memcg: correctly order reading PCG_USED and pc->mem_cgroup
[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                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1836                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1837         } else
1838                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1839
1840         if (csize > PAGE_SIZE) /* change csize and retry */
1841                 return CHARGE_RETRY;
1842
1843         if (!(gfp_mask & __GFP_WAIT))
1844                 return CHARGE_WOULDBLOCK;
1845
1846         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1847                                         gfp_mask, flags);
1848         /*
1849          * try_to_free_mem_cgroup_pages() might not give us a full
1850          * picture of reclaim. Some pages are reclaimed and might be
1851          * moved to swap cache or just unmapped from the cgroup.
1852          * Check the limit again to see if the reclaim reduced the
1853          * current usage of the cgroup before giving up
1854          */
1855         if (ret || mem_cgroup_check_under_limit(mem_over_limit))
1856                 return CHARGE_RETRY;
1857
1858         /*
1859          * At task move, charge accounts can be doubly counted. So, it's
1860          * better to wait until the end of task_move if something is going on.
1861          */
1862         if (mem_cgroup_wait_acct_move(mem_over_limit))
1863                 return CHARGE_RETRY;
1864
1865         /* If we don't need to call oom-killer at el, return immediately */
1866         if (!oom_check)
1867                 return CHARGE_NOMEM;
1868         /* check OOM */
1869         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1870                 return CHARGE_OOM_DIE;
1871
1872         return CHARGE_RETRY;
1873 }
1874
1875 /*
1876  * Unlike exported interface, "oom" parameter is added. if oom==true,
1877  * oom-killer can be invoked.
1878  */
1879 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1880                                    gfp_t gfp_mask,
1881                                    struct mem_cgroup **memcg, bool oom,
1882                                    int page_size)
1883 {
1884         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1885         struct mem_cgroup *mem = NULL;
1886         int ret;
1887         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1888
1889         /*
1890          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1891          * in system level. So, allow to go ahead dying process in addition to
1892          * MEMDIE process.
1893          */
1894         if (unlikely(test_thread_flag(TIF_MEMDIE)
1895                      || fatal_signal_pending(current)))
1896                 goto bypass;
1897
1898         /*
1899          * We always charge the cgroup the mm_struct belongs to.
1900          * The mm_struct's mem_cgroup changes on task migration if the
1901          * thread group leader migrates. It's possible that mm is not
1902          * set, if so charge the init_mm (happens for pagecache usage).
1903          */
1904         if (!*memcg && !mm)
1905                 goto bypass;
1906 again:
1907         if (*memcg) { /* css should be a valid one */
1908                 mem = *memcg;
1909                 VM_BUG_ON(css_is_removed(&mem->css));
1910                 if (mem_cgroup_is_root(mem))
1911                         goto done;
1912                 if (page_size == PAGE_SIZE && consume_stock(mem))
1913                         goto done;
1914                 css_get(&mem->css);
1915         } else {
1916                 struct task_struct *p;
1917
1918                 rcu_read_lock();
1919                 p = rcu_dereference(mm->owner);
1920                 /*
1921                  * Because we don't have task_lock(), "p" can exit.
1922                  * In that case, "mem" can point to root or p can be NULL with
1923                  * race with swapoff. Then, we have small risk of mis-accouning.
1924                  * But such kind of mis-account by race always happens because
1925                  * we don't have cgroup_mutex(). It's overkill and we allo that
1926                  * small race, here.
1927                  * (*) swapoff at el will charge against mm-struct not against
1928                  * task-struct. So, mm->owner can be NULL.
1929                  */
1930                 mem = mem_cgroup_from_task(p);
1931                 if (!mem || mem_cgroup_is_root(mem)) {
1932                         rcu_read_unlock();
1933                         goto done;
1934                 }
1935                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1936                         /*
1937                          * It seems dagerous to access memcg without css_get().
1938                          * But considering how consume_stok works, it's not
1939                          * necessary. If consume_stock success, some charges
1940                          * from this memcg are cached on this cpu. So, we
1941                          * don't need to call css_get()/css_tryget() before
1942                          * calling consume_stock().
1943                          */
1944                         rcu_read_unlock();
1945                         goto done;
1946                 }
1947                 /* after here, we may be blocked. we need to get refcnt */
1948                 if (!css_tryget(&mem->css)) {
1949                         rcu_read_unlock();
1950                         goto again;
1951                 }
1952                 rcu_read_unlock();
1953         }
1954
1955         do {
1956                 bool oom_check;
1957
1958                 /* If killed, bypass charge */
1959                 if (fatal_signal_pending(current)) {
1960                         css_put(&mem->css);
1961                         goto bypass;
1962                 }
1963
1964                 oom_check = false;
1965                 if (oom && !nr_oom_retries) {
1966                         oom_check = true;
1967                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1968                 }
1969
1970                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1971
1972                 switch (ret) {
1973                 case CHARGE_OK:
1974                         break;
1975                 case CHARGE_RETRY: /* not in OOM situation but retry */
1976                         csize = page_size;
1977                         css_put(&mem->css);
1978                         mem = NULL;
1979                         goto again;
1980                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1981                         css_put(&mem->css);
1982                         goto nomem;
1983                 case CHARGE_NOMEM: /* OOM routine works */
1984                         if (!oom) {
1985                                 css_put(&mem->css);
1986                                 goto nomem;
1987                         }
1988                         /* If oom, we never return -ENOMEM */
1989                         nr_oom_retries--;
1990                         break;
1991                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
1992                         css_put(&mem->css);
1993                         goto bypass;
1994                 }
1995         } while (ret != CHARGE_OK);
1996
1997         if (csize > page_size)
1998                 refill_stock(mem, csize - page_size);
1999         css_put(&mem->css);
2000 done:
2001         *memcg = mem;
2002         return 0;
2003 nomem:
2004         *memcg = NULL;
2005         return -ENOMEM;
2006 bypass:
2007         *memcg = NULL;
2008         return 0;
2009 }
2010
2011 /*
2012  * Somemtimes we have to undo a charge we got by try_charge().
2013  * This function is for that and do uncharge, put css's refcnt.
2014  * gotten by try_charge().
2015  */
2016 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2017                                                         unsigned long count)
2018 {
2019         if (!mem_cgroup_is_root(mem)) {
2020                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2021                 if (do_swap_account)
2022                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2023         }
2024 }
2025
2026 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2027                                      int page_size)
2028 {
2029         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2030 }
2031
2032 /*
2033  * A helper function to get mem_cgroup from ID. must be called under
2034  * rcu_read_lock(). The caller must check css_is_removed() or some if
2035  * it's concern. (dropping refcnt from swap can be called against removed
2036  * memcg.)
2037  */
2038 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2039 {
2040         struct cgroup_subsys_state *css;
2041
2042         /* ID 0 is unused ID */
2043         if (!id)
2044                 return NULL;
2045         css = css_lookup(&mem_cgroup_subsys, id);
2046         if (!css)
2047                 return NULL;
2048         return container_of(css, struct mem_cgroup, css);
2049 }
2050
2051 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2052 {
2053         struct mem_cgroup *mem = NULL;
2054         struct page_cgroup *pc;
2055         unsigned short id;
2056         swp_entry_t ent;
2057
2058         VM_BUG_ON(!PageLocked(page));
2059
2060         pc = lookup_page_cgroup(page);
2061         lock_page_cgroup(pc);
2062         if (PageCgroupUsed(pc)) {
2063                 mem = pc->mem_cgroup;
2064                 if (mem && !css_tryget(&mem->css))
2065                         mem = NULL;
2066         } else if (PageSwapCache(page)) {
2067                 ent.val = page_private(page);
2068                 id = lookup_swap_cgroup(ent);
2069                 rcu_read_lock();
2070                 mem = mem_cgroup_lookup(id);
2071                 if (mem && !css_tryget(&mem->css))
2072                         mem = NULL;
2073                 rcu_read_unlock();
2074         }
2075         unlock_page_cgroup(pc);
2076         return mem;
2077 }
2078
2079 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2080                                        struct page_cgroup *pc,
2081                                        enum charge_type ctype,
2082                                        int page_size)
2083 {
2084         int nr_pages = page_size >> PAGE_SHIFT;
2085
2086         /* try_charge() can return NULL to *memcg, taking care of it. */
2087         if (!mem)
2088                 return;
2089
2090         lock_page_cgroup(pc);
2091         if (unlikely(PageCgroupUsed(pc))) {
2092                 unlock_page_cgroup(pc);
2093                 mem_cgroup_cancel_charge(mem, page_size);
2094                 return;
2095         }
2096         /*
2097          * we don't need page_cgroup_lock about tail pages, becase they are not
2098          * accessed by any other context at this point.
2099          */
2100         pc->mem_cgroup = mem;
2101         /*
2102          * We access a page_cgroup asynchronously without lock_page_cgroup().
2103          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2104          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2105          * before USED bit, we need memory barrier here.
2106          * See mem_cgroup_add_lru_list(), etc.
2107          */
2108         smp_wmb();
2109         switch (ctype) {
2110         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2111         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2112                 SetPageCgroupCache(pc);
2113                 SetPageCgroupUsed(pc);
2114                 break;
2115         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2116                 ClearPageCgroupCache(pc);
2117                 SetPageCgroupUsed(pc);
2118                 break;
2119         default:
2120                 break;
2121         }
2122
2123         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2124         unlock_page_cgroup(pc);
2125         /*
2126          * "charge_statistics" updated event counter. Then, check it.
2127          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2128          * if they exceeds softlimit.
2129          */
2130         memcg_check_events(mem, pc->page);
2131 }
2132
2133 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2134
2135 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2136                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2137 /*
2138  * Because tail pages are not marked as "used", set it. We're under
2139  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2140  */
2141 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2142 {
2143         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2144         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2145         unsigned long flags;
2146
2147         /*
2148          * We have no races with charge/uncharge but will have races with
2149          * page state accounting.
2150          */
2151         move_lock_page_cgroup(head_pc, &flags);
2152
2153         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2154         smp_wmb(); /* see __commit_charge() */
2155         if (PageCgroupAcctLRU(head_pc)) {
2156                 enum lru_list lru;
2157                 struct mem_cgroup_per_zone *mz;
2158
2159                 /*
2160                  * LRU flags cannot be copied because we need to add tail
2161                  *.page to LRU by generic call and our hook will be called.
2162                  * We hold lru_lock, then, reduce counter directly.
2163                  */
2164                 lru = page_lru(head);
2165                 mz = page_cgroup_zoneinfo(head_pc);
2166                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2167         }
2168         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2169         move_unlock_page_cgroup(head_pc, &flags);
2170 }
2171 #endif
2172
2173 /**
2174  * __mem_cgroup_move_account - move account of the page
2175  * @pc: page_cgroup of the page.
2176  * @from: mem_cgroup which the page is moved from.
2177  * @to: mem_cgroup which the page is moved to. @from != @to.
2178  * @uncharge: whether we should call uncharge and css_put against @from.
2179  *
2180  * The caller must confirm following.
2181  * - page is not on LRU (isolate_page() is useful.)
2182  * - the pc is locked, used, and ->mem_cgroup points to @from.
2183  *
2184  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2185  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2186  * true, this function does "uncharge" from old cgroup, but it doesn't if
2187  * @uncharge is false, so a caller should do "uncharge".
2188  */
2189
2190 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2191         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge,
2192         int charge_size)
2193 {
2194         int nr_pages = charge_size >> PAGE_SHIFT;
2195
2196         VM_BUG_ON(from == to);
2197         VM_BUG_ON(PageLRU(pc->page));
2198         VM_BUG_ON(!page_is_cgroup_locked(pc));
2199         VM_BUG_ON(!PageCgroupUsed(pc));
2200         VM_BUG_ON(pc->mem_cgroup != from);
2201
2202         if (PageCgroupFileMapped(pc)) {
2203                 /* Update mapped_file data for mem_cgroup */
2204                 preempt_disable();
2205                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2206                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2207                 preempt_enable();
2208         }
2209         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2210         if (uncharge)
2211                 /* This is not "cancel", but cancel_charge does all we need. */
2212                 mem_cgroup_cancel_charge(from, charge_size);
2213
2214         /* caller should have done css_get */
2215         pc->mem_cgroup = to;
2216         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2217         /*
2218          * We charges against "to" which may not have any tasks. Then, "to"
2219          * can be under rmdir(). But in current implementation, caller of
2220          * this function is just force_empty() and move charge, so it's
2221          * garanteed that "to" is never removed. So, we don't check rmdir
2222          * status here.
