1982                         goto again;
1983                 }
1984                 rcu_read_unlock();
1985         }
1986
1987         do {
1988                 bool oom_check;
1989
1990                 /* If killed, bypass charge */
1991                 if (fatal_signal_pending(current)) {
1992                         css_put(&mem->css);
1993                         goto bypass;
1994                 }
1995
1996                 oom_check = false;
1997                 if (oom && !nr_oom_retries) {
1998                         oom_check = true;
1999                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2000                 }
2001
2002                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
2003
2004                 switch (ret) {
2005                 case CHARGE_OK:
2006                         break;
2007                 case CHARGE_RETRY: /* not in OOM situation but retry */
2008                         csize = page_size;
2009                         css_put(&mem->css);
2010                         mem = NULL;
2011                         goto again;
2012                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2013                         css_put(&mem->css);
2014                         goto nomem;
2015                 case CHARGE_NOMEM: /* OOM routine works */
2016                         if (!oom) {
2017                                 css_put(&mem->css);
2018                                 goto nomem;
2019                         }
2020                         /* If oom, we never return -ENOMEM */
2021                         nr_oom_retries--;
2022                         break;
2023                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2024                         css_put(&mem->css);
2025                         goto bypass;
2026                 }
2027         } while (ret != CHARGE_OK);
2028
2029         if (csize > page_size)
2030                 refill_stock(mem, csize - page_size);
2031         css_put(&mem->css);
2032 done:
2033         *memcg = mem;
2034         return 0;
2035 nomem:
2036         *memcg = NULL;
2037         return -ENOMEM;
2038 bypass:
2039         *memcg = NULL;
2040         return 0;
2041 }
2042
2043 /*
2044  * Somemtimes we have to undo a charge we got by try_charge().
2045  * This function is for that and do uncharge, put css's refcnt.
2046  * gotten by try_charge().
2047  */
2048 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2049                                                         unsigned long count)
2050 {
2051         if (!mem_cgroup_is_root(mem)) {
2052                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2053                 if (do_swap_account)
2054                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2055         }
2056 }
2057
2058 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2059                                      int page_size)
2060 {
2061         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2062 }
2063
2064 /*
2065  * A helper function to get mem_cgroup from ID. must be called under
2066  * rcu_read_lock(). The caller must check css_is_removed() or some if
2067  * it's concern. (dropping refcnt from swap can be called against removed
2068  * memcg.)
2069  */
2070 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2071 {
2072         struct cgroup_subsys_state *css;
2073
2074         /* ID 0 is unused ID */
2075         if (!id)
2076                 return NULL;
2077         css = css_lookup(&mem_cgroup_subsys, id);
2078         if (!css)
2079                 return NULL;
2080         return container_of(css, struct mem_cgroup, css);
2081 }
2082
2083 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2084 {
2085         struct mem_cgroup *mem = NULL;
2086         struct page_cgroup *pc;
2087         unsigned short id;
2088         swp_entry_t ent;
2089
2090         VM_BUG_ON(!PageLocked(page));
2091
2092         pc = lookup_page_cgroup(page);
2093         lock_page_cgroup(pc);
2094         if (PageCgroupUsed(pc)) {
2095                 mem = pc->mem_cgroup;
2096                 if (mem && !css_tryget(&mem->css))
2097                         mem = NULL;
2098         } else if (PageSwapCache(page)) {
2099                 ent.val = page_private(page);
2100                 id = lookup_swap_cgroup(ent);
2101                 rcu_read_lock();
2102                 mem = mem_cgroup_lookup(id);
2103                 if (mem && !css_tryget(&mem->css))
2104                         mem = NULL;
2105                 rcu_read_unlock();
2106         }
2107         unlock_page_cgroup(pc);
2108         return mem;
2109 }
2110
2111 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2112                                        struct page_cgroup *pc,
2113                                        enum charge_type ctype,
2114                                        int page_size)
2115 {
2116         int nr_pages = page_size >> PAGE_SHIFT;
2117
2118         lock_page_cgroup(pc);
2119         if (unlikely(PageCgroupUsed(pc))) {
2120                 unlock_page_cgroup(pc);
2121                 mem_cgroup_cancel_charge(mem, page_size);
2122                 return;
2123         }
2124         /*
2125          * we don't need page_cgroup_lock about tail pages, becase they are not
2126          * accessed by any other context at this point.
2127          */
2128         pc->mem_cgroup = mem;
2129         /*
2130          * We access a page_cgroup asynchronously without lock_page_cgroup().
2131          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2132          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2133          * before USED bit, we need memory barrier here.
2134          * See mem_cgroup_add_lru_list(), etc.
2135          */
2136         smp_wmb();
2137         switch (ctype) {
2138         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2139         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2140                 SetPageCgroupCache(pc);
2141                 SetPageCgroupUsed(pc);
2142                 break;
2143         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2144                 ClearPageCgroupCache(pc);
2145                 SetPageCgroupUsed(pc);
2146                 break;
2147         default:
2148                 break;
2149         }
2150
2151         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2152         unlock_page_cgroup(pc);
2153         /*
2154          * "charge_statistics" updated event counter. Then, check it.
2155          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2156          * if they exceeds softlimit.
2157          */
2158         memcg_check_events(mem, pc->page);
2159 }
2160
2161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2162
2163 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2164                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2165 /*
2166  * Because tail pages are not marked as "used", set it. We're under
2167  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2168  */
2169 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2170 {
2171         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2172         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2173         unsigned long flags;
2174
2175         if (mem_cgroup_disabled())
2176                 return;
2177         /*
2178          * We have no races with charge/uncharge but will have races with
2179          * page state accounting.
2180          */
2181         move_lock_page_cgroup(head_pc, &flags);
2182
2183         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2184         smp_wmb(); /* see __commit_charge() */
2185         if (PageCgroupAcctLRU(head_pc)) {
2186                 enum lru_list lru;
2187                 struct mem_cgroup_per_zone *mz;
2188
2189                 /*
2190                  * LRU flags cannot be copied because we need to add tail
2191                  *.page to LRU by generic call and our hook will be called.
2192                  * We hold lru_lock, then, reduce counter directly.
2193                  */
2194                 lru = page_lru(head);
2195                 mz = page_cgroup_zoneinfo(head_pc);
2196                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2197         }
2198         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2199         move_unlock_page_cgroup(head_pc, &flags);
2200 }
2201 #endif
2202
2203 /**
2204  * __mem_cgroup_move_account - move account of the page
2205  * @pc: page_cgroup of the page.
2206  * @from: mem_cgroup which the page is moved from.
2207  * @to: mem_cgroup which the page is moved to. @from != @to.
2208  * @uncharge: whether we should call uncharge and css_put against @from.
2209  *
2210  * The caller must confirm following.
2211  * - page is not on LRU (isolate_page() is useful.)
2212  * - the pc is locked, used, and ->mem_cgroup points to @from.
2213  *
2214  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2215  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2216  * true, this function does "uncharge" from old cgroup, but it doesn't if
2217  * @uncharge is false, so a caller should do "uncharge".
2218  */
2219
2220 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2221         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge,
2222         int charge_size)
2223 {
2224         int nr_pages = charge_size >> PAGE_SHIFT;
2225
2226         VM_BUG_ON(from == to);
2227         VM_BUG_ON(PageLRU(pc->page));
2228         VM_BUG_ON(!page_is_cgroup_locked(pc));
2229         VM_BUG_ON(!PageCgroupUsed(pc));
2230         VM_BUG_ON(pc->mem_cgroup != from);
2231
2232         if (PageCgroupFileMapped(pc)) {
2233                 /* Update mapped_file data for mem_cgroup */
2234                 preempt_disable();
2235                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2236                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2237                 preempt_enable();
2238         }
2239         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2240         if (uncharge)
2241                 /* This is not "cancel", but cancel_charge does all we need. */
2242                 mem_cgroup_cancel_charge(from, charge_size);
2243
2244         /* caller should have done css_get */
2245         pc->mem_cgroup = to;
2246         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2247         /*
2248          * We charges against "to" which may not have any tasks. Then, "to"
2249          * can be under rmdir(). But in current implementation, caller of
2250          * this function is just force_empty() and move charge, so it's
2251          * garanteed that "to" is never removed. So, we don't check rmdir
2252          * status here.
2253          */
2254 }
2255
2256 /*
2257  * check whether the @pc is valid for moving account and call
2258  * __mem_cgroup_move_account()
2259  */
2260 static int mem_cgroup_move_account(struct page_cgroup *pc,
2261                 struct mem_cgroup *from, struct mem_cgroup *to,
2262                 bool uncharge, int charge_size)
2263 {
2264         int ret = -EINVAL;
2265         unsigned long flags;
2266         /*
2267          * The page is isolated from LRU. So, collapse function
2268          * will not handle this page. But page splitting can happen.
2269          * Do this check under compound_page_lock(). The caller should
2270          * hold it.
2271          */
2272         if ((charge_size > PAGE_SIZE) && !PageTransHuge(pc->page))
2273                 return -EBUSY;
2274
2275         lock_page_cgroup(pc);
2276         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2277                 move_lock_page_cgroup(pc, &flags);
2278                 __mem_cgroup_move_account(pc, from, to, uncharge, charge_size);
2279                 move_unlock_page_cgroup(pc, &flags);
2280                 ret = 0;
2281         }
2282         unlock_page_cgroup(pc);
2283         /*
2284          * check events
2285          */
2286         memcg_check_events(to, pc->page);
2287         memcg_check_events(from, pc->page);
2288         return ret;
2289 }
2290
2291 /*
2292  * move charges to its parent.
2293  */
2294
2295 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2296                                   struct mem_cgroup *child,
2297                                   gfp_t gfp_mask)
2298 {
2299         struct page *page = pc->page;
2300         struct cgroup *cg = child->css.cgroup;
2301         struct cgroup *pcg = cg->parent;
2302         struct mem_cgroup *parent;
2303         int page_size = PAGE_SIZE;
2304         unsigned long flags;
2305         int ret;
2306
2307         /* Is ROOT ? */
2308         if (!pcg)
2309                 return -EINVAL;
2310
2311         ret = -EBUSY;
2312         if (!get_page_unless_zero(page))
2313                 goto out;
2314         if (isolate_lru_page(page))
2315                 goto put;
2316
2317         if (PageTransHuge(page))
2318                 page_size = HPAGE_SIZE;
2319
2320         parent = mem_cgroup_from_cont(pcg);
2321         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2322                                 &parent, false, page_size);
2323         if (ret || !parent)
2324                 goto put_back;
2325
2326         if (page_size > PAGE_SIZE)
2327                 flags = compound_lock_irqsave(page);
2328
2329         ret = mem_cgroup_move_account(pc, child, parent, true, page_size);
2330         if (ret)
2331                 mem_cgroup_cancel_charge(parent, page_size);
2332
2333         if (page_size > PAGE_SIZE)
2334                 compound_unlock_irqrestore(page, flags);
2335 put_back:
2336         putback_lru_page(page);
2337 put:
2338         put_page(page);
2339 out:
2340         return ret;
2341 }
2342
2343 /*
2344  * Charge the memory controller for page usage.
2345  * Return
2346  * 0 if the charge was successful
2347  * < 0 if the cgroup is over its limit
2348  */
2349 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2350                                 gfp_t gfp_mask, enum charge_type ctype)
2351 {
2352         struct mem_cgroup *mem = NULL;
2353         int page_size = PAGE_SIZE;
2354         struct page_cgroup *pc;
2355         bool oom = true;
2356         int ret;
2357
2358         if (PageTransHuge(page)) {
2359                 page_size <<= compound_order(page);
2360                 VM_BUG_ON(!PageTransHuge(page));
2361                 /*
2362                  * Never OOM-kill a process for a huge page.  The
2363                  * fault handler will fall back to regular pages.
