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