2223          */
2224 }
2225
2226 /*
2227  * check whether the @pc is valid for moving account and call
2228  * __mem_cgroup_move_account()
2229  */
2230 static int mem_cgroup_move_account(struct page_cgroup *pc,
2231                 struct mem_cgroup *from, struct mem_cgroup *to,
2232                 bool uncharge, int charge_size)
2233 {
2234         int ret = -EINVAL;
2235         unsigned long flags;
2236
2237         if ((charge_size > PAGE_SIZE) && !PageTransHuge(pc->page))
2238                 return -EBUSY;
2239
2240         lock_page_cgroup(pc);
2241         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2242                 move_lock_page_cgroup(pc, &flags);
2243                 __mem_cgroup_move_account(pc, from, to, uncharge, charge_size);
2244                 move_unlock_page_cgroup(pc, &flags);
2245                 ret = 0;
2246         }
2247         unlock_page_cgroup(pc);
2248         /*
2249          * check events
2250          */
2251         memcg_check_events(to, pc->page);
2252         memcg_check_events(from, pc->page);
2253         return ret;
2254 }
2255
2256 /*
2257  * move charges to its parent.
2258  */
2259
2260 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2261                                   struct mem_cgroup *child,
2262                                   gfp_t gfp_mask)
2263 {
2264         struct page *page = pc->page;
2265         struct cgroup *cg = child->css.cgroup;
2266         struct cgroup *pcg = cg->parent;
2267         struct mem_cgroup *parent;
2268         int charge = PAGE_SIZE;
2269         unsigned long flags;
2270         int ret;
2271
2272         /* Is ROOT ? */
2273         if (!pcg)
2274                 return -EINVAL;
2275
2276         ret = -EBUSY;
2277         if (!get_page_unless_zero(page))
2278                 goto out;
2279         if (isolate_lru_page(page))
2280                 goto put;
2281         /* The page is isolated from LRU and we have no race with splitting */
2282         charge = PAGE_SIZE << compound_order(page);
2283
2284         parent = mem_cgroup_from_cont(pcg);
2285         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, charge);
2286         if (ret || !parent)
2287                 goto put_back;
2288
2289         if (charge > PAGE_SIZE)
2290                 flags = compound_lock_irqsave(page);
2291
2292         ret = mem_cgroup_move_account(pc, child, parent, true, charge);
2293         if (ret)
2294                 mem_cgroup_cancel_charge(parent, charge);
2295 put_back:
2296         if (charge > PAGE_SIZE)
2297                 compound_unlock_irqrestore(page, flags);
2298         putback_lru_page(page);
2299 put:
2300         put_page(page);
2301 out:
2302         return ret;
2303 }
2304
2305 /*
2306  * Charge the memory controller for page usage.
2307  * Return
2308  * 0 if the charge was successful
2309  * < 0 if the cgroup is over its limit
2310  */
2311 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2312                                 gfp_t gfp_mask, enum charge_type ctype)
2313 {
2314         struct mem_cgroup *mem = NULL;
2315         struct page_cgroup *pc;
2316         int ret;
2317         int page_size = PAGE_SIZE;
2318
2319         if (PageTransHuge(page)) {
2320                 page_size <<= compound_order(page);
2321                 VM_BUG_ON(!PageTransHuge(page));
2322         }
2323
2324         pc = lookup_page_cgroup(page);
2325         /* can happen at boot */
2326         if (unlikely(!pc))
2327                 return 0;
2328         prefetchw(pc);
2329
2330         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page_size);
2331         if (ret || !mem)
2332                 return ret;
2333
2334         __mem_cgroup_commit_charge(mem, pc, ctype, page_size);
2335         return 0;
2336 }
2337
2338 int mem_cgroup_newpage_charge(struct page *page,
2339                               struct mm_struct *mm, gfp_t gfp_mask)
2340 {
2341         if (mem_cgroup_disabled())
2342                 return 0;
2343         /*
2344          * If already mapped, we don't have to account.
2345          * If page cache, page->mapping has address_space.
2346          * But page->mapping may have out-of-use anon_vma pointer,
2347          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2348          * is NULL.
2349          */
2350         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2351                 return 0;
2352         if (unlikely(!mm))
2353                 mm = &init_mm;
2354         return mem_cgroup_charge_common(page, mm, gfp_mask,
2355                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2356 }
2357
2358 static void
2359 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2360                                         enum charge_type ctype);
2361
2362 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2363                                 gfp_t gfp_mask)
2364 {
2365         int ret;
2366
2367         if (mem_cgroup_disabled())
2368                 return 0;
2369         if (PageCompound(page))
2370                 return 0;
2371         /*
2372          * Corner case handling. This is called from add_to_page_cache()
2373          * in usual. But some FS (shmem) precharges this page before calling it
2374          * and call add_to_page_cache() with GFP_NOWAIT.
2375          *
2376          * For GFP_NOWAIT case, the page may be pre-charged before calling
2377          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2378          * charge twice. (It works but has to pay a bit larger cost.)
2379          * And when the page is SwapCache, it should take swap information
2380          * into account. This is under lock_page() now.
2381          */
2382         if (!(gfp_mask & __GFP_WAIT)) {
2383                 struct page_cgroup *pc;
2384
2385                 pc = lookup_page_cgroup(page);
2386                 if (!pc)
2387                         return 0;
2388                 lock_page_cgroup(pc);
2389                 if (PageCgroupUsed(pc)) {
2390                         unlock_page_cgroup(pc);
2391                         return 0;
2392                 }
2393                 unlock_page_cgroup(pc);
2394         }
2395
2396         if (unlikely(!mm))
2397                 mm = &init_mm;
2398
2399         if (page_is_file_cache(page))
2400                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2401                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2402
2403         /* shmem */
2404         if (PageSwapCache(page)) {
2405                 struct mem_cgroup *mem = NULL;
2406
2407                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2408                 if (!ret)
2409                         __mem_cgroup_commit_charge_swapin(page, mem,
2410                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2411         } else
2412                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2413                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2414
2415         return ret;
2416 }
2417
2418 /*
2419  * While swap-in, try_charge -> commit or cancel, the page is locked.
2420  * And when try_charge() successfully returns, one refcnt to memcg without
2421  * struct page_cgroup is acquired. This refcnt will be consumed by
2422  * "commit()" or removed by "cancel()"
2423  */
2424 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2425                                  struct page *page,
2426                                  gfp_t mask, struct mem_cgroup **ptr)
2427 {
2428         struct mem_cgroup *mem;
2429         int ret;
2430
2431         if (mem_cgroup_disabled())
2432                 return 0;
2433
2434         if (!do_swap_account)
2435                 goto charge_cur_mm;
2436         /*
2437          * A racing thread's fault, or swapoff, may have already updated
2438          * the pte, and even removed page from swap cache: in those cases
2439          * do_swap_page()'s pte_same() test will fail; but there's also a
2440          * KSM case which does need to charge the page.
2441          */
2442         if (!PageSwapCache(page))
2443                 goto charge_cur_mm;
2444         mem = try_get_mem_cgroup_from_page(page);
2445         if (!mem)
2446                 goto charge_cur_mm;
2447         *ptr = mem;
2448         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2449         css_put(&mem->css);
2450         return ret;
2451 charge_cur_mm:
2452         if (unlikely(!mm))
2453                 mm = &init_mm;
2454         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2455 }
2456
2457 static void
2458 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2459                                         enum charge_type ctype)
2460 {
2461         struct page_cgroup *pc;
2462
2463         if (mem_cgroup_disabled())
2464                 return;
2465         if (!ptr)
2466                 return;
2467         cgroup_exclude_rmdir(&ptr->css);
2468         pc = lookup_page_cgroup(page);
2469         mem_cgroup_lru_del_before_commit_swapcache(page);
2470         __mem_cgroup_commit_charge(ptr, pc, ctype, PAGE_SIZE);
2471         mem_cgroup_lru_add_after_commit_swapcache(page);
2472         /*
2473          * Now swap is on-memory. This means this page may be
2474          * counted both as mem and swap....double count.
2475          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2476          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2477          * may call delete_from_swap_cache() before reach here.
2478          */
2479         if (do_swap_account && PageSwapCache(page)) {
2480                 swp_entry_t ent = {.val = page_private(page)};
2481                 unsigned short id;
2482                 struct mem_cgroup *memcg;
2483
2484                 id = swap_cgroup_record(ent, 0);
2485                 rcu_read_lock();
2486                 memcg = mem_cgroup_lookup(id);
2487                 if (memcg) {
2488                         /*
2489                          * This recorded memcg can be obsolete one. So, avoid
2490                          * calling css_tryget
2491                          */
2492                         if (!mem_cgroup_is_root(memcg))
2493                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2494                         mem_cgroup_swap_statistics(memcg, false);
2495                         mem_cgroup_put(memcg);
2496                 }
2497                 rcu_read_unlock();
2498         }
2499         /*
2500          * At swapin, we may charge account against cgroup which has no tasks.
2501          * So, rmdir()->pre_destroy() can be called while we do this charge.
2502          * In that case, we need to call pre_destroy() again. check it here.
2503          */
2504         cgroup_release_and_wakeup_rmdir(&ptr->css);
2505 }
2506
2507 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2508 {
2509         __mem_cgroup_commit_charge_swapin(page, ptr,
2510                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2511 }
2512
2513 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2514 {
2515         if (mem_cgroup_disabled())
2516                 return;
2517         if (!mem)
2518                 return;
2519         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2520 }
2521
2522 static void
2523 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2524               int page_size)
2525 {
2526         struct memcg_batch_info *batch = NULL;
2527         bool uncharge_memsw = true;
2528         /* If swapout, usage of swap doesn't decrease */
2529         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2530                 uncharge_memsw = false;
2531
2532         batch = &current->memcg_batch;
2533         /*
2534          * In usual, we do css_get() when we remember memcg pointer.
2535          * But in this case, we keep res->usage until end of a series of
2536          * uncharges. Then, it's ok to ignore memcg's refcnt.
2537          */
2538         if (!batch->memcg)
2539                 batch->memcg = mem;
2540         /*
2541          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2542          * In those cases, all pages freed continously can be expected to be in
2543          * the same cgroup and we have chance to coalesce uncharges.
2544          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2545          * because we want to do uncharge as soon as possible.
2546          */
2547
2548         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2549                 goto direct_uncharge;
2550
2551         if (page_size != PAGE_SIZE)
2552                 goto direct_uncharge;
2553
2554         /*
2555          * In typical case, batch->memcg == mem. This means we can
2556          * merge a series of uncharges to an uncharge of res_counter.
2557          * If not, we uncharge res_counter ony by one.
2558          */
2559         if (batch->memcg != mem)
2560                 goto direct_uncharge;
2561         /* remember freed charge and uncharge it later */
2562         batch->bytes += PAGE_SIZE;
2563         if (uncharge_memsw)
2564                 batch->memsw_bytes += PAGE_SIZE;
2565         return;
2566 direct_uncharge:
2567         res_counter_uncharge(&mem->res, page_size);
2568         if (uncharge_memsw)
2569                 res_counter_uncharge(&mem->memsw, page_size);
2570         if (unlikely(batch->memcg != mem))
2571                 memcg_oom_recover(mem);
2572         return;
2573 }
2574
2575 /*
2576  * uncharge if !page_mapped(page)
2577  */
2578 static struct mem_cgroup *
2579 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2580 {
2581         int count;
2582         struct page_cgroup *pc;
2583         struct mem_cgroup *mem = NULL;
2584         int page_size = PAGE_SIZE;
2585
2586         if (mem_cgroup_disabled())
2587                 return NULL;
2588
2589         if (PageSwapCache(page))
2590                 return NULL;
2591
2592         if (PageTransHuge(page)) {
2593                 page_size <<= compound_order(page);
2594                 VM_BUG_ON(!PageTransHuge(page));
2595         }
2596
2597         count = page_size >> PAGE_SHIFT;
2598         /*
2599          * Check if our page_cgroup is valid
2600          */
2601         pc = lookup_page_cgroup(page);
2602         if (unlikely(!pc || !PageCgroupUsed(pc)))
2603                 return NULL;
2604
2605         lock_page_cgroup(pc);
2606
2607         mem = pc->mem_cgroup;
2608
2609         if (!PageCgroupUsed(pc))
2610                 goto unlock_out;
2611
2612         switch (ctype) {
2613         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2614         case MEM_CGROUP_CHARGE_TYPE_DROP:
2615                 /* See mem_cgroup_prepare_migration() */
2616                 if (page_mapped(page) || PageCgroupMigration(pc))
2617                         goto unlock_out;
2618                 break;
2619         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2620                 if (!PageAnon(page)) {  /* Shared memory */
2621                         if (page->mapping && !page_is_file_cache(page))
2622                                 goto unlock_out;
2623                 } else if (page_mapped(page)) /* Anon */
2624                                 goto unlock_out;
2625                 break;
2626         default:
2627                 break;
2628         }
2629
2630         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2631
2632         ClearPageCgroupUsed(pc);
2633         /*
2634          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2635          * freed from LRU. This is safe because uncharged page is expected not
2636          * to be reused (freed soon). Exception is SwapCache, it's handled by
2637          * special functions.
