memcg: prevent endless loop when charging huge pages to near-limit group
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2 of the License, or
16  * (at your option) any later version.
17  *
18  * This program is distributed in the hope that it will be useful,
19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
21  * GNU General Public License for more details.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
44 #include <linux/fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
51 #include "internal.h"
52
53 #include <asm/uaccess.h>
54
55 #include <trace/events/vmscan.h>
56
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES      5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
60
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
64
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata = 1;
68 #else
69 static int really_do_swap_account __initdata = 0;
70 #endif
71
72 #else
73 #define do_swap_account         (0)
74 #endif
75
76 /*
77  * Per memcg event counter is incremented at every pagein/pageout. This counter
78  * is used for trigger some periodic events. This is straightforward and better
79  * than using jiffies etc. to handle periodic memcg event.
80  *
81  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
82  */
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
85
86 /*
87  * Statistics for memory cgroup.
88  */
89 enum mem_cgroup_stat_index {
90         /*
91          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92          */
93         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
94         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
95         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
96         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
97         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
98         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
99         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
100         /* incremented at every  pagein/pageout */
101         MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
102         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
103
104         MEM_CGROUP_STAT_NSTATS,
105 };
106
107 struct mem_cgroup_stat_cpu {
108         s64 count[MEM_CGROUP_STAT_NSTATS];
109 };
110
111 /*
112  * per-zone information in memory controller.
113  */
114 struct mem_cgroup_per_zone {
115         /*
116          * spin_lock to protect the per cgroup LRU
117          */
118         struct list_head        lists[NR_LRU_LISTS];
119         unsigned long           count[NR_LRU_LISTS];
120
121         struct zone_reclaim_stat reclaim_stat;
122         struct rb_node          tree_node;      /* RB tree node */
123         unsigned long long      usage_in_excess;/* Set to the value by which */
124                                                 /* the soft limit is exceeded*/
125         bool                    on_tree;
126         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
127                                                 /* use container_of        */
128 };
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
131
132 struct mem_cgroup_per_node {
133         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
134 };
135
136 struct mem_cgroup_lru_info {
137         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
138 };
139
140 /*
141  * Cgroups above their limits are maintained in a RB-Tree, independent of
142  * their hierarchy representation
143  */
144
145 struct mem_cgroup_tree_per_zone {
146         struct rb_root rb_root;
147         spinlock_t lock;
148 };
149
150 struct mem_cgroup_tree_per_node {
151         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
152 };
153
154 struct mem_cgroup_tree {
155         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
156 };
157
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
159
160 struct mem_cgroup_threshold {
161         struct eventfd_ctx *eventfd;
162         u64 threshold;
163 };
164
165 /* For threshold */
166 struct mem_cgroup_threshold_ary {
167         /* An array index points to threshold just below usage. */
168         int current_threshold;
169         /* Size of entries[] */
170         unsigned int size;
171         /* Array of thresholds */
172         struct mem_cgroup_threshold entries[0];
173 };
174
175 struct mem_cgroup_thresholds {
176         /* Primary thresholds array */
177         struct mem_cgroup_threshold_ary *primary;
178         /*
179          * Spare threshold array.
180          * This is needed to make mem_cgroup_unregister_event() "never fail".
181          * It must be able to store at least primary->size - 1 entries.
182          */
183         struct mem_cgroup_threshold_ary *spare;
184 };
185
186 /* for OOM */
187 struct mem_cgroup_eventfd_list {
188         struct list_head list;
189         struct eventfd_ctx *eventfd;
190 };
191
192 static void mem_cgroup_threshold(struct mem_cgroup *mem);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
194
195 /*
196  * The memory controller data structure. The memory controller controls both
197  * page cache and RSS per cgroup. We would eventually like to provide
198  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199  * to help the administrator determine what knobs to tune.
200  *
201  * TODO: Add a water mark for the memory controller. Reclaim will begin when
202  * we hit the water mark. May be even add a low water mark, such that
203  * no reclaim occurs from a cgroup at it's low water mark, this is
204  * a feature that will be implemented much later in the future.
205  */
206 struct mem_cgroup {
207         struct cgroup_subsys_state css;
208         /*
209          * the counter to account for memory usage
210          */
211         struct res_counter res;
212         /*
213          * the counter to account for mem+swap usage.
214          */
215         struct res_counter memsw;
216         /*
217          * Per cgroup active and inactive list, similar to the
218          * per zone LRU lists.
219          */
220         struct mem_cgroup_lru_info info;
221
222         /*
223           protect against reclaim related member.
224         */
225         spinlock_t reclaim_param_lock;
226
227         /*
228          * While reclaiming in a hierarchy, we cache the last child we
229          * reclaimed from.
230          */
231         int last_scanned_child;
232         /*
233          * Should the accounting and control be hierarchical, per subtree?
234          */
235         bool use_hierarchy;
236         atomic_t        oom_lock;
237         atomic_t        refcnt;
238
239         unsigned int    swappiness;
240         /* OOM-Killer disable */
241         int             oom_kill_disable;
242
243         /* set when res.limit == memsw.limit */
244         bool            memsw_is_minimum;
245
246         /* protect arrays of thresholds */
247         struct mutex thresholds_lock;
248
249         /* thresholds for memory usage. RCU-protected */
250         struct mem_cgroup_thresholds thresholds;
251
252         /* thresholds for mem+swap usage. RCU-protected */
253         struct mem_cgroup_thresholds memsw_thresholds;
254
255         /* For oom notifier event fd */
256         struct list_head oom_notify;
257
258         /*
259          * Should we move charges of a task when a task is moved into this
260          * mem_cgroup ? And what type of charges should we move ?