2364                  */
2365                 oom = false;
2366         }
2367
2368         pc = lookup_page_cgroup(page);
2369         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2370
2371         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2372         if (ret || !mem)
2373                 return ret;
2374
2375         __mem_cgroup_commit_charge(mem, pc, ctype, page_size);
2376         return 0;
2377 }
2378
2379 int mem_cgroup_newpage_charge(struct page *page,
2380                               struct mm_struct *mm, gfp_t gfp_mask)
2381 {
2382         if (mem_cgroup_disabled())
2383                 return 0;
2384         /*
2385          * If already mapped, we don't have to account.
2386          * If page cache, page->mapping has address_space.
2387          * But page->mapping may have out-of-use anon_vma pointer,
2388          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2389          * is NULL.
2390          */
2391         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2392                 return 0;
2393         if (unlikely(!mm))
2394                 mm = &init_mm;
2395         return mem_cgroup_charge_common(page, mm, gfp_mask,
2396                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2397 }
2398
2399 static void
2400 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2401                                         enum charge_type ctype);
2402
2403 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2404                                 gfp_t gfp_mask)
2405 {
2406         int ret;
2407
2408         if (mem_cgroup_disabled())
2409                 return 0;
2410         if (PageCompound(page))
2411                 return 0;
2412         /*
2413          * Corner case handling. This is called from add_to_page_cache()
2414          * in usual. But some FS (shmem) precharges this page before calling it
2415          * and call add_to_page_cache() with GFP_NOWAIT.
2416          *
2417          * For GFP_NOWAIT case, the page may be pre-charged before calling
2418          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2419          * charge twice. (It works but has to pay a bit larger cost.)
2420          * And when the page is SwapCache, it should take swap information
2421          * into account. This is under lock_page() now.
2422          */
2423         if (!(gfp_mask & __GFP_WAIT)) {
2424                 struct page_cgroup *pc;
2425
2426                 pc = lookup_page_cgroup(page);
2427                 if (!pc)
2428                         return 0;
2429                 lock_page_cgroup(pc);
2430                 if (PageCgroupUsed(pc)) {
2431                         unlock_page_cgroup(pc);
2432                         return 0;
2433                 }
2434                 unlock_page_cgroup(pc);
2435         }
2436
2437         if (unlikely(!mm))
2438                 mm = &init_mm;
2439
2440         if (page_is_file_cache(page))
2441                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2442                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2443
2444         /* shmem */
2445         if (PageSwapCache(page)) {
2446                 struct mem_cgroup *mem;
2447
2448                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2449                 if (!ret)
2450                         __mem_cgroup_commit_charge_swapin(page, mem,
2451                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2452         } else
2453                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2454                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2455
2456         return ret;
2457 }
2458
2459 /*
2460  * While swap-in, try_charge -> commit or cancel, the page is locked.
2461  * And when try_charge() successfully returns, one refcnt to memcg without
2462  * struct page_cgroup is acquired. This refcnt will be consumed by
2463  * "commit()" or removed by "cancel()"
2464  */
2465 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2466                                  struct page *page,
2467                                  gfp_t mask, struct mem_cgroup **ptr)
2468 {
2469         struct mem_cgroup *mem;
2470         int ret;
2471
2472         *ptr = NULL;
2473
2474         if (mem_cgroup_disabled())
2475                 return 0;
2476
2477         if (!do_swap_account)
2478                 goto charge_cur_mm;
2479         /*
2480          * A racing thread's fault, or swapoff, may have already updated
2481          * the pte, and even removed page from swap cache: in those cases
2482          * do_swap_page()'s pte_same() test will fail; but there's also a
2483          * KSM case which does need to charge the page.
2484          */
2485         if (!PageSwapCache(page))
2486                 goto charge_cur_mm;
2487         mem = try_get_mem_cgroup_from_page(page);
2488         if (!mem)
2489                 goto charge_cur_mm;
2490         *ptr = mem;
2491         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2492         css_put(&mem->css);
2493         return ret;
2494 charge_cur_mm:
2495         if (unlikely(!mm))
2496                 mm = &init_mm;
2497         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2498 }
2499
2500 static void
2501 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2502                                         enum charge_type ctype)
2503 {
2504         struct page_cgroup *pc;
2505
2506         if (mem_cgroup_disabled())
2507                 return;
2508         if (!ptr)
2509                 return;
2510         cgroup_exclude_rmdir(&ptr->css);
2511         pc = lookup_page_cgroup(page);
2512         mem_cgroup_lru_del_before_commit_swapcache(page);
2513         __mem_cgroup_commit_charge(ptr, pc, ctype, PAGE_SIZE);
2514         mem_cgroup_lru_add_after_commit_swapcache(page);
2515         /*
2516          * Now swap is on-memory. This means this page may be
2517          * counted both as mem and swap....double count.
2518          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2519          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2520          * may call delete_from_swap_cache() before reach here.
2521          */
2522         if (do_swap_account && PageSwapCache(page)) {
2523                 swp_entry_t ent = {.val = page_private(page)};
2524                 unsigned short id;
2525                 struct mem_cgroup *memcg;
2526
2527                 id = swap_cgroup_record(ent, 0);
2528                 rcu_read_lock();
2529                 memcg = mem_cgroup_lookup(id);
2530                 if (memcg) {
2531                         /*
2532                          * This recorded memcg can be obsolete one. So, avoid
2533                          * calling css_tryget
2534                          */
2535                         if (!mem_cgroup_is_root(memcg))
2536                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2537                         mem_cgroup_swap_statistics(memcg, false);
2538                         mem_cgroup_put(memcg);
2539                 }
2540                 rcu_read_unlock();
2541         }
2542         /*
2543          * At swapin, we may charge account against cgroup which has no tasks.
2544          * So, rmdir()->pre_destroy() can be called while we do this charge.
2545          * In that case, we need to call pre_destroy() again. check it here.
2546          */
2547         cgroup_release_and_wakeup_rmdir(&ptr->css);
2548 }
2549
2550 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2551 {
2552         __mem_cgroup_commit_charge_swapin(page, ptr,
2553                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2554 }
2555
2556 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2557 {
2558         if (mem_cgroup_disabled())
2559                 return;
2560         if (!mem)
2561                 return;
2562         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2563 }
2564
2565 static void
2566 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2567               int page_size)
2568 {
2569         struct memcg_batch_info *batch = NULL;
2570         bool uncharge_memsw = true;
2571         /* If swapout, usage of swap doesn't decrease */
2572         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2573                 uncharge_memsw = false;
2574
2575         batch = &current->memcg_batch;
2576         /*
2577          * In usual, we do css_get() when we remember memcg pointer.
2578          * But in this case, we keep res->usage until end of a series of
2579          * uncharges. Then, it's ok to ignore memcg's refcnt.
2580          */
2581         if (!batch->memcg)
2582                 batch->memcg = mem;
2583         /*
2584          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2585          * In those cases, all pages freed continously can be expected to be in
2586          * the same cgroup and we have chance to coalesce uncharges.
2587          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2588          * because we want to do uncharge as soon as possible.
2589          */
2590
2591         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2592                 goto direct_uncharge;
2593
2594         if (page_size != PAGE_SIZE)
2595                 goto direct_uncharge;
2596
2597         /*
2598          * In typical case, batch->memcg == mem. This means we can
2599          * merge a series of uncharges to an uncharge of res_counter.
2600          * If not, we uncharge res_counter ony by one.
2601          */
2602         if (batch->memcg != mem)
2603                 goto direct_uncharge;
2604         /* remember freed charge and uncharge it later */
2605         batch->bytes += PAGE_SIZE;
2606         if (uncharge_memsw)
2607                 batch->memsw_bytes += PAGE_SIZE;
2608         return;
2609 direct_uncharge:
2610         res_counter_uncharge(&mem->res, page_size);
2611         if (uncharge_memsw)
2612                 res_counter_uncharge(&mem->memsw, page_size);
2613         if (unlikely(batch->memcg != mem))
2614                 memcg_oom_recover(mem);
2615         return;
2616 }
2617
2618 /*
2619  * uncharge if !page_mapped(page)
2620  */
2621 static struct mem_cgroup *
2622 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2623 {
2624         int count;
2625         struct page_cgroup *pc;
2626         struct mem_cgroup *mem = NULL;
2627         int page_size = PAGE_SIZE;
2628
2629         if (mem_cgroup_disabled())
2630                 return NULL;
2631
2632         if (PageSwapCache(page))
2633                 return NULL;
2634
2635         if (PageTransHuge(page)) {
2636                 page_size <<= compound_order(page);
2637                 VM_BUG_ON(!PageTransHuge(page));
2638         }
2639
2640         count = page_size >> PAGE_SHIFT;
2641         /*
2642          * Check if our page_cgroup is valid
2643          */
2644         pc = lookup_page_cgroup(page);
2645         if (unlikely(!pc || !PageCgroupUsed(pc)))
2646                 return NULL;
2647
2648         lock_page_cgroup(pc);
2649
2650         mem = pc->mem_cgroup;
2651
2652         if (!PageCgroupUsed(pc))
2653                 goto unlock_out;
2654
2655         switch (ctype) {
2656         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2657         case MEM_CGROUP_CHARGE_TYPE_DROP:
2658                 /* See mem_cgroup_prepare_migration() */
2659                 if (page_mapped(page) || PageCgroupMigration(pc))
2660                         goto unlock_out;
2661                 break;
2662         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2663                 if (!PageAnon(page)) {  /* Shared memory */
2664                         if (page->mapping && !page_is_file_cache(page))
2665                                 goto unlock_out;
2666                 } else if (page_mapped(page)) /* Anon */
2667                                 goto unlock_out;
2668                 break;
2669         default:
2670                 break;
2671         }
2672
2673         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2674
2675         ClearPageCgroupUsed(pc);
2676         /*
2677          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2678          * freed from LRU. This is safe because uncharged page is expected not
2679          * to be reused (freed soon). Exception is SwapCache, it's handled by
2680          * special functions.
2681          */
2682
2683         unlock_page_cgroup(pc);
2684         /*
2685          * even after unlock, we have mem->res.usage here and this memcg
2686          * will never be freed.
2687          */
2688         memcg_check_events(mem, page);
2689         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2690                 mem_cgroup_swap_statistics(mem, true);
2691                 mem_cgroup_get(mem);
2692         }
2693         if (!mem_cgroup_is_root(mem))
2694                 __do_uncharge(mem, ctype, page_size);
2695
2696         return mem;
2697
2698 unlock_out:
2699         unlock_page_cgroup(pc);
2700         return NULL;
2701 }
2702
2703 void mem_cgroup_uncharge_page(struct page *page)
2704 {
2705         /* early check. */
2706         if (page_mapped(page))
2707                 return;
2708         if (page->mapping && !PageAnon(page))
2709                 return;
2710         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2711 }
2712
2713 void mem_cgroup_uncharge_cache_page(struct page *page)
2714 {
2715         VM_BUG_ON(page_mapped(page));
2716         VM_BUG_ON(page->mapping);
2717         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2718 }
2719
2720 /*
2721  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2722  * In that cases, pages are freed continuously and we can expect pages
2723  * are in the same memcg. All these calls itself limits the number of
2724  * pages freed at once, then uncharge_start/end() is called properly.
2725  * This may be called prural(2) times in a context,
2726  */
2727
2728 void mem_cgroup_uncharge_start(void)
2729 {
2730         current->memcg_batch.do_batch++;
2731         /* We can do nest. */
2732         if (current->memcg_batch.do_batch == 1) {
2733                 current->memcg_batch.memcg = NULL;
2734                 current->memcg_batch.bytes = 0;
2735                 current->memcg_batch.memsw_bytes = 0;
2736         }
2737 }
2738
2739 void mem_cgroup_uncharge_end(void)
2740 {
2741         struct memcg_batch_info *batch = &current->memcg_batch;
2742
2743         if (!batch->do_batch)
2744                 return;
2745
2746         batch->do_batch--;
2747         if (batch->do_batch) /* If stacked, do nothing. */
2748                 return;
2749
2750         if (!batch->memcg)
2751                 return;
2752         /*
2753          * This "batch->memcg" is valid without any css_get/put etc...