2638          */
2639
2640         unlock_page_cgroup(pc);
2641         /*
2642          * even after unlock, we have mem->res.usage here and this memcg
2643          * will never be freed.
2644          */
2645         memcg_check_events(mem, page);
2646         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2647                 mem_cgroup_swap_statistics(mem, true);
2648                 mem_cgroup_get(mem);
2649         }
2650         if (!mem_cgroup_is_root(mem))
2651                 __do_uncharge(mem, ctype, page_size);
2652
2653         return mem;
2654
2655 unlock_out:
2656         unlock_page_cgroup(pc);
2657         return NULL;
2658 }
2659
2660 void mem_cgroup_uncharge_page(struct page *page)
2661 {
2662         /* early check. */
2663         if (page_mapped(page))
2664                 return;
2665         if (page->mapping && !PageAnon(page))
2666                 return;
2667         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2668 }
2669
2670 void mem_cgroup_uncharge_cache_page(struct page *page)
2671 {
2672         VM_BUG_ON(page_mapped(page));
2673         VM_BUG_ON(page->mapping);
2674         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2675 }
2676
2677 /*
2678  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2679  * In that cases, pages are freed continuously and we can expect pages
2680  * are in the same memcg. All these calls itself limits the number of
2681  * pages freed at once, then uncharge_start/end() is called properly.
2682  * This may be called prural(2) times in a context,
2683  */
2684
2685 void mem_cgroup_uncharge_start(void)
2686 {
2687         current->memcg_batch.do_batch++;
2688         /* We can do nest. */
2689         if (current->memcg_batch.do_batch == 1) {
2690                 current->memcg_batch.memcg = NULL;
2691                 current->memcg_batch.bytes = 0;
2692                 current->memcg_batch.memsw_bytes = 0;
2693         }
2694 }
2695
2696 void mem_cgroup_uncharge_end(void)
2697 {
2698         struct memcg_batch_info *batch = &current->memcg_batch;
2699
2700         if (!batch->do_batch)
2701                 return;
2702
2703         batch->do_batch--;
2704         if (batch->do_batch) /* If stacked, do nothing. */
2705                 return;
2706
2707         if (!batch->memcg)
2708                 return;
2709         /*
2710          * This "batch->memcg" is valid without any css_get/put etc...
2711          * bacause we hide charges behind us.
2712          */
2713         if (batch->bytes)
2714                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2715         if (batch->memsw_bytes)
2716                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2717         memcg_oom_recover(batch->memcg);
2718         /* forget this pointer (for sanity check) */
2719         batch->memcg = NULL;
2720 }
2721
2722 #ifdef CONFIG_SWAP
2723 /*
2724  * called after __delete_from_swap_cache() and drop "page" account.
2725  * memcg information is recorded to swap_cgroup of "ent"
2726  */
2727 void
2728 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2729 {
2730         struct mem_cgroup *memcg;
2731         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2732
2733         if (!swapout) /* this was a swap cache but the swap is unused ! */
2734                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2735
2736         memcg = __mem_cgroup_uncharge_common(page, ctype);
2737
2738         /*
2739          * record memcg information,  if swapout && memcg != NULL,
2740          * mem_cgroup_get() was called in uncharge().
2741          */
2742         if (do_swap_account && swapout && memcg)
2743                 swap_cgroup_record(ent, css_id(&memcg->css));
2744 }
2745 #endif
2746
2747 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2748 /*
2749  * called from swap_entry_free(). remove record in swap_cgroup and
2750  * uncharge "memsw" account.
2751  */
2752 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2753 {
2754         struct mem_cgroup *memcg;
2755         unsigned short id;
2756
2757         if (!do_swap_account)
2758                 return;
2759
2760         id = swap_cgroup_record(ent, 0);
2761         rcu_read_lock();
2762         memcg = mem_cgroup_lookup(id);
2763         if (memcg) {
2764                 /*
2765                  * We uncharge this because swap is freed.
2766                  * This memcg can be obsolete one. We avoid calling css_tryget
2767                  */
2768                 if (!mem_cgroup_is_root(memcg))
2769                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2770                 mem_cgroup_swap_statistics(memcg, false);
2771                 mem_cgroup_put(memcg);
2772         }
2773         rcu_read_unlock();
2774 }
2775
2776 /**
2777  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2778  * @entry: swap entry to be moved
2779  * @from:  mem_cgroup which the entry is moved from
2780  * @to:  mem_cgroup which the entry is moved to
2781  * @need_fixup: whether we should fixup res_counters and refcounts.
2782  *
2783  * It succeeds only when the swap_cgroup's record for this entry is the same
2784  * as the mem_cgroup's id of @from.
2785  *
2786  * Returns 0 on success, -EINVAL on failure.
2787  *
2788  * The caller must have charged to @to, IOW, called res_counter_charge() about
2789  * both res and memsw, and called css_get().
2790  */
2791 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2792                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2793 {
2794         unsigned short old_id, new_id;
2795
2796         old_id = css_id(&from->css);
2797         new_id = css_id(&to->css);
2798
2799         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2800                 mem_cgroup_swap_statistics(from, false);
2801                 mem_cgroup_swap_statistics(to, true);
2802                 /*
2803                  * This function is only called from task migration context now.
2804                  * It postpones res_counter and refcount handling till the end
2805                  * of task migration(mem_cgroup_clear_mc()) for performance
2806                  * improvement. But we cannot postpone mem_cgroup_get(to)
2807                  * because if the process that has been moved to @to does
2808                  * swap-in, the refcount of @to might be decreased to 0.
2809                  */
2810                 mem_cgroup_get(to);
2811                 if (need_fixup) {
2812                         if (!mem_cgroup_is_root(from))
2813                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2814                         mem_cgroup_put(from);
2815                         /*
2816                          * we charged both to->res and to->memsw, so we should
2817                          * uncharge to->res.
2818                          */
2819                         if (!mem_cgroup_is_root(to))
2820                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2821                 }
2822                 return 0;
2823         }
2824         return -EINVAL;
2825 }
2826 #else
2827 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2828                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2829 {
2830         return -EINVAL;
2831 }
2832 #endif
2833
2834 /*
2835  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2836  * page belongs to.
2837  */
2838 int mem_cgroup_prepare_migration(struct page *page,
2839         struct page *newpage, struct mem_cgroup **ptr)
2840 {
2841         struct page_cgroup *pc;
2842         struct mem_cgroup *mem = NULL;
2843         enum charge_type ctype;
2844         int ret = 0;
2845
2846         VM_BUG_ON(PageTransHuge(page));
2847         if (mem_cgroup_disabled())
2848                 return 0;
2849
2850         pc = lookup_page_cgroup(page);
2851         lock_page_cgroup(pc);
2852         if (PageCgroupUsed(pc)) {
2853                 mem = pc->mem_cgroup;
2854                 css_get(&mem->css);
2855                 /*
2856                  * At migrating an anonymous page, its mapcount goes down
2857                  * to 0 and uncharge() will be called. But, even if it's fully
2858                  * unmapped, migration may fail and this page has to be
2859                  * charged again. We set MIGRATION flag here and delay uncharge
2860                  * until end_migration() is called
2861                  *
2862                  * Corner Case Thinking
2863                  * A)
2864                  * When the old page was mapped as Anon and it's unmap-and-freed
2865                  * while migration was ongoing.
2866                  * If unmap finds the old page, uncharge() of it will be delayed
2867                  * until end_migration(). If unmap finds a new page, it's
2868                  * uncharged when it make mapcount to be 1->0. If unmap code
2869                  * finds swap_migration_entry, the new page will not be mapped
2870                  * and end_migration() will find it(mapcount==0).
2871                  *
2872                  * B)
2873                  * When the old page was mapped but migraion fails, the kernel
2874                  * remaps it. A charge for it is kept by MIGRATION flag even
2875                  * if mapcount goes down to 0. We can do remap successfully
2876                  * without charging it again.
2877                  *
2878                  * C)
2879                  * The "old" page is under lock_page() until the end of
2880                  * migration, so, the old page itself will not be swapped-out.
2881                  * If the new page is swapped out before end_migraton, our
2882                  * hook to usual swap-out path will catch the event.
2883                  */
2884                 if (PageAnon(page))
2885                         SetPageCgroupMigration(pc);
2886         }
2887         unlock_page_cgroup(pc);
2888         /*
2889          * If the page is not charged at this point,
2890          * we return here.
2891          */
2892         if (!mem)
2893                 return 0;
2894
2895         *ptr = mem;
2896         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false, PAGE_SIZE);
2897         css_put(&mem->css);/* drop extra refcnt */
2898         if (ret || *ptr == NULL) {
2899                 if (PageAnon(page)) {
2900                         lock_page_cgroup(pc);
2901                         ClearPageCgroupMigration(pc);
2902                         unlock_page_cgroup(pc);
2903                         /*
2904                          * The old page may be fully unmapped while we kept it.
2905                          */
2906                         mem_cgroup_uncharge_page(page);
2907                 }
2908                 return -ENOMEM;
2909         }
2910         /*
2911          * We charge new page before it's used/mapped. So, even if unlock_page()
2912          * is called before end_migration, we can catch all events on this new
2913          * page. In the case new page is migrated but not remapped, new page's
2914          * mapcount will be finally 0 and we call uncharge in end_migration().
2915          */
2916         pc = lookup_page_cgroup(newpage);
2917         if (PageAnon(page))
2918                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2919         else if (page_is_file_cache(page))
2920                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2921         else
2922                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2923         __mem_cgroup_commit_charge(mem, pc, ctype, PAGE_SIZE);
2924         return ret;
2925 }
2926
2927 /* remove redundant charge if migration failed*/
2928 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2929         struct page *oldpage, struct page *newpage, bool migration_ok)
2930 {
2931         struct page *used, *unused;
2932         struct page_cgroup *pc;
2933
2934         if (!mem)
2935                 return;
2936         /* blocks rmdir() */
2937         cgroup_exclude_rmdir(&mem->css);
2938         if (!migration_ok) {
2939                 used = oldpage;
2940                 unused = newpage;
2941         } else {
2942                 used = newpage;
2943                 unused = oldpage;
2944         }
2945         /*
2946          * We disallowed uncharge of pages under migration because mapcount
2947          * of the page goes down to zero, temporarly.
2948          * Clear the flag and check the page should be charged.
2949          */
2950         pc = lookup_page_cgroup(oldpage);
2951         lock_page_cgroup(pc);
2952         ClearPageCgroupMigration(pc);
2953         unlock_page_cgroup(pc);
2954
2955         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2956
2957         /*
2958          * If a page is a file cache, radix-tree replacement is very atomic
2959          * and we can skip this check. When it was an Anon page, its mapcount
2960          * goes down to 0. But because we added MIGRATION flage, it's not
2961          * uncharged yet. There are several case but page->mapcount check
2962          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2963          * check. (see prepare_charge() also)
2964          */
2965         if (PageAnon(used))
2966                 mem_cgroup_uncharge_page(used);
2967         /*
2968          * At migration, we may charge account against cgroup which has no
2969          * tasks.
2970          * So, rmdir()->pre_destroy() can be called while we do this charge.
2971          * In that case, we need to call pre_destroy() again. check it here.