261          */
262         unsigned long   move_charge_at_immigrate;
263         /*
264          * percpu counter.
265          */
266         struct mem_cgroup_stat_cpu *stat;
267         /*
268          * used when a cpu is offlined or other synchronizations
269          * See mem_cgroup_read_stat().
270          */
271         struct mem_cgroup_stat_cpu nocpu_base;
272         spinlock_t pcp_counter_lock;
273 };
274
275 /* Stuffs for move charges at task migration. */
276 /*
277  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278  * left-shifted bitmap of these types.
279  */
280 enum move_type {
281         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
282         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
283         NR_MOVE_TYPE,
284 };
285
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct {
288         spinlock_t        lock; /* for from, to */
289         struct mem_cgroup *from;
290         struct mem_cgroup *to;
291         unsigned long precharge;
292         unsigned long moved_charge;
293         unsigned long moved_swap;
294         struct task_struct *moving_task;        /* a task moving charges */
295         wait_queue_head_t waitq;                /* a waitq for other context */
296 } mc = {
297         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
298         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
299 };
300
301 static bool move_anon(void)
302 {
303         return test_bit(MOVE_CHARGE_TYPE_ANON,
304                                         &mc.to->move_charge_at_immigrate);
305 }
306
307 static bool move_file(void)
308 {
309         return test_bit(MOVE_CHARGE_TYPE_FILE,
310                                         &mc.to->move_charge_at_immigrate);
311 }
312
313 /*
314  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315  * limit reclaim to prevent infinite loops, if they ever occur.
316  */
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
319
320 enum charge_type {
321         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
322         MEM_CGROUP_CHARGE_TYPE_MAPPED,
323         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
324         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
325         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
326         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
327         NR_CHARGE_TYPE,
328 };
329
330 /* only for here (for easy reading.) */
331 #define PCGF_CACHE      (1UL << PCG_CACHE)
332 #define PCGF_USED       (1UL << PCG_USED)
333 #define PCGF_LOCK       (1UL << PCG_LOCK)
334 /* Not used, but added here for completeness */
335 #define PCGF_ACCT       (1UL << PCG_ACCT)
336
337 /* for encoding cft->private value on file */
338 #define _MEM                    (0)
339 #define _MEMSWAP                (1)
340 #define _OOM_TYPE               (2)
341 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
342 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
343 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
344 /* Used for OOM nofiier */
345 #define OOM_CONTROL             (0)
346
347 /*
348  * Reclaim flags for mem_cgroup_hierarchical_reclaim
349  */
350 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
351 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
352 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
353 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
354 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
355 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
356
357 static void mem_cgroup_get(struct mem_cgroup *mem);
358 static void mem_cgroup_put(struct mem_cgroup *mem);
359 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
360 static void drain_all_stock_async(void);
361
362 static struct mem_cgroup_per_zone *
363 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
364 {
365         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
366 }
367
368 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
369 {
370         return &mem->css;
371 }
372
373 static struct mem_cgroup_per_zone *
374 page_cgroup_zoneinfo(struct page_cgroup *pc)
375 {
376         struct mem_cgroup *mem = pc->mem_cgroup;
377         int nid = page_cgroup_nid(pc);
378         int zid = page_cgroup_zid(pc);
379
380         if (!mem)
381                 return NULL;
382
383         return mem_cgroup_zoneinfo(mem, nid, zid);
384 }
385
386 static struct mem_cgroup_tree_per_zone *
387 soft_limit_tree_node_zone(int nid, int zid)
388 {
389         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
390 }
391
392 static struct mem_cgroup_tree_per_zone *
393 soft_limit_tree_from_page(struct page *page)
394 {
395         int nid = page_to_nid(page);
396         int zid = page_zonenum(page);
397
398         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
399 }
400
401 static void
402 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
403                                 struct mem_cgroup_per_zone *mz,
404                                 struct mem_cgroup_tree_per_zone *mctz,
405                                 unsigned long long new_usage_in_excess)
406 {
407         struct rb_node **p = &mctz->rb_root.rb_node;
408         struct rb_node *parent = NULL;
409         struct mem_cgroup_per_zone *mz_node;
410
411         if (mz->on_tree)
412                 return;
413
414         mz->usage_in_excess = new_usage_in_excess;
415         if (!mz->usage_in_excess)
416                 return;
417         while (*p) {
418                 parent = *p;
419                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
420                                         tree_node);
421                 if (mz->usage_in_excess < mz_node->usage_in_excess)
422                         p = &(*p)->rb_left;
423                 /*
424                  * We can't avoid mem cgroups that are over their soft
425                  * limit by the same amount
426                  */
427                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
428                         p = &(*p)->rb_right;
429         }
430         rb_link_node(&mz->tree_node, parent, p);
431         rb_insert_color(&mz->tree_node, &mctz->rb_root);
432         mz->on_tree = true;
433 }
434
435 static void
436 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
437                                 struct mem_cgroup_per_zone *mz,
438                                 struct mem_cgroup_tree_per_zone *mctz)
439 {
440         if (!mz->on_tree)
441                 return;
442         rb_erase(&mz->tree_node, &mctz->rb_root);
443         mz->on_tree = false;
444 }
445
446 static void
447 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
448                                 struct mem_cgroup_per_zone *mz,
449                                 struct mem_cgroup_tree_per_zone *mctz)
450 {
451         spin_lock(&mctz->lock);
452         __mem_cgroup_remove_exceeded(mem, mz, mctz);
453         spin_unlock(&mctz->lock);
454 }
455
456
457 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
458 {
459         unsigned long long excess;
460         struct mem_cgroup_per_zone *mz;
461         struct mem_cgroup_tree_per_zone *mctz;
462         int nid = page_to_nid(page);
463         int zid = page_zonenum(page);
464         mctz = soft_limit_tree_from_page(page);
465
466         /*
467          * Necessary to update all ancestors when hierarchy is used.