2754          * bacause we hide charges behind us.
2755          */
2756         if (batch->bytes)
2757                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2758         if (batch->memsw_bytes)
2759                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2760         memcg_oom_recover(batch->memcg);
2761         /* forget this pointer (for sanity check) */
2762         batch->memcg = NULL;
2763 }
2764
2765 #ifdef CONFIG_SWAP
2766 /*
2767  * called after __delete_from_swap_cache() and drop "page" account.
2768  * memcg information is recorded to swap_cgroup of "ent"
2769  */
2770 void
2771 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2772 {
2773         struct mem_cgroup *memcg;
2774         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2775
2776         if (!swapout) /* this was a swap cache but the swap is unused ! */
2777                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2778
2779         memcg = __mem_cgroup_uncharge_common(page, ctype);
2780
2781         /*
2782          * record memcg information,  if swapout && memcg != NULL,
2783          * mem_cgroup_get() was called in uncharge().
2784          */
2785         if (do_swap_account && swapout && memcg)
2786                 swap_cgroup_record(ent, css_id(&memcg->css));
2787 }
2788 #endif
2789
2790 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2791 /*
2792  * called from swap_entry_free(). remove record in swap_cgroup and
2793  * uncharge "memsw" account.
2794  */
2795 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2796 {
2797         struct mem_cgroup *memcg;
2798         unsigned short id;
2799
2800         if (!do_swap_account)
2801                 return;
2802
2803         id = swap_cgroup_record(ent, 0);
2804         rcu_read_lock();
2805         memcg = mem_cgroup_lookup(id);
2806         if (memcg) {
2807                 /*
2808                  * We uncharge this because swap is freed.
2809                  * This memcg can be obsolete one. We avoid calling css_tryget
2810                  */
2811                 if (!mem_cgroup_is_root(memcg))
2812                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2813                 mem_cgroup_swap_statistics(memcg, false);
2814                 mem_cgroup_put(memcg);
2815         }
2816         rcu_read_unlock();
2817 }
2818
2819 /**
2820  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2821  * @entry: swap entry to be moved
2822  * @from:  mem_cgroup which the entry is moved from
2823  * @to:  mem_cgroup which the entry is moved to
2824  * @need_fixup: whether we should fixup res_counters and refcounts.
2825  *
2826  * It succeeds only when the swap_cgroup's record for this entry is the same
2827  * as the mem_cgroup's id of @from.
2828  *
2829  * Returns 0 on success, -EINVAL on failure.
2830  *
2831  * The caller must have charged to @to, IOW, called res_counter_charge() about
2832  * both res and memsw, and called css_get().
2833  */
2834 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2835                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2836 {
2837         unsigned short old_id, new_id;
2838
2839         old_id = css_id(&from->css);
2840         new_id = css_id(&to->css);
2841
2842         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2843                 mem_cgroup_swap_statistics(from, false);
2844                 mem_cgroup_swap_statistics(to, true);
2845                 /*
2846                  * This function is only called from task migration context now.
2847                  * It postpones res_counter and refcount handling till the end
2848                  * of task migration(mem_cgroup_clear_mc()) for performance
2849                  * improvement. But we cannot postpone mem_cgroup_get(to)
2850                  * because if the process that has been moved to @to does
2851                  * swap-in, the refcount of @to might be decreased to 0.
2852                  */
2853                 mem_cgroup_get(to);
2854                 if (need_fixup) {
2855                         if (!mem_cgroup_is_root(from))
2856                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2857                         mem_cgroup_put(from);
2858                         /*
2859                          * we charged both to->res and to->memsw, so we should
2860                          * uncharge to->res.
2861                          */
2862                         if (!mem_cgroup_is_root(to))
2863                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2864                 }
2865                 return 0;
2866         }
2867         return -EINVAL;
2868 }
2869 #else
2870 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2871                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2872 {
2873         return -EINVAL;
2874 }
2875 #endif
2876
2877 /*
2878  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2879  * page belongs to.
2880  */
2881 int mem_cgroup_prepare_migration(struct page *page,
2882         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2883 {
2884         struct page_cgroup *pc;
2885         struct mem_cgroup *mem = NULL;
2886         enum charge_type ctype;
2887         int ret = 0;
2888
2889         *ptr = NULL;
2890
2891         VM_BUG_ON(PageTransHuge(page));
2892         if (mem_cgroup_disabled())
2893                 return 0;
2894
2895         pc = lookup_page_cgroup(page);
2896         lock_page_cgroup(pc);
2897         if (PageCgroupUsed(pc)) {
2898                 mem = pc->mem_cgroup;
2899                 css_get(&mem->css);
2900                 /*
2901                  * At migrating an anonymous page, its mapcount goes down
2902                  * to 0 and uncharge() will be called. But, even if it's fully
2903                  * unmapped, migration may fail and this page has to be
2904                  * charged again. We set MIGRATION flag here and delay uncharge
2905                  * until end_migration() is called
2906                  *
2907                  * Corner Case Thinking
2908                  * A)
2909                  * When the old page was mapped as Anon and it's unmap-and-freed
2910                  * while migration was ongoing.
2911                  * If unmap finds the old page, uncharge() of it will be delayed
2912                  * until end_migration(). If unmap finds a new page, it's
2913                  * uncharged when it make mapcount to be 1->0. If unmap code
2914                  * finds swap_migration_entry, the new page will not be mapped
2915                  * and end_migration() will find it(mapcount==0).
2916                  *
2917                  * B)
2918                  * When the old page was mapped but migraion fails, the kernel
2919                  * remaps it. A charge for it is kept by MIGRATION flag even
2920                  * if mapcount goes down to 0. We can do remap successfully
2921                  * without charging it again.
2922                  *
2923                  * C)
2924                  * The "old" page is under lock_page() until the end of
2925                  * migration, so, the old page itself will not be swapped-out.
2926                  * If the new page is swapped out before end_migraton, our
2927                  * hook to usual swap-out path will catch the event.
2928                  */
2929                 if (PageAnon(page))
2930                         SetPageCgroupMigration(pc);
2931         }
2932         unlock_page_cgroup(pc);
2933         /*
2934          * If the page is not charged at this point,
2935          * we return here.
2936          */
2937         if (!mem)
2938                 return 0;
2939
2940         *ptr = mem;
2941         ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2942         css_put(&mem->css);/* drop extra refcnt */
2943         if (ret || *ptr == NULL) {
2944                 if (PageAnon(page)) {
2945                         lock_page_cgroup(pc);
2946                         ClearPageCgroupMigration(pc);
2947                         unlock_page_cgroup(pc);
2948                         /*
2949                          * The old page may be fully unmapped while we kept it.
2950                          */
2951                         mem_cgroup_uncharge_page(page);
2952                 }
2953                 return -ENOMEM;
2954         }
2955         /*
2956          * We charge new page before it's used/mapped. So, even if unlock_page()
2957          * is called before end_migration, we can catch all events on this new
2958          * page. In the case new page is migrated but not remapped, new page's
2959          * mapcount will be finally 0 and we call uncharge in end_migration().
2960          */
2961         pc = lookup_page_cgroup(newpage);
2962         if (PageAnon(page))
2963                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2964         else if (page_is_file_cache(page))
2965                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2966         else
2967                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2968         __mem_cgroup_commit_charge(mem, pc, ctype, PAGE_SIZE);
2969         return ret;
2970 }
2971
2972 /* remove redundant charge if migration failed*/
2973 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2974         struct page *oldpage, struct page *newpage, bool migration_ok)
2975 {
2976         struct page *used, *unused;
2977         struct page_cgroup *pc;
2978
2979         if (!mem)
2980                 return;
2981         /* blocks rmdir() */
2982         cgroup_exclude_rmdir(&mem->css);
2983         if (!migration_ok) {
2984                 used = oldpage;
2985                 unused = newpage;
2986         } else {
2987                 used = newpage;
2988                 unused = oldpage;
2989         }
2990         /*
2991          * We disallowed uncharge of pages under migration because mapcount
2992          * of the page goes down to zero, temporarly.
2993          * Clear the flag and check the page should be charged.
2994          */
2995         pc = lookup_page_cgroup(oldpage);
2996         lock_page_cgroup(pc);
2997         ClearPageCgroupMigration(pc);
2998         unlock_page_cgroup(pc);
2999
3000         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3001
3002         /*
3003          * If a page is a file cache, radix-tree replacement is very atomic
3004          * and we can skip this check. When it was an Anon page, its mapcount
3005          * goes down to 0. But because we added MIGRATION flage, it's not
3006          * uncharged yet. There are several case but page->mapcount check
3007          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3008          * check. (see prepare_charge() also)
3009          */
3010         if (PageAnon(used))
3011                 mem_cgroup_uncharge_page(used);
3012         /*
3013          * At migration, we may charge account against cgroup which has no
3014          * tasks.
3015          * So, rmdir()->pre_destroy() can be called while we do this charge.
3016          * In that case, we need to call pre_destroy() again. check it here.
3017          */
3018         cgroup_release_and_wakeup_rmdir(&mem->css);
3019 }
3020
3021 /*
3022  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3023  * Calling hierarchical_reclaim is not enough because we should update
3024  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3025  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3026  * not from the memcg which this page would be charged to.
3027  * try_charge_swapin does all of these works properly.
3028  */
3029 int mem_cgroup_shmem_charge_fallback(struct page *page,
3030                             struct mm_struct *mm,
3031                             gfp_t gfp_mask)
3032 {
3033         struct mem_cgroup *mem;
3034         int ret;
3035
3036         if (mem_cgroup_disabled())
3037                 return 0;
3038
3039         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3040         if (!ret)
3041                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3042
3043         return ret;
3044 }
3045
3046 #ifdef CONFIG_DEBUG_VM
3047 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3048 {
3049         struct page_cgroup *pc;
3050
3051         pc = lookup_page_cgroup(page);
3052         if (likely(pc) && PageCgroupUsed(pc))
3053                 return pc;
3054         return NULL;
3055 }
3056
3057 bool mem_cgroup_bad_page_check(struct page *page)
3058 {
3059         if (mem_cgroup_disabled())
3060                 return false;
3061
3062         return lookup_page_cgroup_used(page) != NULL;
3063 }
3064
3065 void mem_cgroup_print_bad_page(struct page *page)
3066 {
3067         struct page_cgroup *pc;
3068
3069         pc = lookup_page_cgroup_used(page);
3070         if (pc) {
3071                 int ret = -1;
3072                 char *path;
3073
3074                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3075                        pc, pc->flags, pc->mem_cgroup);
3076
3077                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3078                 if (path) {
3079                         rcu_read_lock();
3080                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3081                                                         path, PATH_MAX);
3082                         rcu_read_unlock();
3083                 }
3084
3085                 printk(KERN_CONT "(%s)\n",
3086                                 (ret < 0) ? "cannot get the path" : path);
3087                 kfree(path);
3088         }
3089 }
3090 #endif
3091
3092 static DEFINE_MUTEX(set_limit_mutex);
3093
3094 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3095                                 unsigned long long val)
3096 {
3097         int retry_count;
3098         u64 memswlimit, memlimit;
3099         int ret = 0;
3100         int children = mem_cgroup_count_children(memcg);
3101         u64 curusage, oldusage;
3102         int enlarge;
3103
3104         /*
3105          * For keeping hierarchical_reclaim simple, how long we should retry
3106          * is depends on callers. We set our retry-count to be function
3107          * of # of children which we should visit in this loop.
3108          */
3109         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3110
3111         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3112
3113         enlarge = 0;
3114         while (retry_count) {
3115                 if (signal_pending(current)) {
3116                         ret = -EINTR;
3117                         break;
3118                 }
3119                 /*
3120                  * Rather than hide all in some function, I do this in
3121                  * open coded manner. You see what this really does.