2972          */
2973         cgroup_release_and_wakeup_rmdir(&mem->css);
2974 }
2975
2976 /*
2977  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2978  * Calling hierarchical_reclaim is not enough because we should update
2979  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2980  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2981  * not from the memcg which this page would be charged to.
2982  * try_charge_swapin does all of these works properly.
2983  */
2984 int mem_cgroup_shmem_charge_fallback(struct page *page,
2985                             struct mm_struct *mm,
2986                             gfp_t gfp_mask)
2987 {
2988         struct mem_cgroup *mem = NULL;
2989         int ret;
2990
2991         if (mem_cgroup_disabled())
2992                 return 0;
2993
2994         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2995         if (!ret)
2996                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2997
2998         return ret;
2999 }
3000
3001 static DEFINE_MUTEX(set_limit_mutex);
3002
3003 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3004                                 unsigned long long val)
3005 {
3006         int retry_count;
3007         u64 memswlimit, memlimit;
3008         int ret = 0;
3009         int children = mem_cgroup_count_children(memcg);
3010         u64 curusage, oldusage;
3011         int enlarge;
3012
3013         /*
3014          * For keeping hierarchical_reclaim simple, how long we should retry
3015          * is depends on callers. We set our retry-count to be function
3016          * of # of children which we should visit in this loop.
3017          */
3018         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3019
3020         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3021
3022         enlarge = 0;
3023         while (retry_count) {
3024                 if (signal_pending(current)) {
3025                         ret = -EINTR;
3026                         break;
3027                 }
3028                 /*
3029                  * Rather than hide all in some function, I do this in
3030                  * open coded manner. You see what this really does.
3031                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3032                  */
3033                 mutex_lock(&set_limit_mutex);
3034                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3035                 if (memswlimit < val) {
3036                         ret = -EINVAL;
3037                         mutex_unlock(&set_limit_mutex);
3038                         break;
3039                 }
3040
3041                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3042                 if (memlimit < val)
3043                         enlarge = 1;
3044
3045                 ret = res_counter_set_limit(&memcg->res, val);
3046                 if (!ret) {
3047                         if (memswlimit == val)
3048                                 memcg->memsw_is_minimum = true;
3049                         else
3050                                 memcg->memsw_is_minimum = false;
3051                 }
3052                 mutex_unlock(&set_limit_mutex);
3053
3054                 if (!ret)
3055                         break;
3056
3057                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3058                                                 MEM_CGROUP_RECLAIM_SHRINK);
3059                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3060                 /* Usage is reduced ? */
3061                 if (curusage >= oldusage)
3062                         retry_count--;
3063                 else
3064                         oldusage = curusage;
3065         }
3066         if (!ret && enlarge)
3067                 memcg_oom_recover(memcg);
3068
3069         return ret;
3070 }
3071
3072 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3073                                         unsigned long long val)
3074 {
3075         int retry_count;
3076         u64 memlimit, memswlimit, oldusage, curusage;
3077         int children = mem_cgroup_count_children(memcg);
3078         int ret = -EBUSY;
3079         int enlarge = 0;
3080
3081         /* see mem_cgroup_resize_res_limit */
3082         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3083         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3084         while (retry_count) {
3085                 if (signal_pending(current)) {
3086                         ret = -EINTR;
3087                         break;
3088                 }
3089                 /*
3090                  * Rather than hide all in some function, I do this in
3091                  * open coded manner. You see what this really does.
3092                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3093                  */
3094                 mutex_lock(&set_limit_mutex);
3095                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3096                 if (memlimit > val) {
3097                         ret = -EINVAL;
3098                         mutex_unlock(&set_limit_mutex);
3099                         break;
3100                 }
3101                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3102                 if (memswlimit < val)
3103                         enlarge = 1;
3104                 ret = res_counter_set_limit(&memcg->memsw, val);
3105                 if (!ret) {
3106                         if (memlimit == val)
3107                                 memcg->memsw_is_minimum = true;
3108                         else
3109                                 memcg->memsw_is_minimum = false;
3110                 }
3111                 mutex_unlock(&set_limit_mutex);
3112
3113                 if (!ret)
3114                         break;
3115
3116                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3117                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3118                                                 MEM_CGROUP_RECLAIM_SHRINK);
3119                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3120                 /* Usage is reduced ? */
3121                 if (curusage >= oldusage)
3122                         retry_count--;
3123                 else
3124                         oldusage = curusage;
3125         }
3126         if (!ret && enlarge)
3127                 memcg_oom_recover(memcg);
3128         return ret;
3129 }
3130
3131 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3132                                             gfp_t gfp_mask)
3133 {
3134         unsigned long nr_reclaimed = 0;
3135         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3136         unsigned long reclaimed;
3137         int loop = 0;
3138         struct mem_cgroup_tree_per_zone *mctz;
3139         unsigned long long excess;
3140
3141         if (order > 0)
3142                 return 0;
3143
3144         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3145         /*
3146          * This loop can run a while, specially if mem_cgroup's continuously
3147          * keep exceeding their soft limit and putting the system under
3148          * pressure
3149          */
3150         do {
3151                 if (next_mz)
3152                         mz = next_mz;
3153                 else
3154                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3155                 if (!mz)
3156                         break;
3157
3158                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3159                                                 gfp_mask,
3160                                                 MEM_CGROUP_RECLAIM_SOFT);
3161                 nr_reclaimed += reclaimed;
3162                 spin_lock(&mctz->lock);
3163
3164                 /*
3165                  * If we failed to reclaim anything from this memory cgroup
3166                  * it is time to move on to the next cgroup
3167                  */
3168                 next_mz = NULL;
3169                 if (!reclaimed) {
3170                         do {
3171                                 /*
3172                                  * Loop until we find yet another one.
3173                                  *
3174                                  * By the time we get the soft_limit lock
3175                                  * again, someone might have aded the
3176                                  * group back on the RB tree. Iterate to
3177                                  * make sure we get a different mem.
3178                                  * mem_cgroup_largest_soft_limit_node returns
3179                                  * NULL if no other cgroup is present on
3180                                  * the tree
3181                                  */
3182                                 next_mz =
3183                                 __mem_cgroup_largest_soft_limit_node(mctz);
3184                                 if (next_mz == mz) {
3185                                         css_put(&next_mz->mem->css);
3186                                         next_mz = NULL;
3187                                 } else /* next_mz == NULL or other memcg */
3188                                         break;
3189                         } while (1);
3190                 }
3191                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3192                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3193                 /*
3194                  * One school of thought says that we should not add
3195                  * back the node to the tree if reclaim returns 0.
3196                  * But our reclaim could return 0, simply because due
3197                  * to priority we are exposing a smaller subset of
3198                  * memory to reclaim from. Consider this as a longer
3199                  * term TODO.
3200                  */
3201                 /* If excess == 0, no tree ops */
3202                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3203                 spin_unlock(&mctz->lock);
3204                 css_put(&mz->mem->css);
3205                 loop++;
3206                 /*
3207                  * Could not reclaim anything and there are no more
3208                  * mem cgroups to try or we seem to be looping without
3209                  * reclaiming anything.
3210                  */
3211                 if (!nr_reclaimed &&
3212                         (next_mz == NULL ||
3213                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3214                         break;
3215         } while (!nr_reclaimed);
3216         if (next_mz)
3217                 css_put(&next_mz->mem->css);
3218         return nr_reclaimed;
3219 }
3220
3221 /*
3222  * This routine traverse page_cgroup in given list and drop them all.
3223  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3224  */
3225 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3226                                 int node, int zid, enum lru_list lru)
3227 {
3228         struct zone *zone;
3229         struct mem_cgroup_per_zone *mz;
3230         struct page_cgroup *pc, *busy;
3231         unsigned long flags, loop;
3232         struct list_head *list;
3233         int ret = 0;
3234
3235         zone = &NODE_DATA(node)->node_zones[zid];
3236         mz = mem_cgroup_zoneinfo(mem, node, zid);
3237         list = &mz->lists[lru];
3238
3239         loop = MEM_CGROUP_ZSTAT(mz, lru);
3240         /* give some margin against EBUSY etc...*/
3241         loop += 256;
3242         busy = NULL;
3243         while (loop--) {
3244                 ret = 0;
3245                 spin_lock_irqsave(&zone->lru_lock, flags);
3246                 if (list_empty(list)) {
3247                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3248                         break;
3249                 }
3250                 pc = list_entry(list->prev, struct page_cgroup, lru);
3251                 if (busy == pc) {
3252                         list_move(&pc->lru, list);
3253                         busy = NULL;
3254                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3255                         continue;
3256                 }
3257                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3258
3259                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3260                 if (ret == -ENOMEM)
3261                         break;
3262
3263                 if (ret == -EBUSY || ret == -EINVAL) {
3264                         /* found lock contention or "pc" is obsolete. */
3265                         busy = pc;
3266                         cond_resched();
3267                 } else
3268                         busy = NULL;
3269         }
3270
3271         if (!ret && !list_empty(list))
3272                 return -EBUSY;
3273         return ret;
3274 }
3275
3276 /*
3277  * make mem_cgroup's charge to be 0 if there is no task.
3278  * This enables deleting this mem_cgroup.
3279  */
3280 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3281 {
3282         int ret;
3283         int node, zid, shrink;
3284         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3285         struct cgroup *cgrp = mem->css.cgroup;
3286
3287         css_get(&mem->css);
3288
3289         shrink = 0;
3290         /* should free all ? */
3291         if (free_all)
3292                 goto try_to_free;
3293 move_account:
3294         do {
3295                 ret = -EBUSY;
3296                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3297                         goto out;
3298                 ret = -EINTR;
3299                 if (signal_pending(current))
3300                         goto out;
3301                 /* This is for making all *used* pages to be on LRU. */
3302                 lru_add_drain_all();
3303                 drain_all_stock_sync();
3304                 ret = 0;
3305                 mem_cgroup_start_move(mem);
3306                 for_each_node_state(node, N_HIGH_MEMORY) {
3307                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3308                                 enum lru_list l;
3309                                 for_each_lru(l) {
3310                                         ret = mem_cgroup_force_empty_list(mem,
3311                                                         node, zid, l);
3312                                         if (ret)
3313                                                 break;
3314                                 }
3315                         }
3316                         if (ret)
3317                                 break;
3318                 }
3319                 mem_cgroup_end_move(mem);
3320                 memcg_oom_recover(mem);
3321                 /* it seems parent cgroup doesn't have enough mem */
3322                 if (ret == -ENOMEM)
3323                         goto try_to_free;
3324                 cond_resched();
3325         /* "ret" should also be checked to ensure all lists are empty. */
3326         } while (mem->res.usage > 0 || ret);
3327 out:
3328         css_put(&mem->css);
3329         return ret;
3330
3331 try_to_free:
3332         /* returns EBUSY if there is a task or if we come here twice. */
3333         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3334                 ret = -EBUSY;
3335                 goto out;
3336         }
3337         /* we call try-to-free pages for make this cgroup empty */
3338         lru_add_drain_all();
3339         /* try to free all pages in this cgroup */
3340         shrink = 1;
3341         while (nr_retries && mem->res.usage > 0) {
3342                 int progress;
3343
3344                 if (signal_pending(current)) {
3345                         ret = -EINTR;
3346                         goto out;
3347                 }
3348                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3349                                                 false, get_swappiness(mem));
3350                 if (!progress) {
3351                         nr_retries--;
3352                         /* maybe some writeback is necessary */
3353                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3354                 }
3355
3356         }
3357         lru_add_drain();
3358         /* try move_account...there may be some *locked* pages. */
3359         goto move_account;
3360 }
3361
3362 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3363 {
3364         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3365 }
3366
3367
3368 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3369 {
3370         return mem_cgroup_from_cont(cont)->use_hierarchy;
3371 }
3372
3373 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3374                                         u64 val)
3375 {
3376         int retval = 0;
3377         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3378         struct cgroup *parent = cont->parent;
3379         struct mem_cgroup *parent_mem = NULL;
3380
3381         if (parent)
3382                 parent_mem = mem_cgroup_from_cont(parent);
3383
3384         cgroup_lock();
3385         /*
3386          * If parent's use_hierarchy is set, we can't make any modifications
3387          * in the child subtrees. If it is unset, then the change can
3388          * occur, provided the current cgroup has no children.