468          * because their event counter is not touched.
469          */
470         for (; mem; mem = parent_mem_cgroup(mem)) {
471                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
472                 excess = res_counter_soft_limit_excess(&mem->res);
473                 /*
474                  * We have to update the tree if mz is on RB-tree or
475                  * mem is over its softlimit.
476                  */
477                 if (excess || mz->on_tree) {
478                         spin_lock(&mctz->lock);
479                         /* if on-tree, remove it */
480                         if (mz->on_tree)
481                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
482                         /*
483                          * Insert again. mz->usage_in_excess will be updated.
484                          * If excess is 0, no tree ops.
485                          */
486                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
487                         spin_unlock(&mctz->lock);
488                 }
489         }
490 }
491
492 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
493 {
494         int node, zone;
495         struct mem_cgroup_per_zone *mz;
496         struct mem_cgroup_tree_per_zone *mctz;
497
498         for_each_node_state(node, N_POSSIBLE) {
499                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
500                         mz = mem_cgroup_zoneinfo(mem, node, zone);
501                         mctz = soft_limit_tree_node_zone(node, zone);
502                         mem_cgroup_remove_exceeded(mem, mz, mctz);
503                 }
504         }
505 }
506
507 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
508 {
509         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
510 }
511
512 static struct mem_cgroup_per_zone *
513 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
514 {
515         struct rb_node *rightmost = NULL;
516         struct mem_cgroup_per_zone *mz;
517
518 retry:
519         mz = NULL;
520         rightmost = rb_last(&mctz->rb_root);
521         if (!rightmost)
522                 goto done;              /* Nothing to reclaim from */
523
524         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
525         /*
526          * Remove the node now but someone else can add it back,
527          * we will to add it back at the end of reclaim to its correct
528          * position in the tree.
529          */
530         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
531         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
532                 !css_tryget(&mz->mem->css))
533                 goto retry;
534 done:
535         return mz;
536 }
537
538 static struct mem_cgroup_per_zone *
539 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
540 {
541         struct mem_cgroup_per_zone *mz;
542
543         spin_lock(&mctz->lock);
544         mz = __mem_cgroup_largest_soft_limit_node(mctz);
545         spin_unlock(&mctz->lock);
546         return mz;
547 }
548
549 /*
550  * Implementation Note: reading percpu statistics for memcg.
551  *
552  * Both of vmstat[] and percpu_counter has threshold and do periodic
553  * synchronization to implement "quick" read. There are trade-off between
554  * reading cost and precision of value. Then, we may have a chance to implement
555  * a periodic synchronizion of counter in memcg's counter.
556  *
557  * But this _read() function is used for user interface now. The user accounts
558  * memory usage by memory cgroup and he _always_ requires exact value because
559  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
560  * have to visit all online cpus and make sum. So, for now, unnecessary
561  * synchronization is not implemented. (just implemented for cpu hotplug)
562  *
563  * If there are kernel internal actions which can make use of some not-exact
564  * value, and reading all cpu value can be performance bottleneck in some
565  * common workload, threashold and synchonization as vmstat[] should be
566  * implemented.
567  */
568 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
569                 enum mem_cgroup_stat_index idx)
570 {
571         int cpu;
572         s64 val = 0;
573
574         get_online_cpus();
575         for_each_online_cpu(cpu)
576                 val += per_cpu(mem->stat->count[idx], cpu);
577 #ifdef CONFIG_HOTPLUG_CPU
578         spin_lock(&mem->pcp_counter_lock);
579         val += mem->nocpu_base.count[idx];
580         spin_unlock(&mem->pcp_counter_lock);
581 #endif
582         put_online_cpus();
583         return val;
584 }
585
586 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
587 {
588         s64 ret;
589
590         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
591         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
592         return ret;
593 }
594
595 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
596                                          bool charge)
597 {
598         int val = (charge) ? 1 : -1;
599         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
600 }
601
602 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
603                                          bool file, int nr_pages)
604 {
605         preempt_disable();
606
607         if (file)
608                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
609         else
610                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
611
612         /* pagein of a big page is an event. So, ignore page size */
613         if (nr_pages > 0)
614                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
615         else
616                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
617
618         __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages);
619
620         preempt_enable();
621 }
622
623 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
624                                         enum lru_list idx)
625 {
626         int nid, zid;
627         struct mem_cgroup_per_zone *mz;
628         u64 total = 0;
629
630         for_each_online_node(nid)
631                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
632                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
633                         total += MEM_CGROUP_ZSTAT(mz, idx);
634                 }
635         return total;
636 }
637
638 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
639 {
640         s64 val;
641
642         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
643
644         return !(val & ((1 << event_mask_shift) - 1));
645 }
646
647 /*
648  * Check events in order.
