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