3389          *
3390          * For the root cgroup, parent_mem is NULL, we allow value to be
3391          * set if there are no children.
3392          */
3393         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3394                                 (val == 1 || val == 0)) {
3395                 if (list_empty(&cont->children))
3396                         mem->use_hierarchy = val;
3397                 else
3398                         retval = -EBUSY;
3399         } else
3400                 retval = -EINVAL;
3401         cgroup_unlock();
3402
3403         return retval;
3404 }
3405
3406
3407 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3408                                 enum mem_cgroup_stat_index idx)
3409 {
3410         struct mem_cgroup *iter;
3411         s64 val = 0;
3412
3413         /* each per cpu's value can be minus.Then, use s64 */
3414         for_each_mem_cgroup_tree(iter, mem)
3415                 val += mem_cgroup_read_stat(iter, idx);
3416
3417         if (val < 0) /* race ? */
3418                 val = 0;
3419         return val;
3420 }
3421
3422 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3423 {
3424         u64 val;
3425
3426         if (!mem_cgroup_is_root(mem)) {
3427                 if (!swap)
3428                         return res_counter_read_u64(&mem->res, RES_USAGE);
3429                 else
3430                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3431         }
3432
3433         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3434         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3435
3436         if (swap)
3437                 val += mem_cgroup_get_recursive_idx_stat(mem,
3438                                 MEM_CGROUP_STAT_SWAPOUT);
3439
3440         return val << PAGE_SHIFT;
3441 }
3442
3443 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3444 {
3445         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3446         u64 val;
3447         int type, name;
3448
3449         type = MEMFILE_TYPE(cft->private);
3450         name = MEMFILE_ATTR(cft->private);
3451         switch (type) {
3452         case _MEM:
3453                 if (name == RES_USAGE)
3454                         val = mem_cgroup_usage(mem, false);
3455                 else
3456                         val = res_counter_read_u64(&mem->res, name);
3457                 break;
3458         case _MEMSWAP:
3459                 if (name == RES_USAGE)
3460                         val = mem_cgroup_usage(mem, true);
3461                 else
3462                         val = res_counter_read_u64(&mem->memsw, name);
3463                 break;
3464         default:
3465                 BUG();
3466                 break;
3467         }
3468         return val;
3469 }
3470 /*
3471  * The user of this function is...
3472  * RES_LIMIT.
3473  */
3474 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3475                             const char *buffer)
3476 {
3477         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3478         int type, name;
3479         unsigned long long val;
3480         int ret;
3481
3482         type = MEMFILE_TYPE(cft->private);
3483         name = MEMFILE_ATTR(cft->private);
3484         switch (name) {
3485         case RES_LIMIT:
3486                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3487                         ret = -EINVAL;
3488                         break;
3489                 }
3490                 /* This function does all necessary parse...reuse it */
3491                 ret = res_counter_memparse_write_strategy(buffer, &val);
3492                 if (ret)
3493                         break;
3494                 if (type == _MEM)
3495                         ret = mem_cgroup_resize_limit(memcg, val);
3496                 else
3497                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3498                 break;
3499         case RES_SOFT_LIMIT:
3500                 ret = res_counter_memparse_write_strategy(buffer, &val);
3501                 if (ret)
3502                         break;
3503                 /*
3504                  * For memsw, soft limits are hard to implement in terms
3505                  * of semantics, for now, we support soft limits for
3506                  * control without swap
3507                  */
3508                 if (type == _MEM)
3509                         ret = res_counter_set_soft_limit(&memcg->res, val);
3510                 else
3511                         ret = -EINVAL;
3512                 break;
3513         default:
3514                 ret = -EINVAL; /* should be BUG() ? */
3515                 break;
3516         }
3517         return ret;
3518 }
3519
3520 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3521                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3522 {
3523         struct cgroup *cgroup;
3524         unsigned long long min_limit, min_memsw_limit, tmp;
3525
3526         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3527         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3528         cgroup = memcg->css.cgroup;
3529         if (!memcg->use_hierarchy)
3530                 goto out;
3531
3532         while (cgroup->parent) {
3533                 cgroup = cgroup->parent;
3534                 memcg = mem_cgroup_from_cont(cgroup);
3535                 if (!memcg->use_hierarchy)
3536                         break;
3537                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3538                 min_limit = min(min_limit, tmp);
3539                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3540                 min_memsw_limit = min(min_memsw_limit, tmp);
3541         }
3542 out:
3543         *mem_limit = min_limit;
3544         *memsw_limit = min_memsw_limit;
3545         return;
3546 }
3547
3548 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3549 {
3550         struct mem_cgroup *mem;
3551         int type, name;
3552
3553         mem = mem_cgroup_from_cont(cont);
3554         type = MEMFILE_TYPE(event);
3555         name = MEMFILE_ATTR(event);
3556         switch (name) {
3557         case RES_MAX_USAGE:
3558                 if (type == _MEM)
3559                         res_counter_reset_max(&mem->res);
3560                 else
3561                         res_counter_reset_max(&mem->memsw);
3562                 break;
3563         case RES_FAILCNT:
3564                 if (type == _MEM)
3565                         res_counter_reset_failcnt(&mem->res);
3566                 else
3567                         res_counter_reset_failcnt(&mem->memsw);
3568                 break;
3569         }
3570
3571         return 0;
3572 }
3573
3574 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3575                                         struct cftype *cft)
3576 {
3577         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3578 }
3579
3580 #ifdef CONFIG_MMU
3581 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3582                                         struct cftype *cft, u64 val)
3583 {
3584         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3585
3586         if (val >= (1 << NR_MOVE_TYPE))
3587                 return -EINVAL;
3588         /*
3589          * We check this value several times in both in can_attach() and
3590          * attach(), so we need cgroup lock to prevent this value from being
3591          * inconsistent.
3592          */
3593         cgroup_lock();
3594         mem->move_charge_at_immigrate = val;
3595         cgroup_unlock();
3596
3597         return 0;
3598 }
3599 #else
3600 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3601                                         struct cftype *cft, u64 val)
3602 {
3603         return -ENOSYS;
3604 }
3605 #endif
3606
3607
3608 /* For read statistics */
3609 enum {
3610         MCS_CACHE,
3611         MCS_RSS,
3612         MCS_FILE_MAPPED,
3613         MCS_PGPGIN,
3614         MCS_PGPGOUT,
3615         MCS_SWAP,
3616         MCS_INACTIVE_ANON,
3617         MCS_ACTIVE_ANON,
3618         MCS_INACTIVE_FILE,
3619         MCS_ACTIVE_FILE,
3620         MCS_UNEVICTABLE,
3621         NR_MCS_STAT,
3622 };
3623
3624 struct mcs_total_stat {
3625         s64 stat[NR_MCS_STAT];
3626 };
3627
3628 struct {
3629         char *local_name;
3630         char *total_name;
3631 } memcg_stat_strings[NR_MCS_STAT] = {
3632         {"cache", "total_cache"},
3633         {"rss", "total_rss"},
3634         {"mapped_file", "total_mapped_file"},
3635         {"pgpgin", "total_pgpgin"},
3636         {"pgpgout", "total_pgpgout"},
3637         {"swap", "total_swap"},
3638         {"inactive_anon", "total_inactive_anon"},
3639         {"active_anon", "total_active_anon"},
3640         {"inactive_file", "total_inactive_file"},
3641         {"active_file", "total_active_file"},
3642         {"unevictable", "total_unevictable"}
3643 };
3644
3645
3646 static void
3647 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3648 {
3649         s64 val;
3650
3651         /* per cpu stat */
3652         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3653         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3654         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3655         s->stat[MCS_RSS] += val * PAGE_SIZE;
3656         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3657         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3658         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3659         s->stat[MCS_PGPGIN] += val;
3660         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3661         s->stat[MCS_PGPGOUT] += val;
3662         if (do_swap_account) {
3663                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3664                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3665         }
3666
3667         /* per zone stat */
3668         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3669         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3670         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3671         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3672         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3673         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3674         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3675         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3676         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3677         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3678 }
3679
3680 static void
3681 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3682 {
3683         struct mem_cgroup *iter;
3684
3685         for_each_mem_cgroup_tree(iter, mem)
3686                 mem_cgroup_get_local_stat(iter, s);
3687 }
3688
3689 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3690                                  struct cgroup_map_cb *cb)
3691 {
3692         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3693         struct mcs_total_stat mystat;
3694         int i;
3695
3696         memset(&mystat, 0, sizeof(mystat));
3697         mem_cgroup_get_local_stat(mem_cont, &mystat);
3698
3699         for (i = 0; i < NR_MCS_STAT; i++) {
3700                 if (i == MCS_SWAP && !do_swap_account)
3701                         continue;
3702                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3703         }
3704
3705         /* Hierarchical information */
3706         {
3707                 unsigned long long limit, memsw_limit;
3708                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3709                 cb->fill(cb, "hierarchical_memory_limit", limit);
3710                 if (do_swap_account)
3711                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3712         }
3713
3714         memset(&mystat, 0, sizeof(mystat));
3715         mem_cgroup_get_total_stat(mem_cont, &mystat);
3716         for (i = 0; i < NR_MCS_STAT; i++) {
3717                 if (i == MCS_SWAP && !do_swap_account)
3718                         continue;
3719                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3720         }
3721
3722 #ifdef CONFIG_DEBUG_VM
3723         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3724
3725         {
3726                 int nid, zid;
3727                 struct mem_cgroup_per_zone *mz;
3728                 unsigned long recent_rotated[2] = {0, 0};
3729                 unsigned long recent_scanned[2] = {0, 0};
3730
3731                 for_each_online_node(nid)
3732                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3733                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3734
3735                                 recent_rotated[0] +=
3736                                         mz->reclaim_stat.recent_rotated[0];
3737                                 recent_rotated[1] +=
3738                                         mz->reclaim_stat.recent_rotated[1];
3739                                 recent_scanned[0] +=
3740                                         mz->reclaim_stat.recent_scanned[0];
3741                                 recent_scanned[1] +=
3742                                         mz->reclaim_stat.recent_scanned[1];
3743                         }
3744                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3745                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3746                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3747                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3748         }
3749 #endif
3750
3751         return 0;
3752 }
3753
3754 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3755 {
3756         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3757
3758         return get_swappiness(memcg);
3759 }
3760
3761 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3762                                        u64 val)
3763 {
3764         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3765         struct mem_cgroup *parent;
3766
3767         if (val > 100)
3768                 return -EINVAL;
3769
3770         if (cgrp->parent == NULL)
3771                 return -EINVAL;
3772
3773         parent = mem_cgroup_from_cont(cgrp->parent);
3774
3775         cgroup_lock();
3776
3777         /* If under hierarchy, only empty-root can set this value */
3778         if ((parent->use_hierarchy) ||
3779             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3780                 cgroup_unlock();
3781                 return -EINVAL;
3782         }
3783
3784         spin_lock(&memcg->reclaim_param_lock);
3785         memcg->swappiness = val;
3786         spin_unlock(&memcg->reclaim_param_lock);
3787
3788         cgroup_unlock();
3789
3790         return 0;
3791 }
3792
3793 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3794 {
3795         struct mem_cgroup_threshold_ary *t;
3796         u64 usage;
3797         int i;
3798
3799         rcu_read_lock();
3800         if (!swap)
3801                 t = rcu_dereference(memcg->thresholds.primary);
3802         else
3803                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3804
3805         if (!t)
3806                 goto unlock;
3807
3808         usage = mem_cgroup_usage(memcg, swap);
3809
3810         /*
3811          * current_threshold points to threshold just below usage.
3812          * If it's not true, a threshold was crossed after last
3813          * call of __mem_cgroup_threshold().
3814          */
3815         i = t->current_threshold;
3816
3817         /*
3818          * Iterate backward over array of thresholds starting from
3819          * current_threshold and check if a threshold is crossed.
3820          * If none of thresholds below usage is crossed, we read
3821          * only one element of the array here.
3822          */
3823         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3824                 eventfd_signal(t->entries[i].eventfd, 1);
3825
3826         /* i = current_threshold + 1 */
3827         i++;
3828
3829         /*
3830          * Iterate forward over array of thresholds starting from
3831          * current_threshold+1 and check if a threshold is crossed.
3832          * If none of thresholds above usage is crossed, we read
3833          * only one element of the array here.