649  *
650  */
651 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
652 {
653         /* threshold event is triggered in finer grain than soft limit */
654         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
655                 mem_cgroup_threshold(mem);
656                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
657                         mem_cgroup_update_tree(mem, page);
658         }
659 }
660
661 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
662 {
663         return container_of(cgroup_subsys_state(cont,
664                                 mem_cgroup_subsys_id), struct mem_cgroup,
665                                 css);
666 }
667
668 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
669 {
670         /*
671          * mm_update_next_owner() may clear mm->owner to NULL
672          * if it races with swapoff, page migration, etc.
673          * So this can be called with p == NULL.
674          */
675         if (unlikely(!p))
676                 return NULL;
677
678         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
679                                 struct mem_cgroup, css);
680 }
681
682 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
683 {
684         struct mem_cgroup *mem = NULL;
685
686         if (!mm)
687                 return NULL;
688         /*
689          * Because we have no locks, mm->owner's may be being moved to other
690          * cgroup. We use css_tryget() here even if this looks
691          * pessimistic (rather than adding locks here).
692          */
693         rcu_read_lock();
694         do {
695                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
696                 if (unlikely(!mem))
697                         break;
698         } while (!css_tryget(&mem->css));
699         rcu_read_unlock();
700         return mem;
701 }
702
703 /* The caller has to guarantee "mem" exists before calling this */
704 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
705 {
706         struct cgroup_subsys_state *css;
707         int found;
708
709         if (!mem) /* ROOT cgroup has the smallest ID */
710                 return root_mem_cgroup; /*css_put/get against root is ignored*/
711         if (!mem->use_hierarchy) {
712                 if (css_tryget(&mem->css))
713                         return mem;
714                 return NULL;
715         }
716         rcu_read_lock();
717         /*
718          * searching a memory cgroup which has the smallest ID under given
719          * ROOT cgroup. (ID >= 1)
720          */
721         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
722         if (css && css_tryget(css))
723                 mem = container_of(css, struct mem_cgroup, css);
724         else
725                 mem = NULL;
726         rcu_read_unlock();
727         return mem;
728 }
729
730 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
731                                         struct mem_cgroup *root,
732                                         bool cond)
733 {
734         int nextid = css_id(&iter->css) + 1;
735         int found;
736         int hierarchy_used;
737         struct cgroup_subsys_state *css;
738
739         hierarchy_used = iter->use_hierarchy;
740
741         css_put(&iter->css);
742         /* If no ROOT, walk all, ignore hierarchy */
743         if (!cond || (root && !hierarchy_used))
744                 return NULL;
745
746         if (!root)
747                 root = root_mem_cgroup;
748
749         do {
750                 iter = NULL;
751                 rcu_read_lock();
752
753                 css = css_get_next(&mem_cgroup_subsys, nextid,
754                                 &root->css, &found);
755                 if (css && css_tryget(css))
756                         iter = container_of(css, struct mem_cgroup, css);
757                 rcu_read_unlock();
758                 /* If css is NULL, no more cgroups will be found */
759                 nextid = found + 1;
760         } while (css && !iter);
761
762         return iter;
763 }
764 /*
765  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
766  * be careful that "break" loop is not allowed. We have reference count.
767  * Instead of that modify "cond" to be false and "continue" to exit the loop.
768  */
769 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
770         for (iter = mem_cgroup_start_loop(root);\
771              iter != NULL;\
772              iter = mem_cgroup_get_next(iter, root, cond))
773
774 #define for_each_mem_cgroup_tree(iter, root) \
775         for_each_mem_cgroup_tree_cond(iter, root, true)
776
777 #define for_each_mem_cgroup_all(iter) \
778         for_each_mem_cgroup_tree_cond(iter, NULL, true)
779
780
781 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
782 {
783         return (mem == root_mem_cgroup);
784 }
785
786 /*
787  * Following LRU functions are allowed to be used without PCG_LOCK.
788  * Operations are called by routine of global LRU independently from memcg.
789  * What we have to take care of here is validness of pc->mem_cgroup.
790  *
791  * Changes to pc->mem_cgroup happens when
792  * 1. charge
793  * 2. moving account
794  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
795  * It is added to LRU before charge.
796  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
797  * When moving account, the page is not on LRU. It's isolated.
798  */
799
800 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
801 {
802         struct page_cgroup *pc;
803         struct mem_cgroup_per_zone *mz;
804
805         if (mem_cgroup_disabled())
806                 return;
807         pc = lookup_page_cgroup(page);
808         /* can happen while we handle swapcache. */
809         if (!TestClearPageCgroupAcctLRU(pc))
810                 return;
811         VM_BUG_ON(!pc->mem_cgroup);
812         /*
813          * We don't check PCG_USED bit. It's cleared when the "page" is finally
814          * removed from global LRU.