3122                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3123                  */
3124                 mutex_lock(&set_limit_mutex);
3125                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3126                 if (memswlimit < val) {
3127                         ret = -EINVAL;
3128                         mutex_unlock(&set_limit_mutex);
3129                         break;
3130                 }
3131
3132                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3133                 if (memlimit < val)
3134                         enlarge = 1;
3135
3136                 ret = res_counter_set_limit(&memcg->res, val);
3137                 if (!ret) {
3138                         if (memswlimit == val)
3139                                 memcg->memsw_is_minimum = true;
3140                         else
3141                                 memcg->memsw_is_minimum = false;
3142                 }
3143                 mutex_unlock(&set_limit_mutex);
3144
3145                 if (!ret)
3146                         break;
3147
3148                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3149                                                 MEM_CGROUP_RECLAIM_SHRINK);
3150                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3151                 /* Usage is reduced ? */
3152                 if (curusage >= oldusage)
3153                         retry_count--;
3154                 else
3155                         oldusage = curusage;
3156         }
3157         if (!ret && enlarge)
3158                 memcg_oom_recover(memcg);
3159
3160         return ret;
3161 }
3162
3163 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3164                                         unsigned long long val)
3165 {
3166         int retry_count;
3167         u64 memlimit, memswlimit, oldusage, curusage;
3168         int children = mem_cgroup_count_children(memcg);
3169         int ret = -EBUSY;
3170         int enlarge = 0;
3171
3172         /* see mem_cgroup_resize_res_limit */
3173         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3174         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3175         while (retry_count) {
3176                 if (signal_pending(current)) {
3177                         ret = -EINTR;
3178                         break;
3179                 }
3180                 /*
3181                  * Rather than hide all in some function, I do this in
3182                  * open coded manner. You see what this really does.
3183                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3184                  */
3185                 mutex_lock(&set_limit_mutex);
3186                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3187                 if (memlimit > val) {
3188                         ret = -EINVAL;
3189                         mutex_unlock(&set_limit_mutex);
3190                         break;
3191                 }
3192                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3193                 if (memswlimit < val)
3194                         enlarge = 1;
3195                 ret = res_counter_set_limit(&memcg->memsw, val);
3196                 if (!ret) {
3197                         if (memlimit == val)
3198                                 memcg->memsw_is_minimum = true;
3199                         else
3200                                 memcg->memsw_is_minimum = false;
3201                 }
3202                 mutex_unlock(&set_limit_mutex);
3203
3204                 if (!ret)
3205                         break;
3206
3207                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3208                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3209                                                 MEM_CGROUP_RECLAIM_SHRINK);
3210                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3211                 /* Usage is reduced ? */
3212                 if (curusage >= oldusage)
3213                         retry_count--;
3214                 else
3215                         oldusage = curusage;
3216         }
3217         if (!ret && enlarge)
3218                 memcg_oom_recover(memcg);
3219         return ret;
3220 }
3221
3222 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3223                                             gfp_t gfp_mask)
3224 {
3225         unsigned long nr_reclaimed = 0;
3226         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3227         unsigned long reclaimed;
3228         int loop = 0;
3229         struct mem_cgroup_tree_per_zone *mctz;
3230         unsigned long long excess;
3231
3232         if (order > 0)
3233                 return 0;
3234
3235         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3236         /*
3237          * This loop can run a while, specially if mem_cgroup's continuously
3238          * keep exceeding their soft limit and putting the system under
3239          * pressure
3240          */
3241         do {
3242                 if (next_mz)
3243                         mz = next_mz;
3244                 else
3245                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3246                 if (!mz)
3247                         break;
3248
3249                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3250                                                 gfp_mask,
3251                                                 MEM_CGROUP_RECLAIM_SOFT);
3252                 nr_reclaimed += reclaimed;
3253                 spin_lock(&mctz->lock);
3254
3255                 /*
3256                  * If we failed to reclaim anything from this memory cgroup
3257                  * it is time to move on to the next cgroup
3258                  */
3259                 next_mz = NULL;
3260                 if (!reclaimed) {
3261                         do {
3262                                 /*
3263                                  * Loop until we find yet another one.
3264                                  *
3265                                  * By the time we get the soft_limit lock
3266                                  * again, someone might have aded the
3267                                  * group back on the RB tree. Iterate to
3268                                  * make sure we get a different mem.
3269                                  * mem_cgroup_largest_soft_limit_node returns
3270                                  * NULL if no other cgroup is present on
3271                                  * the tree
3272                                  */
3273                                 next_mz =
3274                                 __mem_cgroup_largest_soft_limit_node(mctz);
3275                                 if (next_mz == mz) {
3276                                         css_put(&next_mz->mem->css);
3277                                         next_mz = NULL;
3278                                 } else /* next_mz == NULL or other memcg */
3279                                         break;
3280                         } while (1);
3281                 }
3282                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3283                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3284                 /*
3285                  * One school of thought says that we should not add
3286                  * back the node to the tree if reclaim returns 0.
3287                  * But our reclaim could return 0, simply because due
3288                  * to priority we are exposing a smaller subset of
3289                  * memory to reclaim from. Consider this as a longer
3290                  * term TODO.
3291                  */
3292                 /* If excess == 0, no tree ops */
3293                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3294                 spin_unlock(&mctz->lock);
3295                 css_put(&mz->mem->css);
3296                 loop++;
3297                 /*
3298                  * Could not reclaim anything and there are no more
3299                  * mem cgroups to try or we seem to be looping without
3300                  * reclaiming anything.
3301                  */
3302                 if (!nr_reclaimed &&
3303                         (next_mz == NULL ||
3304                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3305                         break;
3306         } while (!nr_reclaimed);
3307         if (next_mz)
3308                 css_put(&next_mz->mem->css);
3309         return nr_reclaimed;
3310 }
3311
3312 /*
3313  * This routine traverse page_cgroup in given list and drop them all.
3314  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3315  */
3316 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3317                                 int node, int zid, enum lru_list lru)
3318 {
3319         struct zone *zone;
3320         struct mem_cgroup_per_zone *mz;
3321         struct page_cgroup *pc, *busy;
3322         unsigned long flags, loop;
3323         struct list_head *list;
3324         int ret = 0;
3325
3326         zone = &NODE_DATA(node)->node_zones[zid];
3327         mz = mem_cgroup_zoneinfo(mem, node, zid);
3328         list = &mz->lists[lru];
3329
3330         loop = MEM_CGROUP_ZSTAT(mz, lru);
3331         /* give some margin against EBUSY etc...*/
3332         loop += 256;
3333         busy = NULL;
3334         while (loop--) {
3335                 ret = 0;
3336                 spin_lock_irqsave(&zone->lru_lock, flags);
3337                 if (list_empty(list)) {
3338                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3339                         break;
3340                 }
3341                 pc = list_entry(list->prev, struct page_cgroup, lru);
3342                 if (busy == pc) {
3343                         list_move(&pc->lru, list);
3344                         busy = NULL;
3345                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3346                         continue;
3347                 }
3348                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3349
3350                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3351                 if (ret == -ENOMEM)
3352                         break;
3353
3354                 if (ret == -EBUSY || ret == -EINVAL) {
3355                         /* found lock contention or "pc" is obsolete. */
3356                         busy = pc;
3357                         cond_resched();
3358                 } else
3359                         busy = NULL;
3360         }
3361
3362         if (!ret && !list_empty(list))
3363                 return -EBUSY;
3364         return ret;
3365 }
3366
3367 /*
3368  * make mem_cgroup's charge to be 0 if there is no task.
3369  * This enables deleting this mem_cgroup.
3370  */
3371 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3372 {
3373         int ret;
3374         int node, zid, shrink;
3375         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3376         struct cgroup *cgrp = mem->css.cgroup;
3377
3378         css_get(&mem->css);
3379
3380         shrink = 0;
3381         /* should free all ? */
3382         if (free_all)
3383                 goto try_to_free;
3384 move_account:
3385         do {
3386                 ret = -EBUSY;
3387                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3388                         goto out;
3389                 ret = -EINTR;
3390                 if (signal_pending(current))
3391                         goto out;
3392                 /* This is for making all *used* pages to be on LRU. */
3393                 lru_add_drain_all();
3394                 drain_all_stock_sync();
3395                 ret = 0;
3396                 mem_cgroup_start_move(mem);
3397                 for_each_node_state(node, N_HIGH_MEMORY) {
3398                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3399                                 enum lru_list l;
3400                                 for_each_lru(l) {
3401                                         ret = mem_cgroup_force_empty_list(mem,
3402                                                         node, zid, l);
3403                                         if (ret)
3404                                                 break;
3405                                 }
3406                         }
3407                         if (ret)
3408                                 break;
3409                 }
3410                 mem_cgroup_end_move(mem);
3411                 memcg_oom_recover(mem);
3412                 /* it seems parent cgroup doesn't have enough mem */
3413                 if (ret == -ENOMEM)
3414                         goto try_to_free;
3415                 cond_resched();
3416         /* "ret" should also be checked to ensure all lists are empty. */
3417         } while (mem->res.usage > 0 || ret);
3418 out:
3419         css_put(&mem->css);
3420         return ret;
3421
3422 try_to_free:
3423         /* returns EBUSY if there is a task or if we come here twice. */
3424         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3425                 ret = -EBUSY;
3426                 goto out;
3427         }
3428         /* we call try-to-free pages for make this cgroup empty */
3429         lru_add_drain_all();
3430         /* try to free all pages in this cgroup */
3431         shrink = 1;
3432         while (nr_retries && mem->res.usage > 0) {
3433                 int progress;
3434
3435                 if (signal_pending(current)) {
3436                         ret = -EINTR;
3437                         goto out;
3438                 }
3439                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3440                                                 false, get_swappiness(mem));
3441                 if (!progress) {
3442                         nr_retries--;
3443                         /* maybe some writeback is necessary */
3444                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3445                 }
3446
3447         }
3448         lru_add_drain();
3449         /* try move_account...there may be some *locked* pages. */
3450         goto move_account;
3451 }
3452
3453 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3454 {
3455         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3456 }
3457
3458
3459 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3460 {
3461         return mem_cgroup_from_cont(cont)->use_hierarchy;
3462 }
3463
3464 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3465                                         u64 val)
3466 {
3467         int retval = 0;
3468         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3469         struct cgroup *parent = cont->parent;
3470         struct mem_cgroup *parent_mem = NULL;
3471
3472         if (parent)
3473                 parent_mem = mem_cgroup_from_cont(parent);
3474
3475         cgroup_lock();
3476         /*
3477          * If parent's use_hierarchy is set, we can't make any modifications
3478          * in the child subtrees. If it is unset, then the change can
3479          * occur, provided the current cgroup has no children.
3480          *
3481          * For the root cgroup, parent_mem is NULL, we allow value to be
3482          * set if there are no children.
3483          */
3484         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3485                                 (val == 1 || val == 0)) {
3486                 if (list_empty(&cont->children))
3487                         mem->use_hierarchy = val;
3488                 else
3489                         retval = -EBUSY;
3490         } else
3491                 retval = -EINVAL;
3492         cgroup_unlock();
3493
3494         return retval;
3495 }
3496
3497
3498 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3499                                 enum mem_cgroup_stat_index idx)
3500 {
3501         struct mem_cgroup *iter;
3502         s64 val = 0;
3503
3504         /* each per cpu's value can be minus.Then, use s64 */
3505         for_each_mem_cgroup_tree(iter, mem)
3506                 val += mem_cgroup_read_stat(iter, idx);
3507
3508         if (val < 0) /* race ? */
3509                 val = 0;
3510         return val;
3511 }
3512
3513 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3514 {
3515         u64 val;
3516
3517         if (!mem_cgroup_is_root(mem)) {
3518                 if (!swap)
3519                         return res_counter_read_u64(&mem->res, RES_USAGE);
3520                 else
3521                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3522         }
3523
3524         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3525         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3526
3527         if (swap)
3528                 val += mem_cgroup_get_recursive_idx_stat(mem,
3529                                 MEM_CGROUP_STAT_SWAPOUT);
3530
3531         return val << PAGE_SHIFT;
3532 }
3533
3534 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3535 {
3536         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3537         u64 val;
3538         int type, name;
3539
3540         type = MEMFILE_TYPE(cft->private);
3541         name = MEMFILE_ATTR(cft->private);
3542         switch (type) {
3543         case _MEM:
3544                 if (name == RES_USAGE)
3545                         val = mem_cgroup_usage(mem, false);
3546                 else
3547                         val = res_counter_read_u64(&mem->res, name);
3548                 break;
3549         case _MEMSWAP:
3550                 if (name == RES_USAGE)
3551                         val = mem_cgroup_usage(mem, true);
3552                 else
3553                         val = res_counter_read_u64(&mem->memsw, name);
3554                 break;
3555         default:
3556                 BUG();
3557                 break;
3558         }
3559         return val;
3560 }
3561 /*
3562  * The user of this function is...