3834          */
3835         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3836                 eventfd_signal(t->entries[i].eventfd, 1);
3837
3838         /* Update current_threshold */
3839         t->current_threshold = i - 1;
3840 unlock:
3841         rcu_read_unlock();
3842 }
3843
3844 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3845 {
3846         while (memcg) {
3847                 __mem_cgroup_threshold(memcg, false);
3848                 if (do_swap_account)
3849                         __mem_cgroup_threshold(memcg, true);
3850
3851                 memcg = parent_mem_cgroup(memcg);
3852         }
3853 }
3854
3855 static int compare_thresholds(const void *a, const void *b)
3856 {
3857         const struct mem_cgroup_threshold *_a = a;
3858         const struct mem_cgroup_threshold *_b = b;
3859
3860         return _a->threshold - _b->threshold;
3861 }
3862
3863 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3864 {
3865         struct mem_cgroup_eventfd_list *ev;
3866
3867         list_for_each_entry(ev, &mem->oom_notify, list)
3868                 eventfd_signal(ev->eventfd, 1);
3869         return 0;
3870 }
3871
3872 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3873 {
3874         struct mem_cgroup *iter;
3875
3876         for_each_mem_cgroup_tree(iter, mem)
3877                 mem_cgroup_oom_notify_cb(iter);
3878 }
3879
3880 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3881         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3882 {
3883         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3884         struct mem_cgroup_thresholds *thresholds;
3885         struct mem_cgroup_threshold_ary *new;
3886         int type = MEMFILE_TYPE(cft->private);
3887         u64 threshold, usage;
3888         int i, size, ret;
3889
3890         ret = res_counter_memparse_write_strategy(args, &threshold);
3891         if (ret)
3892                 return ret;
3893
3894         mutex_lock(&memcg->thresholds_lock);
3895
3896         if (type == _MEM)
3897                 thresholds = &memcg->thresholds;
3898         else if (type == _MEMSWAP)
3899                 thresholds = &memcg->memsw_thresholds;
3900         else
3901                 BUG();
3902
3903         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3904
3905         /* Check if a threshold crossed before adding a new one */
3906         if (thresholds->primary)
3907                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3908
3909         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3910
3911         /* Allocate memory for new array of thresholds */
3912         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3913                         GFP_KERNEL);
3914         if (!new) {
3915                 ret = -ENOMEM;
3916                 goto unlock;
3917         }
3918         new->size = size;
3919
3920         /* Copy thresholds (if any) to new array */
3921         if (thresholds->primary) {
3922                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3923                                 sizeof(struct mem_cgroup_threshold));
3924         }
3925
3926         /* Add new threshold */
3927         new->entries[size - 1].eventfd = eventfd;
3928         new->entries[size - 1].threshold = threshold;
3929
3930         /* Sort thresholds. Registering of new threshold isn't time-critical */
3931         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3932                         compare_thresholds, NULL);
3933
3934         /* Find current threshold */
3935         new->current_threshold = -1;
3936         for (i = 0; i < size; i++) {
3937                 if (new->entries[i].threshold < usage) {
3938                         /*
3939                          * new->current_threshold will not be used until
3940                          * rcu_assign_pointer(), so it's safe to increment
3941                          * it here.
3942                          */
3943                         ++new->current_threshold;
3944                 }
3945         }
3946
3947         /* Free old spare buffer and save old primary buffer as spare */
3948         kfree(thresholds->spare);
3949         thresholds->spare = thresholds->primary;
3950
3951         rcu_assign_pointer(thresholds->primary, new);
3952
3953         /* To be sure that nobody uses thresholds */
3954         synchronize_rcu();
3955
3956 unlock:
3957         mutex_unlock(&memcg->thresholds_lock);
3958
3959         return ret;
3960 }
3961
3962 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3963         struct cftype *cft, struct eventfd_ctx *eventfd)
3964 {
3965         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3966         struct mem_cgroup_thresholds *thresholds;
3967         struct mem_cgroup_threshold_ary *new;
3968         int type = MEMFILE_TYPE(cft->private);
3969         u64 usage;
3970         int i, j, size;
3971
3972         mutex_lock(&memcg->thresholds_lock);
3973         if (type == _MEM)
3974                 thresholds = &memcg->thresholds;
3975         else if (type == _MEMSWAP)
3976                 thresholds = &memcg->memsw_thresholds;
3977         else
3978                 BUG();
3979
3980         /*
3981          * Something went wrong if we trying to unregister a threshold
3982          * if we don't have thresholds
3983          */
3984         BUG_ON(!thresholds);
3985
3986         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3987
3988         /* Check if a threshold crossed before removing */
3989         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3990
3991         /* Calculate new number of threshold */
3992         size = 0;
3993         for (i = 0; i < thresholds->primary->size; i++) {
3994                 if (thresholds->primary->entries[i].eventfd != eventfd)
3995                         size++;
3996         }
3997
3998         new = thresholds->spare;
3999
4000         /* Set thresholds array to NULL if we don't have thresholds */
4001         if (!size) {
4002                 kfree(new);
4003                 new = NULL;
4004                 goto swap_buffers;
4005         }
4006
4007         new->size = size;
4008
4009         /* Copy thresholds and find current threshold */
4010         new->current_threshold = -1;
4011         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4012                 if (thresholds->primary->entries[i].eventfd == eventfd)
4013                         continue;
4014
4015                 new->entries[j] = thresholds->primary->entries[i];
4016                 if (new->entries[j].threshold < usage) {
4017                         /*
4018                          * new->current_threshold will not be used
4019                          * until rcu_assign_pointer(), so it's safe to increment
4020                          * it here.
4021                          */
4022                         ++new->current_threshold;
4023                 }
4024                 j++;
4025         }
4026
4027 swap_buffers:
4028         /* Swap primary and spare array */
4029         thresholds->spare = thresholds->primary;
4030         rcu_assign_pointer(thresholds->primary, new);
4031
4032         /* To be sure that nobody uses thresholds */
4033         synchronize_rcu();
4034
4035         mutex_unlock(&memcg->thresholds_lock);
4036 }
4037
4038 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4039         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4040 {
4041         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4042         struct mem_cgroup_eventfd_list *event;
4043         int type = MEMFILE_TYPE(cft->private);
4044
4045         BUG_ON(type != _OOM_TYPE);
4046         event = kmalloc(sizeof(*event), GFP_KERNEL);
4047         if (!event)
4048                 return -ENOMEM;
4049
4050         mutex_lock(&memcg_oom_mutex);
4051
4052         event->eventfd = eventfd;
4053         list_add(&event->list, &memcg->oom_notify);
4054
4055         /* already in OOM ? */
4056         if (atomic_read(&memcg->oom_lock))
4057                 eventfd_signal(eventfd, 1);
4058         mutex_unlock(&memcg_oom_mutex);
4059
4060         return 0;
4061 }
4062
4063 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4064         struct cftype *cft, struct eventfd_ctx *eventfd)
4065 {
4066         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4067         struct mem_cgroup_eventfd_list *ev, *tmp;
4068         int type = MEMFILE_TYPE(cft->private);
4069
4070         BUG_ON(type != _OOM_TYPE);
4071
4072         mutex_lock(&memcg_oom_mutex);
4073
4074         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4075                 if (ev->eventfd == eventfd) {
4076                         list_del(&ev->list);
4077                         kfree(ev);
4078                 }
4079         }
4080
4081         mutex_unlock(&memcg_oom_mutex);
4082 }
4083
4084 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4085         struct cftype *cft,  struct cgroup_map_cb *cb)
4086 {
4087         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4088
4089         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4090
4091         if (atomic_read(&mem->oom_lock))
4092                 cb->fill(cb, "under_oom", 1);
4093         else
4094                 cb->fill(cb, "under_oom", 0);
4095         return 0;
4096 }
4097
4098 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4099         struct cftype *cft, u64 val)
4100 {
4101         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4102         struct mem_cgroup *parent;
4103
4104         /* cannot set to root cgroup and only 0 and 1 are allowed */
4105         if (!cgrp->parent || !((val == 0) || (val == 1)))
4106                 return -EINVAL;
4107
4108         parent = mem_cgroup_from_cont(cgrp->parent);
4109
4110         cgroup_lock();
4111         /* oom-kill-disable is a flag for subhierarchy. */
4112         if ((parent->use_hierarchy) ||
4113             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4114                 cgroup_unlock();
4115                 return -EINVAL;
4116         }
4117         mem->oom_kill_disable = val;
4118         if (!val)
4119                 memcg_oom_recover(mem);
4120         cgroup_unlock();
4121         return 0;
4122 }
4123
4124 static struct cftype mem_cgroup_files[] = {
4125         {
4126                 .name = "usage_in_bytes",
4127                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4128                 .read_u64 = mem_cgroup_read,
4129                 .register_event = mem_cgroup_usage_register_event,
4130                 .unregister_event = mem_cgroup_usage_unregister_event,
4131         },
4132         {
4133                 .name = "max_usage_in_bytes",
4134                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4135                 .trigger = mem_cgroup_reset,
4136                 .read_u64 = mem_cgroup_read,
4137         },
4138         {
4139                 .name = "limit_in_bytes",
4140                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4141                 .write_string = mem_cgroup_write,
4142                 .read_u64 = mem_cgroup_read,
4143         },
4144         {
4145                 .name = "soft_limit_in_bytes",
4146                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4147                 .write_string = mem_cgroup_write,
4148                 .read_u64 = mem_cgroup_read,
4149         },
4150         {
4151                 .name = "failcnt",
4152                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4153                 .trigger = mem_cgroup_reset,
4154                 .read_u64 = mem_cgroup_read,
4155         },
4156         {
4157                 .name = "stat",
4158                 .read_map = mem_control_stat_show,
4159         },
4160         {
4161                 .name = "force_empty",
4162                 .trigger = mem_cgroup_force_empty_write,
4163         },
4164         {
4165                 .name = "use_hierarchy",
4166                 .write_u64 = mem_cgroup_hierarchy_write,
4167                 .read_u64 = mem_cgroup_hierarchy_read,
4168         },
4169         {
4170                 .name = "swappiness",
4171                 .read_u64 = mem_cgroup_swappiness_read,
4172                 .write_u64 = mem_cgroup_swappiness_write,
4173         },
4174         {
4175                 .name = "move_charge_at_immigrate",
4176                 .read_u64 = mem_cgroup_move_charge_read,
4177                 .write_u64 = mem_cgroup_move_charge_write,
4178         },
4179         {
4180                 .name = "oom_control",
4181                 .read_map = mem_cgroup_oom_control_read,
4182                 .write_u64 = mem_cgroup_oom_control_write,
4183                 .register_event = mem_cgroup_oom_register_event,
4184                 .unregister_event = mem_cgroup_oom_unregister_event,
4185                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4186         },
4187 };
4188
4189 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4190 static struct cftype memsw_cgroup_files[] = {
4191         {
4192                 .name = "memsw.usage_in_bytes",
4193                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4194                 .read_u64 = mem_cgroup_read,
4195                 .register_event = mem_cgroup_usage_register_event,
4196                 .unregister_event = mem_cgroup_usage_unregister_event,
4197         },
4198         {
4199                 .name = "memsw.max_usage_in_bytes",
4200                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4201                 .trigger = mem_cgroup_reset,
4202                 .read_u64 = mem_cgroup_read,
4203         },
4204         {
4205                 .name = "memsw.limit_in_bytes",
4206                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4207                 .write_string = mem_cgroup_write,
4208                 .read_u64 = mem_cgroup_read,
4209         },
4210         {
4211                 .name = "memsw.failcnt",
4212                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4213                 .trigger = mem_cgroup_reset,
4214                 .read_u64 = mem_cgroup_read,
4215         },
4216 };
4217
4218 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4219 {
4220         if (!do_swap_account)
4221                 return 0;
4222         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4223                                 ARRAY_SIZE(memsw_cgroup_files));
4224 };
4225 #else
4226 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4227 {
4228         return 0;
4229 }
4230 #endif
4231
4232 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4233 {
4234         struct mem_cgroup_per_node *pn;
4235         struct mem_cgroup_per_zone *mz;
4236         enum lru_list l;
4237         int zone, tmp = node;
4238         /*
4239          * This routine is called against possible nodes.