815          */
816         mz = page_cgroup_zoneinfo(pc);
817         /* huge page split is done under lru_lock. so, we have no races. */
818         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
819         if (mem_cgroup_is_root(pc->mem_cgroup))
820                 return;
821         VM_BUG_ON(list_empty(&pc->lru));
822         list_del_init(&pc->lru);
823 }
824
825 void mem_cgroup_del_lru(struct page *page)
826 {
827         mem_cgroup_del_lru_list(page, page_lru(page));
828 }
829
830 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
831 {
832         struct mem_cgroup_per_zone *mz;
833         struct page_cgroup *pc;
834
835         if (mem_cgroup_disabled())
836                 return;
837
838         pc = lookup_page_cgroup(page);
839         /* unused or root page is not rotated. */
840         if (!PageCgroupUsed(pc))
841                 return;
842         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
843         smp_rmb();
844         if (mem_cgroup_is_root(pc->mem_cgroup))
845                 return;
846         mz = page_cgroup_zoneinfo(pc);
847         list_move(&pc->lru, &mz->lists[lru]);
848 }
849
850 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
851 {
852         struct page_cgroup *pc;
853         struct mem_cgroup_per_zone *mz;
854
855         if (mem_cgroup_disabled())
856                 return;
857         pc = lookup_page_cgroup(page);
858         VM_BUG_ON(PageCgroupAcctLRU(pc));
859         if (!PageCgroupUsed(pc))
860                 return;
861         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
862         smp_rmb();
863         mz = page_cgroup_zoneinfo(pc);
864         /* huge page split is done under lru_lock. so, we have no races. */
865         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
866         SetPageCgroupAcctLRU(pc);
867         if (mem_cgroup_is_root(pc->mem_cgroup))
868                 return;
869         list_add(&pc->lru, &mz->lists[lru]);
870 }
871
872 /*
873  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
874  * lru because the page may.be reused after it's fully uncharged (because of
875  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
876  * it again. This function is only used to charge SwapCache. It's done under
877  * lock_page and expected that zone->lru_lock is never held.
878  */
879 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
880 {
881         unsigned long flags;
882         struct zone *zone = page_zone(page);
883         struct page_cgroup *pc = lookup_page_cgroup(page);
884
885         spin_lock_irqsave(&zone->lru_lock, flags);
886         /*
887          * Forget old LRU when this page_cgroup is *not* used. This Used bit
888          * is guarded by lock_page() because the page is SwapCache.
889          */
890         if (!PageCgroupUsed(pc))
891                 mem_cgroup_del_lru_list(page, page_lru(page));
892         spin_unlock_irqrestore(&zone->lru_lock, flags);
893 }
894
895 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
896 {
897         unsigned long flags;
898         struct zone *zone = page_zone(page);
899         struct page_cgroup *pc = lookup_page_cgroup(page);
900
901         spin_lock_irqsave(&zone->lru_lock, flags);
902         /* link when the page is linked to LRU but page_cgroup isn't */
903         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
904                 mem_cgroup_add_lru_list(page, page_lru(page));
905         spin_unlock_irqrestore(&zone->lru_lock, flags);
906 }
907
908
909 void mem_cgroup_move_lists(struct page *page,
910                            enum lru_list from, enum lru_list to)
911 {
912         if (mem_cgroup_disabled())
913                 return;
914         mem_cgroup_del_lru_list(page, from);
915         mem_cgroup_add_lru_list(page, to);
916 }
917
918 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
919 {
920         int ret;
921         struct mem_cgroup *curr = NULL;
922         struct task_struct *p;
923
924         p = find_lock_task_mm(task);
925         if (!p)
926                 return 0;
927         curr = try_get_mem_cgroup_from_mm(p->mm);
928         task_unlock(p);
929         if (!curr)
930                 return 0;
931         /*
932          * We should check use_hierarchy of "mem" not "curr". Because checking
933          * use_hierarchy of "curr" here make this function true if hierarchy is
934          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
935          * hierarchy(even if use_hierarchy is disabled in "mem").