3563  * RES_LIMIT.
3564  */
3565 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3566                             const char *buffer)
3567 {
3568         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3569         int type, name;
3570         unsigned long long val;
3571         int ret;
3572
3573         type = MEMFILE_TYPE(cft->private);
3574         name = MEMFILE_ATTR(cft->private);
3575         switch (name) {
3576         case RES_LIMIT:
3577                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3578                         ret = -EINVAL;
3579                         break;
3580                 }
3581                 /* This function does all necessary parse...reuse it */
3582                 ret = res_counter_memparse_write_strategy(buffer, &val);
3583                 if (ret)
3584                         break;
3585                 if (type == _MEM)
3586                         ret = mem_cgroup_resize_limit(memcg, val);
3587                 else
3588                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3589                 break;
3590         case RES_SOFT_LIMIT:
3591                 ret = res_counter_memparse_write_strategy(buffer, &val);
3592                 if (ret)
3593                         break;
3594                 /*
3595                  * For memsw, soft limits are hard to implement in terms
3596                  * of semantics, for now, we support soft limits for
3597                  * control without swap
3598                  */
3599                 if (type == _MEM)
3600                         ret = res_counter_set_soft_limit(&memcg->res, val);
3601                 else
3602                         ret = -EINVAL;
3603                 break;
3604         default:
3605                 ret = -EINVAL; /* should be BUG() ? */
3606                 break;
3607         }
3608         return ret;
3609 }
3610
3611 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3612                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3613 {
3614         struct cgroup *cgroup;
3615         unsigned long long min_limit, min_memsw_limit, tmp;
3616
3617         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3618         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3619         cgroup = memcg->css.cgroup;
3620         if (!memcg->use_hierarchy)
3621                 goto out;
3622
3623         while (cgroup->parent) {
3624                 cgroup = cgroup->parent;
3625                 memcg = mem_cgroup_from_cont(cgroup);
3626                 if (!memcg->use_hierarchy)
3627                         break;
3628                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3629                 min_limit = min(min_limit, tmp);
3630                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3631                 min_memsw_limit = min(min_memsw_limit, tmp);
3632         }
3633 out:
3634         *mem_limit = min_limit;
3635         *memsw_limit = min_memsw_limit;
3636         return;
3637 }
3638
3639 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3640 {
3641         struct mem_cgroup *mem;
3642         int type, name;
3643
3644         mem = mem_cgroup_from_cont(cont);
3645         type = MEMFILE_TYPE(event);
3646         name = MEMFILE_ATTR(event);
3647         switch (name) {
3648         case RES_MAX_USAGE:
3649                 if (type == _MEM)
3650                         res_counter_reset_max(&mem->res);
3651                 else
3652                         res_counter_reset_max(&mem->memsw);
3653                 break;
3654         case RES_FAILCNT:
3655                 if (type == _MEM)
3656                         res_counter_reset_failcnt(&mem->res);
3657                 else
3658                         res_counter_reset_failcnt(&mem->memsw);
3659                 break;
3660         }
3661
3662         return 0;
3663 }
3664
3665 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3666                                         struct cftype *cft)
3667 {
3668         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3669 }
3670
3671 #ifdef CONFIG_MMU
3672 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3673                                         struct cftype *cft, u64 val)
3674 {
3675         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3676
3677         if (val >= (1 << NR_MOVE_TYPE))
3678                 return -EINVAL;
3679         /*
3680          * We check this value several times in both in can_attach() and
3681          * attach(), so we need cgroup lock to prevent this value from being
3682          * inconsistent.
3683          */
3684         cgroup_lock();
3685         mem->move_charge_at_immigrate = val;
3686         cgroup_unlock();
3687
3688         return 0;
3689 }
3690 #else
3691 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3692                                         struct cftype *cft, u64 val)
3693 {
3694         return -ENOSYS;
3695 }
3696 #endif
3697
3698
3699 /* For read statistics */
3700 enum {
3701         MCS_CACHE,
3702         MCS_RSS,
3703         MCS_FILE_MAPPED,
3704         MCS_PGPGIN,
3705         MCS_PGPGOUT,
3706         MCS_SWAP,
3707         MCS_INACTIVE_ANON,
3708         MCS_ACTIVE_ANON,
3709         MCS_INACTIVE_FILE,
3710         MCS_ACTIVE_FILE,
3711         MCS_UNEVICTABLE,
3712         NR_MCS_STAT,
3713 };
3714
3715 struct mcs_total_stat {
3716         s64 stat[NR_MCS_STAT];
3717 };
3718
3719 struct {
3720         char *local_name;
3721         char *total_name;
3722 } memcg_stat_strings[NR_MCS_STAT] = {
3723         {"cache", "total_cache"},
3724         {"rss", "total_rss"},
3725         {"mapped_file", "total_mapped_file"},
3726         {"pgpgin", "total_pgpgin"},
3727         {"pgpgout", "total_pgpgout"},
3728         {"swap", "total_swap"},
3729         {"inactive_anon", "total_inactive_anon"},
3730         {"active_anon", "total_active_anon"},
3731         {"inactive_file", "total_inactive_file"},
3732         {"active_file", "total_active_file"},
3733         {"unevictable", "total_unevictable"}
3734 };
3735
3736
3737 static void
3738 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3739 {
3740         s64 val;
3741
3742         /* per cpu stat */
3743         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3744         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3745         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3746         s->stat[MCS_RSS] += val * PAGE_SIZE;
3747         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3748         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3749         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3750         s->stat[MCS_PGPGIN] += val;
3751         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3752         s->stat[MCS_PGPGOUT] += val;
3753         if (do_swap_account) {
3754                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3755                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3756         }
3757
3758         /* per zone stat */
3759         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3760         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3761         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3762         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3763         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3764         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3765         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3766         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3767         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3768         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3769 }
3770
3771 static void
3772 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3773 {
3774         struct mem_cgroup *iter;
3775
3776         for_each_mem_cgroup_tree(iter, mem)
3777                 mem_cgroup_get_local_stat(iter, s);
3778 }
3779
3780 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3781                                  struct cgroup_map_cb *cb)
3782 {
3783         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3784         struct mcs_total_stat mystat;
3785         int i;
3786
3787         memset(&mystat, 0, sizeof(mystat));
3788         mem_cgroup_get_local_stat(mem_cont, &mystat);
3789
3790         for (i = 0; i < NR_MCS_STAT; i++) {
3791                 if (i == MCS_SWAP && !do_swap_account)
3792                         continue;
3793                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3794         }
3795
3796         /* Hierarchical information */
3797         {
3798                 unsigned long long limit, memsw_limit;
3799                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3800                 cb->fill(cb, "hierarchical_memory_limit", limit);
3801                 if (do_swap_account)
3802                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3803         }
3804
3805         memset(&mystat, 0, sizeof(mystat));
3806         mem_cgroup_get_total_stat(mem_cont, &mystat);
3807         for (i = 0; i < NR_MCS_STAT; i++) {
3808                 if (i == MCS_SWAP && !do_swap_account)
3809                         continue;
3810                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3811         }
3812
3813 #ifdef CONFIG_DEBUG_VM
3814         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3815
3816         {
3817                 int nid, zid;
3818                 struct mem_cgroup_per_zone *mz;
3819                 unsigned long recent_rotated[2] = {0, 0};
3820                 unsigned long recent_scanned[2] = {0, 0};
3821
3822                 for_each_online_node(nid)
3823                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3824                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3825
3826                                 recent_rotated[0] +=
3827                                         mz->reclaim_stat.recent_rotated[0];
3828                                 recent_rotated[1] +=
3829                                         mz->reclaim_stat.recent_rotated[1];
3830                                 recent_scanned[0] +=
3831                                         mz->reclaim_stat.recent_scanned[0];
3832                                 recent_scanned[1] +=
3833                                         mz->reclaim_stat.recent_scanned[1];
3834                         }
3835                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3836                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3837                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3838                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3839         }
3840 #endif
3841
3842         return 0;
3843 }
3844
3845 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3846 {
3847         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3848
3849         return get_swappiness(memcg);
3850 }
3851
3852 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3853                                        u64 val)
3854 {
3855         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3856         struct mem_cgroup *parent;
3857
3858         if (val > 100)
3859                 return -EINVAL;
3860
3861         if (cgrp->parent == NULL)
3862                 return -EINVAL;
3863
3864         parent = mem_cgroup_from_cont(cgrp->parent);
3865
3866         cgroup_lock();
3867
3868         /* If under hierarchy, only empty-root can set this value */
3869         if ((parent->use_hierarchy) ||
3870             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3871                 cgroup_unlock();
3872                 return -EINVAL;
3873         }
3874
3875         spin_lock(&memcg->reclaim_param_lock);
3876         memcg->swappiness = val;
3877         spin_unlock(&memcg->reclaim_param_lock);
3878
3879         cgroup_unlock();
3880
3881         return 0;
3882 }
3883
3884 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3885 {
3886         struct mem_cgroup_threshold_ary *t;
3887         u64 usage;
3888         int i;
3889
3890         rcu_read_lock();
3891         if (!swap)
3892                 t = rcu_dereference(memcg->thresholds.primary);
3893         else
3894                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3895
3896         if (!t)
3897                 goto unlock;
3898
3899         usage = mem_cgroup_usage(memcg, swap);
3900
3901         /*
3902          * current_threshold points to threshold just below usage.
3903          * If it's not true, a threshold was crossed after last
3904          * call of __mem_cgroup_threshold().
3905          */
3906         i = t->current_threshold;
3907
3908         /*
3909          * Iterate backward over array of thresholds starting from
3910          * current_threshold and check if a threshold is crossed.
3911          * If none of thresholds below usage is crossed, we read
3912          * only one element of the array here.
3913          */
3914         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3915                 eventfd_signal(t->entries[i].eventfd, 1);
3916
3917         /* i = current_threshold + 1 */
3918         i++;
3919
3920         /*
3921          * Iterate forward over array of thresholds starting from
3922          * current_threshold+1 and check if a threshold is crossed.
3923          * If none of thresholds above usage is crossed, we read
3924          * only one element of the array here.