4240          * But it's BUG to call kmalloc() against offline node.
4241          *
4242          * TODO: this routine can waste much memory for nodes which will
4243          *       never be onlined. It's better to use memory hotplug callback
4244          *       function.
4245          */
4246         if (!node_state(node, N_NORMAL_MEMORY))
4247                 tmp = -1;
4248         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4249         if (!pn)
4250                 return 1;
4251
4252         mem->info.nodeinfo[node] = pn;
4253         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4254                 mz = &pn->zoneinfo[zone];
4255                 for_each_lru(l)
4256                         INIT_LIST_HEAD(&mz->lists[l]);
4257                 mz->usage_in_excess = 0;
4258                 mz->on_tree = false;
4259                 mz->mem = mem;
4260         }
4261         return 0;
4262 }
4263
4264 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4265 {
4266         kfree(mem->info.nodeinfo[node]);
4267 }
4268
4269 static struct mem_cgroup *mem_cgroup_alloc(void)
4270 {
4271         struct mem_cgroup *mem;
4272         int size = sizeof(struct mem_cgroup);
4273
4274         /* Can be very big if MAX_NUMNODES is very big */
4275         if (size < PAGE_SIZE)
4276                 mem = kzalloc(size, GFP_KERNEL);
4277         else
4278                 mem = vzalloc(size);
4279
4280         if (!mem)
4281                 return NULL;
4282
4283         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4284         if (!mem->stat)
4285                 goto out_free;
4286         spin_lock_init(&mem->pcp_counter_lock);
4287         return mem;
4288
4289 out_free:
4290         if (size < PAGE_SIZE)
4291                 kfree(mem);
4292         else
4293                 vfree(mem);
4294         return NULL;
4295 }
4296
4297 /*
4298  * At destroying mem_cgroup, references from swap_cgroup can remain.
4299  * (scanning all at force_empty is too costly...)
4300  *
4301  * Instead of clearing all references at force_empty, we remember
4302  * the number of reference from swap_cgroup and free mem_cgroup when
4303  * it goes down to 0.
4304  *
4305  * Removal of cgroup itself succeeds regardless of refs from swap.
4306  */
4307
4308 static void __mem_cgroup_free(struct mem_cgroup *mem)
4309 {
4310         int node;
4311
4312         mem_cgroup_remove_from_trees(mem);
4313         free_css_id(&mem_cgroup_subsys, &mem->css);
4314
4315         for_each_node_state(node, N_POSSIBLE)
4316                 free_mem_cgroup_per_zone_info(mem, node);
4317
4318         free_percpu(mem->stat);
4319         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4320                 kfree(mem);
4321         else
4322                 vfree(mem);
4323 }
4324
4325 static void mem_cgroup_get(struct mem_cgroup *mem)
4326 {
4327         atomic_inc(&mem->refcnt);
4328 }
4329
4330 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4331 {
4332         if (atomic_sub_and_test(count, &mem->refcnt)) {
4333                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4334                 __mem_cgroup_free(mem);
4335                 if (parent)
4336                         mem_cgroup_put(parent);
4337         }
4338 }
4339
4340 static void mem_cgroup_put(struct mem_cgroup *mem)
4341 {
4342         __mem_cgroup_put(mem, 1);
4343 }
4344
4345 /*
4346  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4347  */
4348 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4349 {
4350         if (!mem->res.parent)
4351                 return NULL;
4352         return mem_cgroup_from_res_counter(mem->res.parent, res);
4353 }
4354
4355 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4356 static void __init enable_swap_cgroup(void)
4357 {
4358         if (!mem_cgroup_disabled() && really_do_swap_account)
4359                 do_swap_account = 1;
4360 }
4361 #else
4362 static void __init enable_swap_cgroup(void)
4363 {
4364 }
4365 #endif
4366
4367 static int mem_cgroup_soft_limit_tree_init(void)
4368 {
4369         struct mem_cgroup_tree_per_node *rtpn;
4370         struct mem_cgroup_tree_per_zone *rtpz;
4371         int tmp, node, zone;
4372
4373         for_each_node_state(node, N_POSSIBLE) {
4374                 tmp = node;
4375                 if (!node_state(node, N_NORMAL_MEMORY))
4376                         tmp = -1;
4377                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4378                 if (!rtpn)
4379                         return 1;
4380
4381                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4382
4383                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4384                         rtpz = &rtpn->rb_tree_per_zone[zone];
4385                         rtpz->rb_root = RB_ROOT;
4386                         spin_lock_init(&rtpz->lock);
4387                 }
4388         }
4389         return 0;
4390 }
4391
4392 static struct cgroup_subsys_state * __ref
4393 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4394 {
4395         struct mem_cgroup *mem, *parent;
4396         long error = -ENOMEM;
4397         int node;
4398
4399         mem = mem_cgroup_alloc();
4400         if (!mem)
4401                 return ERR_PTR(error);
4402
4403         for_each_node_state(node, N_POSSIBLE)
4404                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4405                         goto free_out;
4406
4407         /* root ? */
4408         if (cont->parent == NULL) {
4409                 int cpu;
4410                 enable_swap_cgroup();
4411                 parent = NULL;
4412                 root_mem_cgroup = mem;
4413                 if (mem_cgroup_soft_limit_tree_init())
4414                         goto free_out;
4415                 for_each_possible_cpu(cpu) {
4416                         struct memcg_stock_pcp *stock =
4417                                                 &per_cpu(memcg_stock, cpu);
4418                         INIT_WORK(&stock->work, drain_local_stock);
4419                 }
4420                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4421         } else {
4422                 parent = mem_cgroup_from_cont(cont->parent);
4423                 mem->use_hierarchy = parent->use_hierarchy;
4424                 mem->oom_kill_disable = parent->oom_kill_disable;
4425         }
4426
4427         if (parent && parent->use_hierarchy) {
4428                 res_counter_init(&mem->res, &parent->res);
4429                 res_counter_init(&mem->memsw, &parent->memsw);
4430                 /*
4431                  * We increment refcnt of the parent to ensure that we can
4432                  * safely access it on res_counter_charge/uncharge.
4433                  * This refcnt will be decremented when freeing this
4434                  * mem_cgroup(see mem_cgroup_put).
4435                  */
4436                 mem_cgroup_get(parent);
4437         } else {
4438                 res_counter_init(&mem->res, NULL);
4439                 res_counter_init(&mem->memsw, NULL);
4440         }
4441         mem->last_scanned_child = 0;
4442         spin_lock_init(&mem->reclaim_param_lock);
4443         INIT_LIST_HEAD(&mem->oom_notify);
4444
4445         if (parent)
4446                 mem->swappiness = get_swappiness(parent);
4447         atomic_set(&mem->refcnt, 1);
4448         mem->move_charge_at_immigrate = 0;
4449         mutex_init(&mem->thresholds_lock);
4450         return &mem->css;
4451 free_out:
4452         __mem_cgroup_free(mem);
4453         root_mem_cgroup = NULL;
4454         return ERR_PTR(error);
4455 }
4456
4457 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4458                                         struct cgroup *cont)
4459 {
4460         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4461
4462         return mem_cgroup_force_empty(mem, false);
4463 }
4464
4465 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4466                                 struct cgroup *cont)
4467 {
4468         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4469
4470         mem_cgroup_put(mem);
4471 }
4472
4473 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4474                                 struct cgroup *cont)
4475 {
4476         int ret;
4477
4478         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4479                                 ARRAY_SIZE(mem_cgroup_files));
4480
4481         if (!ret)
4482                 ret = register_memsw_files(cont, ss);
4483         return ret;
4484 }
4485
4486 #ifdef CONFIG_MMU
4487 /* Handlers for move charge at task migration. */
4488 #define PRECHARGE_COUNT_AT_ONCE 256
4489 static int mem_cgroup_do_precharge(unsigned long count)
4490 {
4491         int ret = 0;
4492         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4493         struct mem_cgroup *mem = mc.to;
4494
4495         if (mem_cgroup_is_root(mem)) {
4496                 mc.precharge += count;
4497                 /* we don't need css_get for root */
4498                 return ret;
4499         }
4500         /* try to charge at once */
4501         if (count > 1) {
4502                 struct res_counter *dummy;
4503                 /*
4504                  * "mem" cannot be under rmdir() because we've already checked
4505                  * by cgroup_lock_live_cgroup() that it is not removed and we
4506                  * are still under the same cgroup_mutex. So we can postpone
4507                  * css_get().
4508                  */
4509                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4510                         goto one_by_one;
4511                 if (do_swap_account && res_counter_charge(&mem->memsw,
4512                                                 PAGE_SIZE * count, &dummy)) {
4513                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4514                         goto one_by_one;
4515                 }
4516                 mc.precharge += count;
4517                 return ret;
4518         }
4519 one_by_one:
4520         /* fall back to one by one charge */
4521         while (count--) {
4522                 if (signal_pending(current)) {
4523                         ret = -EINTR;
4524                         break;
4525                 }
4526                 if (!batch_count--) {
4527                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4528                         cond_resched();
4529                 }
4530                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4531                                               PAGE_SIZE);
4532                 if (ret || !mem)
4533                         /* mem_cgroup_clear_mc() will do uncharge later */
4534                         return -ENOMEM;
4535                 mc.precharge++;
4536         }
4537         return ret;
4538 }
4539
4540 /**
4541  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4542  * @vma: the vma the pte to be checked belongs
4543  * @addr: the address corresponding to the pte to be checked
4544  * @ptent: the pte to be checked
4545  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4546  *
4547  * Returns
4548  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4549  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4550  *     move charge. if @target is not NULL, the page is stored in target->page
4551  *     with extra refcnt got(Callers should handle it).
4552  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4553  *     target for charge migration. if @target is not NULL, the entry is stored
4554  *     in target->ent.
4555  *
4556  * Called with pte lock held.
4557  */
4558 union mc_target {
4559         struct page     *page;
4560         swp_entry_t     ent;
4561 };
4562
4563 enum mc_target_type {
4564         MC_TARGET_NONE, /* not used */
4565         MC_TARGET_PAGE,
4566         MC_TARGET_SWAP,
4567 };
4568
4569 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4570                                                 unsigned long addr, pte_t ptent)
4571 {
4572         struct page *page = vm_normal_page(vma, addr, ptent);
4573
4574         if (!page || !page_mapped(page))
4575                 return NULL;
4576         if (PageAnon(page)) {
4577                 /* we don't move shared anon */
4578                 if (!move_anon() || page_mapcount(page) > 2)
4579                         return NULL;
4580         } else if (!move_file())
4581                 /* we ignore mapcount for file pages */
4582                 return NULL;
4583         if (!get_page_unless_zero(page))
4584                 return NULL;
4585
4586         return page;
4587 }
4588
4589 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4590                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4591 {
4592         int usage_count;
4593         struct page *page = NULL;
4594         swp_entry_t ent = pte_to_swp_entry(ptent);
4595
4596         if (!move_anon() || non_swap_entry(ent))
4597                 return NULL;
4598         usage_count = mem_cgroup_count_swap_user(ent, &page);
4599         if (usage_count > 1) { /* we don't move shared anon */
4600                 if (page)
4601                         put_page(page);
4602                 return NULL;
4603         }
4604         if (do_swap_account)
4605                 entry->val = ent.val;
4606
4607         return page;
4608 }
4609
4610 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4611                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4612 {
4613         struct page *page = NULL;
4614         struct inode *inode;
4615         struct address_space *mapping;
4616         pgoff_t pgoff;
4617
4618         if (!vma->vm_file) /* anonymous vma */
4619                 return NULL;
4620         if (!move_file())
4621                 return NULL;
4622
4623         inode = vma->vm_file->f_path.dentry->d_inode;
4624         mapping = vma->vm_file->f_mapping;
4625         if (pte_none(ptent))
4626                 pgoff = linear_page_index(vma, addr);
4627         else /* pte_file(ptent) is true */
4628                 pgoff = pte_to_pgoff(ptent);
4629
4630         /* page is moved even if it's not RSS of this task(page-faulted). */
4631         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4632                 page = find_get_page(mapping, pgoff);
4633         } else { /* shmem/tmpfs file. we should take account of swap too. */
4634                 swp_entry_t ent;
4635                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4636                 if (do_swap_account)
4637                         entry->val = ent.val;
4638         }
4639
4640         return page;
4641 }
4642
4643 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4644                 unsigned long addr, pte_t ptent, union mc_target *target)
4645 {
4646         struct page *page = NULL;
4647         struct page_cgroup *pc;
4648         int ret = 0;
4649         swp_entry_t ent = { .val = 0 };
4650
4651         if (pte_present(ptent))
4652                 page = mc_handle_present_pte(vma, addr, ptent);
4653         else if (is_swap_pte(ptent))
4654                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4655         else if (pte_none(ptent) || pte_file(ptent))
4656                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4657
4658         if (!page && !ent.val)
4659                 return 0;
4660         if (page) {
4661                 pc = lookup_page_cgroup(page);
4662                 /*
4663                  * Do only loose check w/o page_cgroup lock.