936          */
937         if (mem->use_hierarchy)
938                 ret = css_is_ancestor(&curr->css, &mem->css);
939         else
940                 ret = (curr == mem);
941         css_put(&curr->css);
942         return ret;
943 }
944
945 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
946 {
947         unsigned long active;
948         unsigned long inactive;
949         unsigned long gb;
950         unsigned long inactive_ratio;
951
952         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
953         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
954
955         gb = (inactive + active) >> (30 - PAGE_SHIFT);
956         if (gb)
957                 inactive_ratio = int_sqrt(10 * gb);
958         else
959                 inactive_ratio = 1;
960
961         if (present_pages) {
962                 present_pages[0] = inactive;
963                 present_pages[1] = active;
964         }
965
966         return inactive_ratio;
967 }
968
969 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
970 {
971         unsigned long active;
972         unsigned long inactive;
973         unsigned long present_pages[2];
974         unsigned long inactive_ratio;
975
976         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
977
978         inactive = present_pages[0];
979         active = present_pages[1];
980
981         if (inactive * inactive_ratio < active)
982                 return 1;
983
984         return 0;
985 }
986
987 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
988 {
989         unsigned long active;
990         unsigned long inactive;
991
992         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
993         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
994
995         return (active > inactive);
996 }
997
998 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
999                                        struct zone *zone,
1000                                        enum lru_list lru)
1001 {
1002         int nid = zone_to_nid(zone);
1003         int zid = zone_idx(zone);
1004         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1005
1006         return MEM_CGROUP_ZSTAT(mz, lru);
1007 }
1008
1009 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1010                                                       struct zone *zone)
1011 {
1012         int nid = zone_to_nid(zone);
1013         int zid = zone_idx(zone);
1014         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1015
1016         return &mz->reclaim_stat;
1017 }
1018
1019 struct zone_reclaim_stat *
1020 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1021 {
1022         struct page_cgroup *pc;
1023         struct mem_cgroup_per_zone *mz;
1024
1025         if (mem_cgroup_disabled())
1026                 return NULL;
1027
1028         pc = lookup_page_cgroup(page);
1029         if (!PageCgroupUsed(pc))
1030                 return NULL;
1031         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1032         smp_rmb();
1033         mz = page_cgroup_zoneinfo(pc);
1034         if (!mz)
1035                 return NULL;
1036
1037         return &mz->reclaim_stat;
1038 }
1039
1040 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1041                                         struct list_head *dst,
1042                                         unsigned long *scanned, int order,
1043                                         int mode, struct zone *z,
1044                                         struct mem_cgroup *mem_cont,
1045                                         int active, int file)
1046 {
1047         unsigned long nr_taken = 0;
1048         struct page *page;
1049         unsigned long scan;
1050         LIST_HEAD(pc_list);
1051         struct list_head *src;
1052         struct page_cgroup *pc, *tmp;
1053         int nid = zone_to_nid(z);
1054         int zid = zone_idx(z);
1055         struct mem_cgroup_per_zone *mz;
1056         int lru = LRU_FILE * file + active;
1057         int ret;
1058
1059         BUG_ON(!mem_cont);
1060         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1061         src = &mz->lists[lru];
1062
1063         scan = 0;
1064         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1065                 if (scan >= nr_to_scan)
1066                         break;
1067
1068                 page = pc->page;
1069                 if (unlikely(!PageCgroupUsed(pc)))
1070                         continue;
1071                 if (unlikely(!PageLRU(page)))
1072                         continue;
1073
1074                 scan++;
1075                 ret = __isolate_lru_page(page, mode, file);
1076                 switch (ret) {
1077                 case 0:
1078                         list_move(&page->lru, dst);
1079                         mem_cgroup_del_lru(page);
1080                         nr_taken += hpage_nr_pages(page);
1081                         break;
1082                 case -EBUSY:
1083                         /* we don't affect global LRU but rotate in our LRU */
1084                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1085                         break;
1086                 default:
1087                         break;
1088                 }
1089         }
1090
1091         *scanned = scan;
1092
1093         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1094                                       0, 0, 0, mode);
1095
1096         return nr_taken;
1097 }
1098
1099 #define mem_cgroup_from_res_counter(counter, member)    \
1100         container_of(counter, struct mem_cgroup, member)
1101
1102 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1103 {
1104         if (do_swap_account) {
1105                 if (res_counter_check_under_limit(&mem->res) &&
1106                         res_counter_check_under_limit(&mem->memsw))
1107                         return true;
1108         } else
1109                 if (res_counter_check_under_limit(&mem->res))
1110                         return true;
1111         return false;
1112 }
1113
1114 /**
1115  * mem_cgroup_check_margin - check if the memory cgroup allows charging
1116  * @mem: memory cgroup to check
1117  * @bytes: the number of bytes the caller intends to charge
1118  *
1119  * Returns a boolean value on whether @mem can be charged @bytes or
1120  * whether this would exceed the limit.
1121  */
1122 static bool mem_cgroup_check_margin(struct mem_cgroup *mem, unsigned long bytes)
1123 {
1124         if (!res_counter_check_margin(&mem->res, bytes))
1125                 return false;
1126         if (do_swap_account && !res_counter_check_margin(&mem->memsw, bytes))
1127                 return false;
1128         return true;
1129 }
1130
1131 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1132 {
1133         struct cgroup *cgrp = memcg->css.cgroup;
1134         unsigned int swappiness;
1135
1136         /* root ? */
1137         if (cgrp->parent == NULL)
1138                 return vm_swappiness;
1139
1140         spin_lock(&memcg->reclaim_param_lock);
1141         swappiness = memcg->swappiness;
1142         spin_unlock(&memcg->reclaim_param_lock);
1143
1144         return swappiness;
1145 }
1146
1147 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1148 {
1149         int cpu;
1150
1151         get_online_cpus();
1152         spin_lock(&mem->pcp_counter_lock);
1153         for_each_online_cpu(cpu)
1154                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1155         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1156         spin_unlock(&mem->pcp_counter_lock);
1157         put_online_cpus();
1158
1159         synchronize_rcu();
1160 }
1161
1162 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1163 {
1164         int cpu;
1165
1166         if (!mem)
1167                 return;
1168         get_online_cpus();
1169         spin_lock(&mem->pcp_counter_lock);
1170         for_each_online_cpu(cpu)
1171                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1172         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1173         spin_unlock(&mem->pcp_counter_lock);
1174         put_online_cpus();
1175 }
1176 /*
1177  * 2 routines for checking "mem" is under move_account() or not.
1178  *
1179  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1180  *                        for avoiding race in accounting. If true,
1181  *                        pc->mem_cgroup may be overwritten.
1182  *
1183  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1184  *                        under hierarchy of moving cgroups. This is for
1185  *                        waiting at hith-memory prressure caused by "move".