3925          */
3926         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3927                 eventfd_signal(t->entries[i].eventfd, 1);
3928
3929         /* Update current_threshold */
3930         t->current_threshold = i - 1;
3931 unlock:
3932         rcu_read_unlock();
3933 }
3934
3935 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3936 {
3937         while (memcg) {
3938                 __mem_cgroup_threshold(memcg, false);
3939                 if (do_swap_account)
3940                         __mem_cgroup_threshold(memcg, true);
3941
3942                 memcg = parent_mem_cgroup(memcg);
3943         }
3944 }
3945
3946 static int compare_thresholds(const void *a, const void *b)
3947 {
3948         const struct mem_cgroup_threshold *_a = a;
3949         const struct mem_cgroup_threshold *_b = b;
3950
3951         return _a->threshold - _b->threshold;
3952 }
3953
3954 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3955 {
3956         struct mem_cgroup_eventfd_list *ev;
3957
3958         list_for_each_entry(ev, &mem->oom_notify, list)
3959                 eventfd_signal(ev->eventfd, 1);
3960         return 0;
3961 }
3962
3963 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3964 {
3965         struct mem_cgroup *iter;
3966
3967         for_each_mem_cgroup_tree(iter, mem)
3968                 mem_cgroup_oom_notify_cb(iter);
3969 }
3970
3971 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3972         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3973 {
3974         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3975         struct mem_cgroup_thresholds *thresholds;
3976         struct mem_cgroup_threshold_ary *new;
3977         int type = MEMFILE_TYPE(cft->private);
3978         u64 threshold, usage;
3979         int i, size, ret;
3980
3981         ret = res_counter_memparse_write_strategy(args, &threshold);
3982         if (ret)
3983                 return ret;
3984
3985         mutex_lock(&memcg->thresholds_lock);
3986
3987         if (type == _MEM)
3988                 thresholds = &memcg->thresholds;
3989         else if (type == _MEMSWAP)
3990                 thresholds = &memcg->memsw_thresholds;
3991         else
3992                 BUG();
3993
3994         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3995
3996         /* Check if a threshold crossed before adding a new one */
3997         if (thresholds->primary)
3998                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3999
4000         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4001
4002         /* Allocate memory for new array of thresholds */
4003         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4004                         GFP_KERNEL);
4005         if (!new) {
4006                 ret = -ENOMEM;
4007                 goto unlock;
4008         }
4009         new->size = size;
4010
4011         /* Copy thresholds (if any) to new array */
4012         if (thresholds->primary) {
4013                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4014                                 sizeof(struct mem_cgroup_threshold));
4015         }
4016
4017         /* Add new threshold */
4018         new->entries[size - 1].eventfd = eventfd;
4019         new->entries[size - 1].threshold = threshold;
4020
4021         /* Sort thresholds. Registering of new threshold isn't time-critical */
4022         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4023                         compare_thresholds, NULL);
4024
4025         /* Find current threshold */
4026         new->current_threshold = -1;
4027         for (i = 0; i < size; i++) {
4028                 if (new->entries[i].threshold < usage) {
4029                         /*
4030                          * new->current_threshold will not be used until
4031                          * rcu_assign_pointer(), so it's safe to increment
4032                          * it here.
4033                          */
4034                         ++new->current_threshold;
4035                 }
4036         }
4037
4038         /* Free old spare buffer and save old primary buffer as spare */
4039         kfree(thresholds->spare);
4040         thresholds->spare = thresholds->primary;
4041
4042         rcu_assign_pointer(thresholds->primary, new);
4043
4044         /* To be sure that nobody uses thresholds */
4045         synchronize_rcu();
4046
4047 unlock:
4048         mutex_unlock(&memcg->thresholds_lock);
4049
4050         return ret;
4051 }
4052
4053 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4054         struct cftype *cft, struct eventfd_ctx *eventfd)
4055 {
4056         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4057         struct mem_cgroup_thresholds *thresholds;
4058         struct mem_cgroup_threshold_ary *new;
4059         int type = MEMFILE_TYPE(cft->private);
4060         u64 usage;
4061         int i, j, size;
4062
4063         mutex_lock(&memcg->thresholds_lock);
4064         if (type == _MEM)
4065                 thresholds = &memcg->thresholds;
4066         else if (type == _MEMSWAP)
4067                 thresholds = &memcg->memsw_thresholds;
4068         else
4069                 BUG();
4070
4071         /*
4072          * Something went wrong if we trying to unregister a threshold
4073          * if we don't have thresholds
4074          */
4075         BUG_ON(!thresholds);
4076
4077         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4078
4079         /* Check if a threshold crossed before removing */
4080         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4081
4082         /* Calculate new number of threshold */
4083         size = 0;
4084         for (i = 0; i < thresholds->primary->size; i++) {
4085                 if (thresholds->primary->entries[i].eventfd != eventfd)
4086                         size++;
4087         }
4088
4089         new = thresholds->spare;
4090
4091         /* Set thresholds array to NULL if we don't have thresholds */
4092         if (!size) {
4093                 kfree(new);
4094                 new = NULL;
4095                 goto swap_buffers;
4096         }
4097
4098         new->size = size;
4099
4100         /* Copy thresholds and find current threshold */
4101         new->current_threshold = -1;
4102         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4103                 if (thresholds->primary->entries[i].eventfd == eventfd)
4104                         continue;
4105
4106                 new->entries[j] = thresholds->primary->entries[i];
4107                 if (new->entries[j].threshold < usage) {
4108                         /*
4109                          * new->current_threshold will not be used
4110                          * until rcu_assign_pointer(), so it's safe to increment
4111                          * it here.
4112                          */
4113                         ++new->current_threshold;
4114                 }
4115                 j++;
4116         }
4117
4118 swap_buffers:
4119         /* Swap primary and spare array */
4120         thresholds->spare = thresholds->primary;
4121         rcu_assign_pointer(thresholds->primary, new);
4122
4123         /* To be sure that nobody uses thresholds */
4124         synchronize_rcu();
4125
4126         mutex_unlock(&memcg->thresholds_lock);
4127 }
4128
4129 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4130         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4131 {
4132         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4133         struct mem_cgroup_eventfd_list *event;
4134         int type = MEMFILE_TYPE(cft->private);
4135
4136         BUG_ON(type != _OOM_TYPE);
4137         event = kmalloc(sizeof(*event), GFP_KERNEL);
4138         if (!event)
4139                 return -ENOMEM;
4140
4141         mutex_lock(&memcg_oom_mutex);
4142
4143         event->eventfd = eventfd;
4144         list_add(&event->list, &memcg->oom_notify);
4145
4146         /* already in OOM ? */
4147         if (atomic_read(&memcg->oom_lock))
4148                 eventfd_signal(eventfd, 1);
4149         mutex_unlock(&memcg_oom_mutex);
4150
4151         return 0;
4152 }
4153
4154 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4155         struct cftype *cft, struct eventfd_ctx *eventfd)
4156 {
4157         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4158         struct mem_cgroup_eventfd_list *ev, *tmp;
4159         int type = MEMFILE_TYPE(cft->private);
4160
4161         BUG_ON(type != _OOM_TYPE);
4162
4163         mutex_lock(&memcg_oom_mutex);
4164
4165         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4166                 if (ev->eventfd == eventfd) {
4167                         list_del(&ev->list);
4168                         kfree(ev);
4169                 }
4170         }
4171
4172         mutex_unlock(&memcg_oom_mutex);
4173 }
4174
4175 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4176         struct cftype *cft,  struct cgroup_map_cb *cb)
4177 {
4178         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4179
4180         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4181
4182         if (atomic_read(&mem->oom_lock))
4183                 cb->fill(cb, "under_oom", 1);
4184         else
4185                 cb->fill(cb, "under_oom", 0);
4186         return 0;
4187 }
4188
4189 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4190         struct cftype *cft, u64 val)
4191 {
4192         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4193         struct mem_cgroup *parent;
4194
4195         /* cannot set to root cgroup and only 0 and 1 are allowed */
4196         if (!cgrp->parent || !((val == 0) || (val == 1)))
4197                 return -EINVAL;
4198
4199         parent = mem_cgroup_from_cont(cgrp->parent);
4200
4201         cgroup_lock();
4202         /* oom-kill-disable is a flag for subhierarchy. */
4203         if ((parent->use_hierarchy) ||
4204             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4205                 cgroup_unlock();
4206                 return -EINVAL;
4207         }
4208         mem->oom_kill_disable = val;
4209         if (!val)
4210                 memcg_oom_recover(mem);
4211         cgroup_unlock();
4212         return 0;
4213 }
4214
4215 static struct cftype mem_cgroup_files[] = {
4216         {
4217                 .name = "usage_in_bytes",
4218                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4219                 .read_u64 = mem_cgroup_read,
4220                 .register_event = mem_cgroup_usage_register_event,
4221                 .unregister_event = mem_cgroup_usage_unregister_event,
4222         },
4223         {
4224                 .name = "max_usage_in_bytes",
4225                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4226                 .trigger = mem_cgroup_reset,
4227                 .read_u64 = mem_cgroup_read,
4228         },
4229         {
4230                 .name = "limit_in_bytes",
4231                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4232                 .write_string = mem_cgroup_write,
4233                 .read_u64 = mem_cgroup_read,
4234         },
4235         {
4236                 .name = "soft_limit_in_bytes",
4237                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4238                 .write_string = mem_cgroup_write,
4239                 .read_u64 = mem_cgroup_read,
4240         },
4241         {
4242                 .name = "failcnt",
4243                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4244                 .trigger = mem_cgroup_reset,
4245                 .read_u64 = mem_cgroup_read,
4246         },
4247         {
4248                 .name = "stat",
4249                 .read_map = mem_control_stat_show,
4250         },
4251         {
4252                 .name = "force_empty",
4253                 .trigger = mem_cgroup_force_empty_write,
4254         },
4255         {
4256                 .name = "use_hierarchy",
4257                 .write_u64 = mem_cgroup_hierarchy_write,
4258                 .read_u64 = mem_cgroup_hierarchy_read,
4259         },
4260         {
4261                 .name = "swappiness",
4262                 .read_u64 = mem_cgroup_swappiness_read,
4263                 .write_u64 = mem_cgroup_swappiness_write,
4264         },
4265         {
4266                 .name = "move_charge_at_immigrate",
4267                 .read_u64 = mem_cgroup_move_charge_read,
4268                 .write_u64 = mem_cgroup_move_charge_write,
4269         },
4270         {
4271                 .name = "oom_control",
4272                 .read_map = mem_cgroup_oom_control_read,
4273                 .write_u64 = mem_cgroup_oom_control_write,
4274                 .register_event = mem_cgroup_oom_register_event,
4275                 .unregister_event = mem_cgroup_oom_unregister_event,
4276                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4277         },
4278 };
4279
4280 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4281 static struct cftype memsw_cgroup_files[] = {
4282         {
4283                 .name = "memsw.usage_in_bytes",
4284                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4285                 .read_u64 = mem_cgroup_read,
4286                 .register_event = mem_cgroup_usage_register_event,
4287                 .unregister_event = mem_cgroup_usage_unregister_event,
4288         },
4289         {
4290                 .name = "memsw.max_usage_in_bytes",
4291                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4292                 .trigger = mem_cgroup_reset,
4293                 .read_u64 = mem_cgroup_read,
4294         },
4295         {
4296                 .name = "memsw.limit_in_bytes",
4297                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4298                 .write_string = mem_cgroup_write,
4299                 .read_u64 = mem_cgroup_read,
4300         },
4301         {
4302                 .name = "memsw.failcnt",
4303                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4304                 .trigger = mem_cgroup_reset,
4305                 .read_u64 = mem_cgroup_read,
4306         },
4307 };
4308
4309 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4310 {
4311         if (!do_swap_account)
4312                 return 0;
4313         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4314                                 ARRAY_SIZE(memsw_cgroup_files));
4315 };
4316 #else
4317 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4318 {
4319         return 0;
4320 }
4321 #endif
4322
4323 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4324 {
4325         struct mem_cgroup_per_node *pn;
4326         struct mem_cgroup_per_zone *mz;
4327         enum lru_list l;
4328         int zone, tmp = node;
4329         /*
4330          * This routine is called against possible nodes.
4331          * But it's BUG to call kmalloc() against offline node.
4332          *
4333          * TODO: this routine can waste much memory for nodes which will
4334          *       never be onlined. It's better to use memory hotplug callback
4335          *       function.
