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