1186  */
1187
1188 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1189 {
1190         VM_BUG_ON(!rcu_read_lock_held());
1191         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1192 }
1193
1194 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1195 {
1196         struct mem_cgroup *from;
1197         struct mem_cgroup *to;
1198         bool ret = false;
1199         /*
1200          * Unlike task_move routines, we access mc.to, mc.from not under
1201          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1202          */
1203         spin_lock(&mc.lock);
1204         from = mc.from;
1205         to = mc.to;
1206         if (!from)
1207                 goto unlock;
1208         if (from == mem || to == mem
1209             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1210             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1211                 ret = true;
1212 unlock:
1213         spin_unlock(&mc.lock);
1214         return ret;
1215 }
1216
1217 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1218 {
1219         if (mc.moving_task && current != mc.moving_task) {
1220                 if (mem_cgroup_under_move(mem)) {
1221                         DEFINE_WAIT(wait);
1222                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1223                         /* moving charge context might have finished. */
1224                         if (mc.moving_task)
1225                                 schedule();
1226                         finish_wait(&mc.waitq, &wait);
1227                         return true;
1228                 }
1229         }
1230         return false;
1231 }
1232
1233 /**
1234  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1235  * @memcg: The memory cgroup that went over limit
1236  * @p: Task that is going to be killed
1237  *
1238  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1239  * enabled
1240  */
1241 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1242 {
1243         struct cgroup *task_cgrp;
1244         struct cgroup *mem_cgrp;
1245         /*
1246          * Need a buffer in BSS, can't rely on allocations. The code relies
1247          * on the assumption that OOM is serialized for memory controller.
1248          * If this assumption is broken, revisit this code.
1249          */
1250         static char memcg_name[PATH_MAX];
1251         int ret;
1252
1253         if (!memcg || !p)
1254                 return;
1255
1256
1257         rcu_read_lock();
1258
1259         mem_cgrp = memcg->css.cgroup;
1260         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1261
1262         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1263         if (ret < 0) {
1264                 /*
1265                  * Unfortunately, we are unable to convert to a useful name
1266                  * But we'll still print out the usage information
1267                  */
1268                 rcu_read_unlock();
1269                 goto done;
1270         }
1271         rcu_read_unlock();
1272
1273         printk(KERN_INFO "Task in %s killed", memcg_name);
1274
1275         rcu_read_lock();
1276         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1277         if (ret < 0) {
1278                 rcu_read_unlock();
1279                 goto done;
1280         }
1281         rcu_read_unlock();
1282
1283         /*
1284          * Continues from above, so we don't need an KERN_ level
1285          */
1286         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1287 done:
1288
1289         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1290                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1291                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1292                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1293         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1294                 "failcnt %llu\n",
1295                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1296                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1297                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1298 }
1299
1300 /*
1301  * This function returns the number of memcg under hierarchy tree. Returns
1302  * 1(self count) if no children.
1303  */
1304 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1305 {
1306         int num = 0;
1307         struct mem_cgroup *iter;
1308
1309         for_each_mem_cgroup_tree(iter, mem)
1310                 num++;
1311         return num;
1312 }
1313
1314 /*
1315  * Return the memory (and swap, if configured) limit for a memcg.
1316  */
1317 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1318 {
1319         u64 limit;
1320         u64 memsw;
1321
1322         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1323         limit += total_swap_pages << PAGE_SHIFT;
1324
1325         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1326         /*
1327          * If memsw is finite and limits the amount of swap space available
1328          * to this memcg, return that limit.
1329          */
1330         return min(limit, memsw);
1331 }
1332
1333 /*
1334  * Visit the first child (need not be the first child as per the ordering
1335  * of the cgroup list, since we track last_scanned_child) of @mem and use
1336  * that to reclaim free pages from.
1337  */
1338 static struct mem_cgroup *
1339 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1340 {
1341         struct mem_cgroup *ret = NULL;
1342         struct cgroup_subsys_state *css;
1343         int nextid, found;
1344
1345         if (!root_mem->use_hierarchy) {
1346                 css_get(&root_mem->css);
1347                 ret = root_mem;
1348         }
1349
1350         while (!ret) {
1351                 rcu_read_lock();
1352                 nextid = root_mem->last_scanned_child + 1;
1353                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1354                                    &found);
1355                 if (css && css_tryget(css))
1356                         ret = container_of(css, struct mem_cgroup, css);
1357
1358                 rcu_read_unlock();
1359                 /* Updates scanning parameter */
1360                 spin_lock(&root_mem->reclaim_param_lock);
1361                 if (!css) {
1362                         /* this means start scan from ID:1 */
1363                         root_mem->last_scanned_child = 0;
1364                 } else
1365                         root_mem->last_scanned_child = found;
1366                 spin_unlock(&root_mem->reclaim_param_lock);
1367         }
1368
1369         return ret;
1370 }
1371
1372 /*
1373  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1374  * we reclaimed from, so that we don't end up penalizing one child extensively
1375  * based on its position in the children list.
1376  *
1377  * root_mem is the original ancestor that we've been reclaim from.
1378  *
1379  * We give up and return to the caller when we visit root_mem twice.
1380  * (other groups can be removed while we're walking....)
1381  *
1382  * If shrink==true, for avoiding to free too much, this returns immedieately.