4336          */
4337         if (!node_state(node, N_NORMAL_MEMORY))
4338                 tmp = -1;
4339         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4340         if (!pn)
4341                 return 1;
4342
4343         mem->info.nodeinfo[node] = pn;
4344         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4345                 mz = &pn->zoneinfo[zone];
4346                 for_each_lru(l)
4347                         INIT_LIST_HEAD(&mz->lists[l]);
4348                 mz->usage_in_excess = 0;
4349                 mz->on_tree = false;
4350                 mz->mem = mem;
4351         }
4352         return 0;
4353 }
4354
4355 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4356 {
4357         kfree(mem->info.nodeinfo[node]);
4358 }
4359
4360 static struct mem_cgroup *mem_cgroup_alloc(void)
4361 {
4362         struct mem_cgroup *mem;
4363         int size = sizeof(struct mem_cgroup);
4364
4365         /* Can be very big if MAX_NUMNODES is very big */
4366         if (size < PAGE_SIZE)
4367                 mem = kzalloc(size, GFP_KERNEL);
4368         else
4369                 mem = vzalloc(size);
4370
4371         if (!mem)
4372                 return NULL;
4373
4374         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4375         if (!mem->stat)
4376                 goto out_free;
4377         spin_lock_init(&mem->pcp_counter_lock);
4378         return mem;
4379
4380 out_free:
4381         if (size < PAGE_SIZE)
4382                 kfree(mem);
4383         else
4384                 vfree(mem);
4385         return NULL;
4386 }
4387
4388 /*
4389  * At destroying mem_cgroup, references from swap_cgroup can remain.
4390  * (scanning all at force_empty is too costly...)
4391  *
4392  * Instead of clearing all references at force_empty, we remember
4393  * the number of reference from swap_cgroup and free mem_cgroup when
4394  * it goes down to 0.
4395  *
4396  * Removal of cgroup itself succeeds regardless of refs from swap.
4397  */
4398
4399 static void __mem_cgroup_free(struct mem_cgroup *mem)
4400 {
4401         int node;
4402
4403         mem_cgroup_remove_from_trees(mem);
4404         free_css_id(&mem_cgroup_subsys, &mem->css);
4405
4406         for_each_node_state(node, N_POSSIBLE)
4407                 free_mem_cgroup_per_zone_info(mem, node);
4408
4409         free_percpu(mem->stat);
4410         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4411                 kfree(mem);
4412         else
4413                 vfree(mem);
4414 }
4415
4416 static void mem_cgroup_get(struct mem_cgroup *mem)
4417 {
4418         atomic_inc(&mem->refcnt);
4419 }
4420
4421 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4422 {
4423         if (atomic_sub_and_test(count, &mem->refcnt)) {
4424                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4425                 __mem_cgroup_free(mem);
4426                 if (parent)
4427                         mem_cgroup_put(parent);
4428         }
4429 }
4430
4431 static void mem_cgroup_put(struct mem_cgroup *mem)
4432 {
4433         __mem_cgroup_put(mem, 1);
4434 }
4435
4436 /*
4437  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4438  */
4439 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4440 {
4441         if (!mem->res.parent)
4442                 return NULL;
4443         return mem_cgroup_from_res_counter(mem->res.parent, res);
4444 }
4445
4446 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4447 static void __init enable_swap_cgroup(void)
4448 {
4449         if (!mem_cgroup_disabled() && really_do_swap_account)
4450                 do_swap_account = 1;
4451 }
4452 #else
4453 static void __init enable_swap_cgroup(void)
4454 {
4455 }
4456 #endif
4457
4458 static int mem_cgroup_soft_limit_tree_init(void)
4459 {
4460         struct mem_cgroup_tree_per_node *rtpn;
4461         struct mem_cgroup_tree_per_zone *rtpz;
4462         int tmp, node, zone;
4463
4464         for_each_node_state(node, N_POSSIBLE) {
4465                 tmp = node;
4466                 if (!node_state(node, N_NORMAL_MEMORY))
4467                         tmp = -1;
4468                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4469                 if (!rtpn)
4470                         return 1;
4471
4472                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4473
4474                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4475                         rtpz = &rtpn->rb_tree_per_zone[zone];
4476                         rtpz->rb_root = RB_ROOT;
4477                         spin_lock_init(&rtpz->lock);
4478                 }
4479         }
4480         return 0;
4481 }
4482
4483 static struct cgroup_subsys_state * __ref
4484 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4485 {
4486         struct mem_cgroup *mem, *parent;
4487         long error = -ENOMEM;
4488         int node;
4489
4490         mem = mem_cgroup_alloc();
4491         if (!mem)
4492                 return ERR_PTR(error);
4493
4494         for_each_node_state(node, N_POSSIBLE)
4495                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4496                         goto free_out;
4497
4498         /* root ? */
4499         if (cont->parent == NULL) {
4500                 int cpu;
4501                 enable_swap_cgroup();
4502                 parent = NULL;
4503                 root_mem_cgroup = mem;
4504                 if (mem_cgroup_soft_limit_tree_init())
4505                         goto free_out;
4506                 for_each_possible_cpu(cpu) {
4507                         struct memcg_stock_pcp *stock =
4508                                                 &per_cpu(memcg_stock, cpu);
4509                         INIT_WORK(&stock->work, drain_local_stock);
4510                 }
4511                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4512         } else {
4513                 parent = mem_cgroup_from_cont(cont->parent);
4514                 mem->use_hierarchy = parent->use_hierarchy;
4515                 mem->oom_kill_disable = parent->oom_kill_disable;
4516         }
4517
4518         if (parent && parent->use_hierarchy) {
4519                 res_counter_init(&mem->res, &parent->res);
4520                 res_counter_init(&mem->memsw, &parent->memsw);
4521                 /*
4522                  * We increment refcnt of the parent to ensure that we can
4523                  * safely access it on res_counter_charge/uncharge.
4524                  * This refcnt will be decremented when freeing this
4525                  * mem_cgroup(see mem_cgroup_put).
4526                  */
4527                 mem_cgroup_get(parent);
4528         } else {
4529                 res_counter_init(&mem->res, NULL);
4530                 res_counter_init(&mem->memsw, NULL);
4531         }
4532         mem->last_scanned_child = 0;
4533         spin_lock_init(&mem->reclaim_param_lock);
4534         INIT_LIST_HEAD(&mem->oom_notify);
4535
4536         if (parent)
4537                 mem->swappiness = get_swappiness(parent);
4538         atomic_set(&mem->refcnt, 1);
4539         mem->move_charge_at_immigrate = 0;
4540         mutex_init(&mem->thresholds_lock);
4541         return &mem->css;
4542 free_out:
4543         __mem_cgroup_free(mem);
4544         root_mem_cgroup = NULL;
4545         return ERR_PTR(error);
4546 }
4547
4548 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4549                                         struct cgroup *cont)
4550 {
4551         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4552
4553         return mem_cgroup_force_empty(mem, false);
4554 }
4555
4556 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4557                                 struct cgroup *cont)
4558 {
4559         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4560
4561         mem_cgroup_put(mem);
4562 }
4563
4564 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4565                                 struct cgroup *cont)
4566 {
4567         int ret;
4568
4569         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4570                                 ARRAY_SIZE(mem_cgroup_files));
4571
4572         if (!ret)
4573                 ret = register_memsw_files(cont, ss);
4574         return ret;
4575 }
4576
4577 #ifdef CONFIG_MMU
4578 /* Handlers for move charge at task migration. */
4579 #define PRECHARGE_COUNT_AT_ONCE 256
4580 static int mem_cgroup_do_precharge(unsigned long count)
4581 {
4582         int ret = 0;
4583         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4584         struct mem_cgroup *mem = mc.to;
4585
4586         if (mem_cgroup_is_root(mem)) {
4587                 mc.precharge += count;
4588                 /* we don't need css_get for root */
4589                 return ret;
4590         }
4591         /* try to charge at once */
4592         if (count > 1) {
4593                 struct res_counter *dummy;
4594                 /*
4595                  * "mem" cannot be under rmdir() because we've already checked
4596                  * by cgroup_lock_live_cgroup() that it is not removed and we
4597                  * are still under the same cgroup_mutex. So we can postpone
4598                  * css_get().
4599                  */
4600                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4601                         goto one_by_one;
4602                 if (do_swap_account && res_counter_charge(&mem->memsw,
4603                                                 PAGE_SIZE * count, &dummy)) {
4604                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4605                         goto one_by_one;
4606                 }
4607                 mc.precharge += count;
4608                 return ret;
4609         }
4610 one_by_one:
4611         /* fall back to one by one charge */
4612         while (count--) {
4613                 if (signal_pending(current)) {
4614                         ret = -EINTR;
4615                         break;
4616                 }
4617                 if (!batch_count--) {
4618                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4619                         cond_resched();
4620                 }
4621                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4622                                               PAGE_SIZE);
4623                 if (ret || !mem)
4624                         /* mem_cgroup_clear_mc() will do uncharge later */
4625                         return -ENOMEM;
4626                 mc.precharge++;
4627         }
4628         return ret;
4629 }
4630
4631 /**
4632  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4633  * @vma: the vma the pte to be checked belongs
4634  * @addr: the address corresponding to the pte to be checked
4635  * @ptent: the pte to be checked
4636  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4637  *
4638  * Returns
4639  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4640  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4641  *     move charge. if @target is not NULL, the page is stored in target->page
4642  *     with extra refcnt got(Callers should handle it).
4643  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4644  *     target for charge migration. if @target is not NULL, the entry is stored
4645  *     in target->ent.
4646  *
4647  * Called with pte lock held.
4648  */
4649 union mc_target {
4650         struct page     *page;
4651         swp_entry_t     ent;
4652 };
4653
4654 enum mc_target_type {
4655         MC_TARGET_NONE, /* not used */
4656         MC_TARGET_PAGE,
4657         MC_TARGET_SWAP,
4658 };
4659
4660 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4661                                                 unsigned long addr, pte_t ptent)
4662 {
4663         struct page *page = vm_normal_page(vma, addr, ptent);
4664
4665         if (!page || !page_mapped(page))
4666                 return NULL;
4667         if (PageAnon(page)) {
4668                 /* we don't move shared anon */
4669                 if (!move_anon() || page_mapcount(page) > 2)
4670                         return NULL;
4671         } else if (!move_file())
4672                 /* we ignore mapcount for file pages */
4673                 return NULL;
4674         if (!get_page_unless_zero(page))
4675                 return NULL;
4676
4677         return page;
4678 }
4679
4680 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4681                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4682 {
4683         int usage_count;
4684         struct page *page = NULL;
4685         swp_entry_t ent = pte_to_swp_entry(ptent);
4686
4687         if (!move_anon() || non_swap_entry(ent))
4688                 return NULL;
4689         usage_count = mem_cgroup_count_swap_user(ent, &page);
4690         if (usage_count > 1) { /* we don't move shared anon */
4691                 if (page)
4692                         put_page(page);
4693                 return NULL;
4694         }
4695         if (do_swap_account)
4696                 entry->val = ent.val;
4697
4698         return page;
4699 }
4700
4701 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4702                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4703 {
4704         struct page *page = NULL;
4705         struct inode *inode;
4706         struct address_space *mapping;
4707         pgoff_t pgoff;
4708
4709         if (!vma->vm_file) /* anonymous vma */
4710                 return NULL;
4711         if (!move_file())
4712                 return NULL;
4713
4714         inode = vma->vm_file->f_path.dentry->d_inode;
4715         mapping = vma->vm_file->f_mapping;
4716         if (pte_none(ptent))
4717                 pgoff = linear_page_index(vma, addr);
4718         else /* pte_file(ptent) is true */
4719                 pgoff = pte_to_pgoff(ptent);
4720
4721         /* page is moved even if it's not RSS of this task(page-faulted). */
4722         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4723                 page = find_get_page(mapping, pgoff);
4724         } else { /* shmem/tmpfs file. we should take account of swap too. */
4725                 swp_entry_t ent;
4726                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4727                 if (do_swap_account)
4728                         entry->val = ent.val;
4729         }
4730
4731         return page;
4732 }
4733
4734 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4735                 unsigned long addr, pte_t ptent, union mc_target *target)
4736 {
4737         struct page *page = NULL;
4738         struct page_cgroup *pc;
4739         int ret = 0;
4740         swp_entry_t ent = { .val = 0 };
4741
4742         if (pte_present(ptent))
4743                 page = mc_handle_present_pte(vma, addr, ptent);
4744         else if (is_swap_pte(ptent))
4745                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4746         else if (pte_none(ptent) || pte_file(ptent))
4747                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4748
4749         if (!page && !ent.val)
4750                 return 0;
4751         if (page) {
4752                 pc = lookup_page_cgroup(page);
4753                 /*
4754                  * Do only loose check w/o page_cgroup lock.