4664                  * mem_cgroup_move_account() checks the pc is valid or not under
4665                  * the lock.
4666                  */
4667                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4668                         ret = MC_TARGET_PAGE;
4669                         if (target)
4670                                 target->page = page;
4671                 }
4672                 if (!ret || !target)
4673                         put_page(page);
4674         }
4675         /* There is a swap entry and a page doesn't exist or isn't charged */
4676         if (ent.val && !ret &&
4677                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4678                 ret = MC_TARGET_SWAP;
4679                 if (target)
4680                         target->ent = ent;
4681         }
4682         return ret;
4683 }
4684
4685 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4686                                         unsigned long addr, unsigned long end,
4687                                         struct mm_walk *walk)
4688 {
4689         struct vm_area_struct *vma = walk->private;
4690         pte_t *pte;
4691         spinlock_t *ptl;
4692
4693         VM_BUG_ON(pmd_trans_huge(*pmd));
4694         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4695         for (; addr != end; pte++, addr += PAGE_SIZE)
4696                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4697                         mc.precharge++; /* increment precharge temporarily */
4698         pte_unmap_unlock(pte - 1, ptl);
4699         cond_resched();
4700
4701         return 0;
4702 }
4703
4704 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4705 {
4706         unsigned long precharge;
4707         struct vm_area_struct *vma;
4708
4709         down_read(&mm->mmap_sem);
4710         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4711                 struct mm_walk mem_cgroup_count_precharge_walk = {
4712                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4713                         .mm = mm,
4714                         .private = vma,
4715                 };
4716                 if (is_vm_hugetlb_page(vma))
4717                         continue;
4718                 walk_page_range(vma->vm_start, vma->vm_end,
4719                                         &mem_cgroup_count_precharge_walk);
4720         }
4721         up_read(&mm->mmap_sem);
4722
4723         precharge = mc.precharge;
4724         mc.precharge = 0;
4725
4726         return precharge;
4727 }
4728
4729 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4730 {
4731         unsigned long precharge = mem_cgroup_count_precharge(mm);
4732
4733         VM_BUG_ON(mc.moving_task);
4734         mc.moving_task = current;
4735         return mem_cgroup_do_precharge(precharge);
4736 }
4737
4738 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4739 static void __mem_cgroup_clear_mc(void)
4740 {
4741         struct mem_cgroup *from = mc.from;
4742         struct mem_cgroup *to = mc.to;
4743
4744         /* we must uncharge all the leftover precharges from mc.to */
4745         if (mc.precharge) {
4746                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4747                 mc.precharge = 0;
4748         }
4749         /*
4750          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4751          * we must uncharge here.
4752          */
4753         if (mc.moved_charge) {
4754                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4755                 mc.moved_charge = 0;
4756         }
4757         /* we must fixup refcnts and charges */
4758         if (mc.moved_swap) {
4759                 /* uncharge swap account from the old cgroup */
4760                 if (!mem_cgroup_is_root(mc.from))
4761                         res_counter_uncharge(&mc.from->memsw,
4762                                                 PAGE_SIZE * mc.moved_swap);
4763                 __mem_cgroup_put(mc.from, mc.moved_swap);
4764
4765                 if (!mem_cgroup_is_root(mc.to)) {
4766                         /*
4767                          * we charged both to->res and to->memsw, so we should
4768                          * uncharge to->res.
4769                          */
4770                         res_counter_uncharge(&mc.to->res,
4771                                                 PAGE_SIZE * mc.moved_swap);
4772                 }
4773                 /* we've already done mem_cgroup_get(mc.to) */
4774                 mc.moved_swap = 0;
4775         }
4776         memcg_oom_recover(from);
4777         memcg_oom_recover(to);
4778         wake_up_all(&mc.waitq);
4779 }
4780
4781 static void mem_cgroup_clear_mc(void)
4782 {
4783         struct mem_cgroup *from = mc.from;
4784
4785         /*
4786          * we must clear moving_task before waking up waiters at the end of
4787          * task migration.
4788          */
4789         mc.moving_task = NULL;
4790         __mem_cgroup_clear_mc();
4791         spin_lock(&mc.lock);
4792         mc.from = NULL;
4793         mc.to = NULL;
4794         spin_unlock(&mc.lock);
4795         mem_cgroup_end_move(from);
4796 }
4797
4798 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4799                                 struct cgroup *cgroup,
4800                                 struct task_struct *p,
4801                                 bool threadgroup)
4802 {
4803         int ret = 0;
4804         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4805
4806         if (mem->move_charge_at_immigrate) {
4807                 struct mm_struct *mm;
4808                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4809
4810                 VM_BUG_ON(from == mem);
4811
4812                 mm = get_task_mm(p);
4813                 if (!mm)
4814                         return 0;
4815                 /* We move charges only when we move a owner of the mm */
4816                 if (mm->owner == p) {
4817                         VM_BUG_ON(mc.from);
4818                         VM_BUG_ON(mc.to);
4819                         VM_BUG_ON(mc.precharge);
4820                         VM_BUG_ON(mc.moved_charge);
4821                         VM_BUG_ON(mc.moved_swap);
4822                         mem_cgroup_start_move(from);
4823                         spin_lock(&mc.lock);
4824                         mc.from = from;
4825                         mc.to = mem;
4826                         spin_unlock(&mc.lock);
4827                         /* We set mc.moving_task later */
4828
4829                         ret = mem_cgroup_precharge_mc(mm);
4830                         if (ret)
4831                                 mem_cgroup_clear_mc();
4832                 }
4833                 mmput(mm);
4834         }
4835         return ret;
4836 }
4837
4838 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4839                                 struct cgroup *cgroup,
4840                                 struct task_struct *p,
4841                                 bool threadgroup)
4842 {
4843         mem_cgroup_clear_mc();
4844 }
4845
4846 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4847                                 unsigned long addr, unsigned long end,
4848                                 struct mm_walk *walk)
4849 {
4850         int ret = 0;
4851         struct vm_area_struct *vma = walk->private;
4852         pte_t *pte;
4853         spinlock_t *ptl;
4854
4855 retry:
4856         VM_BUG_ON(pmd_trans_huge(*pmd));
4857         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4858         for (; addr != end; addr += PAGE_SIZE) {
4859                 pte_t ptent = *(pte++);
4860                 union mc_target target;
4861                 int type;
4862                 struct page *page;
4863                 struct page_cgroup *pc;
4864                 swp_entry_t ent;
4865
4866                 if (!mc.precharge)
4867                         break;
4868
4869                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4870                 switch (type) {
4871                 case MC_TARGET_PAGE:
4872                         page = target.page;
4873                         if (isolate_lru_page(page))
4874                                 goto put;
4875                         pc = lookup_page_cgroup(page);
4876                         if (!mem_cgroup_move_account(pc,
4877                                         mc.from, mc.to, false, PAGE_SIZE)) {
4878                                 mc.precharge--;
4879                                 /* we uncharge from mc.from later. */
4880                                 mc.moved_charge++;
4881                         }
4882                         putback_lru_page(page);
4883 put:                    /* is_target_pte_for_mc() gets the page */
4884                         put_page(page);
4885                         break;
4886                 case MC_TARGET_SWAP:
4887                         ent = target.ent;
4888                         if (!mem_cgroup_move_swap_account(ent,
4889                                                 mc.from, mc.to, false)) {
4890                                 mc.precharge--;
4891                                 /* we fixup refcnts and charges later. */
4892                                 mc.moved_swap++;
4893                         }
4894                         break;
4895                 default:
4896                         break;
4897                 }
4898         }
4899         pte_unmap_unlock(pte - 1, ptl);
4900         cond_resched();
4901
4902         if (addr != end) {
4903                 /*
4904                  * We have consumed all precharges we got in can_attach().
4905                  * We try charge one by one, but don't do any additional
4906                  * charges to mc.to if we have failed in charge once in attach()
4907                  * phase.
4908                  */
4909                 ret = mem_cgroup_do_precharge(1);
4910                 if (!ret)
4911                         goto retry;
4912         }
4913
4914         return ret;
4915 }
4916
4917 static void mem_cgroup_move_charge(struct mm_struct *mm)
4918 {
4919         struct vm_area_struct *vma;
4920
4921         lru_add_drain_all();
4922 retry:
4923         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4924                 /*
4925                  * Someone who are holding the mmap_sem might be waiting in
4926                  * waitq. So we cancel all extra charges, wake up all waiters,
4927                  * and retry. Because we cancel precharges, we might not be able
4928                  * to move enough charges, but moving charge is a best-effort
4929                  * feature anyway, so it wouldn't be a big problem.
4930                  */
4931                 __mem_cgroup_clear_mc();
4932                 cond_resched();
4933                 goto retry;
4934         }
4935         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4936                 int ret;
4937                 struct mm_walk mem_cgroup_move_charge_walk = {
4938                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4939                         .mm = mm,
4940                         .private = vma,
4941                 };
4942                 if (is_vm_hugetlb_page(vma))
4943                         continue;
4944                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4945                                                 &mem_cgroup_move_charge_walk);
4946                 if (ret)
4947                         /*
4948                          * means we have consumed all precharges and failed in
4949                          * doing additional charge. Just abandon here.
4950                          */
4951                         break;
4952         }
4953         up_read(&mm->mmap_sem);
4954 }
4955
4956 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4957                                 struct cgroup *cont,
4958                                 struct cgroup *old_cont,
4959                                 struct task_struct *p,
4960                                 bool threadgroup)
4961 {
4962         struct mm_struct *mm;
4963
4964         if (!mc.to)
4965                 /* no need to move charge */
4966                 return;
4967
4968         mm = get_task_mm(p);
4969         if (mm) {
4970                 mem_cgroup_move_charge(mm);
4971                 mmput(mm);
4972         }
4973         mem_cgroup_clear_mc();
4974 }
4975 #else   /* !CONFIG_MMU */
4976 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4977                                 struct cgroup *cgroup,
4978                                 struct task_struct *p,
4979                                 bool threadgroup)
4980 {
4981         return 0;
4982 }
4983 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4984                                 struct cgroup *cgroup,
4985                                 struct task_struct *p,
4986                                 bool threadgroup)
4987 {
4988 }
4989 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4990                                 struct cgroup *cont,
4991                                 struct cgroup *old_cont,
4992                                 struct task_struct *p,
4993                                 bool threadgroup)
4994 {
4995 }
4996 #endif
4997
4998 struct cgroup_subsys mem_cgroup_subsys = {
4999         .name = "memory",
5000         .subsys_id = mem_cgroup_subsys_id,
5001         .create = mem_cgroup_create,
5002         .pre_destroy = mem_cgroup_pre_destroy,
5003         .destroy = mem_cgroup_destroy,
5004         .populate = mem_cgroup_populate,
5005         .can_attach = mem_cgroup_can_attach,
5006         .cancel_attach = mem_cgroup_cancel_attach,
5007         .attach = mem_cgroup_move_task,
5008         .early_init = 0,
5009         .use_id = 1,
5010 };
5011
5012 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5013 static int __init enable_swap_account(char *s)
5014 {
5015         /* consider enabled if no parameter or 1 is given */
5016         if (!s || !strcmp(s, "1"))
5017                 really_do_swap_account = 1;
5018         else if (!strcmp(s, "0"))
5019                 really_do_swap_account = 0;
5020         return 1;
5021 }
5022 __setup("swapaccount", enable_swap_account);
5023
5024 static int __init disable_swap_account(char *s)
5025 {
5026         enable_swap_account("0");
5027         return 1;
5028 }
5029 __setup("noswapaccount", disable_swap_account);
5030 #endif