4755                  * mem_cgroup_move_account() checks the pc is valid or not under
4756                  * the lock.
4757                  */
4758                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4759                         ret = MC_TARGET_PAGE;
4760                         if (target)
4761                                 target->page = page;
4762                 }
4763                 if (!ret || !target)
4764                         put_page(page);
4765         }
4766         /* There is a swap entry and a page doesn't exist or isn't charged */
4767         if (ent.val && !ret &&
4768                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4769                 ret = MC_TARGET_SWAP;
4770                 if (target)
4771                         target->ent = ent;
4772         }
4773         return ret;
4774 }
4775
4776 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4777                                         unsigned long addr, unsigned long end,
4778                                         struct mm_walk *walk)
4779 {
4780         struct vm_area_struct *vma = walk->private;
4781         pte_t *pte;
4782         spinlock_t *ptl;
4783
4784         split_huge_page_pmd(walk->mm, pmd);
4785
4786         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4787         for (; addr != end; pte++, addr += PAGE_SIZE)
4788                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4789                         mc.precharge++; /* increment precharge temporarily */
4790         pte_unmap_unlock(pte - 1, ptl);
4791         cond_resched();
4792
4793         return 0;
4794 }
4795
4796 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4797 {
4798         unsigned long precharge;
4799         struct vm_area_struct *vma;
4800
4801         down_read(&mm->mmap_sem);
4802         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4803                 struct mm_walk mem_cgroup_count_precharge_walk = {
4804                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4805                         .mm = mm,
4806                         .private = vma,
4807                 };
4808                 if (is_vm_hugetlb_page(vma))
4809                         continue;
4810                 walk_page_range(vma->vm_start, vma->vm_end,
4811                                         &mem_cgroup_count_precharge_walk);
4812         }
4813         up_read(&mm->mmap_sem);
4814
4815         precharge = mc.precharge;
4816         mc.precharge = 0;
4817
4818         return precharge;
4819 }
4820
4821 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4822 {
4823         unsigned long precharge = mem_cgroup_count_precharge(mm);
4824
4825         VM_BUG_ON(mc.moving_task);
4826         mc.moving_task = current;
4827         return mem_cgroup_do_precharge(precharge);
4828 }
4829
4830 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4831 static void __mem_cgroup_clear_mc(void)
4832 {
4833         struct mem_cgroup *from = mc.from;
4834         struct mem_cgroup *to = mc.to;
4835
4836         /* we must uncharge all the leftover precharges from mc.to */
4837         if (mc.precharge) {
4838                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4839                 mc.precharge = 0;
4840         }
4841         /*
4842          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4843          * we must uncharge here.
4844          */
4845         if (mc.moved_charge) {
4846                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4847                 mc.moved_charge = 0;
4848         }
4849         /* we must fixup refcnts and charges */
4850         if (mc.moved_swap) {
4851                 /* uncharge swap account from the old cgroup */
4852                 if (!mem_cgroup_is_root(mc.from))
4853                         res_counter_uncharge(&mc.from->memsw,
4854                                                 PAGE_SIZE * mc.moved_swap);
4855                 __mem_cgroup_put(mc.from, mc.moved_swap);
4856
4857                 if (!mem_cgroup_is_root(mc.to)) {
4858                         /*
4859                          * we charged both to->res and to->memsw, so we should
4860                          * uncharge to->res.
4861                          */
4862                         res_counter_uncharge(&mc.to->res,
4863                                                 PAGE_SIZE * mc.moved_swap);
4864                 }
4865                 /* we've already done mem_cgroup_get(mc.to) */
4866                 mc.moved_swap = 0;
4867         }
4868         memcg_oom_recover(from);
4869         memcg_oom_recover(to);
4870         wake_up_all(&mc.waitq);
4871 }
4872
4873 static void mem_cgroup_clear_mc(void)
4874 {
4875         struct mem_cgroup *from = mc.from;
4876
4877         /*
4878          * we must clear moving_task before waking up waiters at the end of
4879          * task migration.
4880          */
4881         mc.moving_task = NULL;
4882         __mem_cgroup_clear_mc();
4883         spin_lock(&mc.lock);
4884         mc.from = NULL;
4885         mc.to = NULL;
4886         spin_unlock(&mc.lock);
4887         mem_cgroup_end_move(from);
4888 }
4889
4890 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4891                                 struct cgroup *cgroup,
4892                                 struct task_struct *p,
4893                                 bool threadgroup)
4894 {
4895         int ret = 0;
4896         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4897
4898         if (mem->move_charge_at_immigrate) {
4899                 struct mm_struct *mm;
4900                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4901
4902                 VM_BUG_ON(from == mem);
4903
4904                 mm = get_task_mm(p);
4905                 if (!mm)
4906                         return 0;
4907                 /* We move charges only when we move a owner of the mm */
4908                 if (mm->owner == p) {
4909                         VM_BUG_ON(mc.from);
4910                         VM_BUG_ON(mc.to);
4911                         VM_BUG_ON(mc.precharge);
4912                         VM_BUG_ON(mc.moved_charge);
4913                         VM_BUG_ON(mc.moved_swap);
4914                         mem_cgroup_start_move(from);
4915                         spin_lock(&mc.lock);
4916                         mc.from = from;
4917                         mc.to = mem;
4918                         spin_unlock(&mc.lock);
4919                         /* We set mc.moving_task later */
4920
4921                         ret = mem_cgroup_precharge_mc(mm);
4922                         if (ret)
4923                                 mem_cgroup_clear_mc();
4924                 }
4925                 mmput(mm);
4926         }
4927         return ret;
4928 }
4929
4930 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4931                                 struct cgroup *cgroup,
4932                                 struct task_struct *p,
4933                                 bool threadgroup)
4934 {
4935         mem_cgroup_clear_mc();
4936 }
4937
4938 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4939                                 unsigned long addr, unsigned long end,
4940                                 struct mm_walk *walk)
4941 {
4942         int ret = 0;
4943         struct vm_area_struct *vma = walk->private;
4944         pte_t *pte;
4945         spinlock_t *ptl;
4946
4947         split_huge_page_pmd(walk->mm, pmd);
4948 retry:
4949         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4950         for (; addr != end; addr += PAGE_SIZE) {
4951                 pte_t ptent = *(pte++);
4952                 union mc_target target;
4953                 int type;
4954                 struct page *page;
4955                 struct page_cgroup *pc;
4956                 swp_entry_t ent;
4957
4958                 if (!mc.precharge)
4959                         break;
4960
4961                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4962                 switch (type) {
4963                 case MC_TARGET_PAGE:
4964                         page = target.page;
4965                         if (isolate_lru_page(page))
4966                                 goto put;
4967                         pc = lookup_page_cgroup(page);
4968                         if (!mem_cgroup_move_account(pc,
4969                                         mc.from, mc.to, false, PAGE_SIZE)) {
4970                                 mc.precharge--;
4971                                 /* we uncharge from mc.from later. */
4972                                 mc.moved_charge++;
4973                         }
4974                         putback_lru_page(page);
4975 put:                    /* is_target_pte_for_mc() gets the page */
4976                         put_page(page);
4977                         break;
4978                 case MC_TARGET_SWAP:
4979                         ent = target.ent;
4980                         if (!mem_cgroup_move_swap_account(ent,
4981                                                 mc.from, mc.to, false)) {
4982                                 mc.precharge--;
4983                                 /* we fixup refcnts and charges later. */
4984                                 mc.moved_swap++;
4985                         }
4986                         break;
4987                 default:
4988                         break;
4989                 }
4990         }
4991         pte_unmap_unlock(pte - 1, ptl);
4992         cond_resched();
4993
4994         if (addr != end) {
4995                 /*
4996                  * We have consumed all precharges we got in can_attach().
4997                  * We try charge one by one, but don't do any additional
4998                  * charges to mc.to if we have failed in charge once in attach()
4999                  * phase.
5000                  */
5001                 ret = mem_cgroup_do_precharge(1);
5002                 if (!ret)
5003                         goto retry;
5004         }
5005
5006         return ret;
5007 }
5008
5009 static void mem_cgroup_move_charge(struct mm_struct *mm)
5010 {
5011         struct vm_area_struct *vma;
5012
5013         lru_add_drain_all();
5014 retry:
5015         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5016                 /*
5017                  * Someone who are holding the mmap_sem might be waiting in
5018                  * waitq. So we cancel all extra charges, wake up all waiters,
5019                  * and retry. Because we cancel precharges, we might not be able
5020                  * to move enough charges, but moving charge is a best-effort
5021                  * feature anyway, so it wouldn't be a big problem.
5022                  */
5023                 __mem_cgroup_clear_mc();
5024                 cond_resched();
5025                 goto retry;
5026         }
5027         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5028                 int ret;
5029                 struct mm_walk mem_cgroup_move_charge_walk = {
5030                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5031                         .mm = mm,
5032                         .private = vma,
5033                 };
5034                 if (is_vm_hugetlb_page(vma))
5035                         continue;
5036                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5037                                                 &mem_cgroup_move_charge_walk);
5038                 if (ret)
5039                         /*
5040                          * means we have consumed all precharges and failed in
5041                          * doing additional charge. Just abandon here.
5042                          */
5043                         break;
5044         }
5045         up_read(&mm->mmap_sem);
5046 }
5047
5048 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5049                                 struct cgroup *cont,
5050                                 struct cgroup *old_cont,
5051                                 struct task_struct *p,
5052                                 bool threadgroup)
5053 {
5054         struct mm_struct *mm;
5055
5056         if (!mc.to)
5057                 /* no need to move charge */
5058                 return;
5059
5060         mm = get_task_mm(p);
5061         if (mm) {
5062                 mem_cgroup_move_charge(mm);
5063                 mmput(mm);
5064         }
5065         mem_cgroup_clear_mc();
5066 }
5067 #else   /* !CONFIG_MMU */
5068 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5069                                 struct cgroup *cgroup,
5070                                 struct task_struct *p,
5071                                 bool threadgroup)
5072 {
5073         return 0;
5074 }
5075 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5076                                 struct cgroup *cgroup,
5077                                 struct task_struct *p,
5078                                 bool threadgroup)
5079 {
5080 }
5081 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5082                                 struct cgroup *cont,
5083                                 struct cgroup *old_cont,
5084                                 struct task_struct *p,
5085                                 bool threadgroup)
5086 {
5087 }
5088 #endif
5089
5090 struct cgroup_subsys mem_cgroup_subsys = {
5091         .name = "memory",
5092         .subsys_id = mem_cgroup_subsys_id,
5093         .create = mem_cgroup_create,
5094         .pre_destroy = mem_cgroup_pre_destroy,
5095         .destroy = mem_cgroup_destroy,
5096         .populate = mem_cgroup_populate,
5097         .can_attach = mem_cgroup_can_attach,
5098         .cancel_attach = mem_cgroup_cancel_attach,
5099         .attach = mem_cgroup_move_task,
5100         .early_init = 0,
5101         .use_id = 1,
5102 };
5103
5104 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5105 static int __init enable_swap_account(char *s)
5106 {
5107         /* consider enabled if no parameter or 1 is given */
5108         if (!(*s) || !strcmp(s, "=1"))
5109                 really_do_swap_account = 1;
5110         else if (!strcmp(s, "=0"))
5111                 really_do_swap_account = 0;
5112         return 1;
5113 }
5114 __setup("swapaccount", enable_swap_account);
5115
5116 static int __init disable_swap_account(char *s)
5117 {
5118         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5119         enable_swap_account("=0");
5120         return 1;
5121 }
5122 __setup("noswapaccount", disable_swap_account);
5123 #endif