1383  */
1384 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1385                                                 struct zone *zone,
1386                                                 gfp_t gfp_mask,
1387                                                 unsigned long reclaim_options)
1388 {
1389         struct mem_cgroup *victim;
1390         int ret, total = 0;
1391         int loop = 0;
1392         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1393         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1394         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1395         unsigned long excess = mem_cgroup_get_excess(root_mem);
1396
1397         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1398         if (root_mem->memsw_is_minimum)
1399                 noswap = true;
1400
1401         while (1) {
1402                 victim = mem_cgroup_select_victim(root_mem);
1403                 if (victim == root_mem) {
1404                         loop++;
1405                         if (loop >= 1)
1406                                 drain_all_stock_async();
1407                         if (loop >= 2) {
1408                                 /*
1409                                  * If we have not been able to reclaim
1410                                  * anything, it might because there are
1411                                  * no reclaimable pages under this hierarchy
1412                                  */
1413                                 if (!check_soft || !total) {
1414                                         css_put(&victim->css);
1415                                         break;
1416                                 }
1417                                 /*
1418                                  * We want to do more targetted reclaim.
1419                                  * excess >> 2 is not to excessive so as to
1420                                  * reclaim too much, nor too less that we keep
1421                                  * coming back to reclaim from this cgroup
1422                                  */
1423                                 if (total >= (excess >> 2) ||
1424                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1425                                         css_put(&victim->css);
1426                                         break;
1427                                 }
1428                         }
1429                 }
1430                 if (!mem_cgroup_local_usage(victim)) {
1431                         /* this cgroup's local usage == 0 */
1432                         css_put(&victim->css);
1433                         continue;
1434                 }
1435                 /* we use swappiness of local cgroup */
1436                 if (check_soft)
1437                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1438                                 noswap, get_swappiness(victim), zone);
1439                 else
1440                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1441                                                 noswap, get_swappiness(victim));
1442                 css_put(&victim->css);
1443                 /*
1444                  * At shrinking usage, we can't check we should stop here or
1445                  * reclaim more. It's depends on callers. last_scanned_child
1446                  * will work enough for keeping fairness under tree.
1447                  */
1448                 if (shrink)
1449                         return ret;
1450                 total += ret;
1451                 if (check_soft) {
1452                         if (res_counter_check_under_soft_limit(&root_mem->res))
1453                                 return total;
1454                 } else if (mem_cgroup_check_under_limit(root_mem))
1455                         return 1 + total;
1456         }
1457         return total;
1458 }
1459
1460 /*
1461  * Check OOM-Killer is already running under our hierarchy.
1462  * If someone is running, return false.
1463  */
1464 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1465 {
1466         int x, lock_count = 0;
1467         struct mem_cgroup *iter;
1468
1469         for_each_mem_cgroup_tree(iter, mem) {
1470                 x = atomic_inc_return(&iter->oom_lock);
1471                 lock_count = max(x, lock_count);
1472         }
1473
1474         if (lock_count == 1)
1475                 return true;
1476         return false;
1477 }
1478
1479 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1480 {
1481         struct mem_cgroup *iter;
1482
1483         /*
1484          * When a new child is created while the hierarchy is under oom,
1485          * mem_cgroup_oom_lock() may not be called. We have to use
1486          * atomic_add_unless() here.
1487          */
1488         for_each_mem_cgroup_tree(iter, mem)
1489                 atomic_add_unless(&iter->oom_lock, -1, 0);
1490         return 0;
1491 }
1492
1493
1494 static DEFINE_MUTEX(memcg_oom_mutex);
1495 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1496
1497 struct oom_wait_info {
1498         struct mem_cgroup *mem;
1499         wait_queue_t    wait;
1500 };
1501
1502 static int memcg_oom_wake_function(wait_queue_t *wait,
1503         unsigned mode, int sync, void *arg)
1504 {
1505         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1506         struct oom_wait_info *oom_wait_info;
1507
1508         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1509
1510         if (oom_wait_info->mem == wake_mem)
1511                 goto wakeup;
1512         /* if no hierarchy, no match */
1513         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1514                 return 0;
1515         /*
1516          * Both of oom_wait_info->mem and wake_mem are stable under us.
1517          * Then we can use css_is_ancestor without taking care of RCU.
1518          */
1519         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1520             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1521                 return 0;
1522
1523 wakeup:
1524         return autoremove_wake_function(wait, mode, sync, arg);
1525 }
1526
1527 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1528 {
1529         /* for filtering, pass "mem" as argument. */
1530         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1531 }
1532
1533 static void memcg_oom_recover(struct mem_cgroup *mem)
1534 {
1535         if (mem && atomic_read(&mem->oom_lock))
1536                 memcg_wakeup_oom(mem);
1537 }
1538
1539 /*
1540  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1541  */
1542 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1543 {
1544         struct oom_wait_info owait;
1545         bool locked, need_to_kill;
1546
1547         owait.mem = mem;
1548         owait.wait.flags = 0;
1549         owait.wait.func = memcg_oom_wake_function;
1550         owait.wait.private = current;
1551         INIT_LIST_HEAD(&owait.wait.task_list);
1552         need_to_kill = true;
1553         /* At first, try to OOM lock hierarchy under mem.*/
1554         mutex_lock(&memcg_oom_mutex);
1555         locked = mem_cgroup_oom_lock(mem);
1556         /*
1557          * Even if signal_pending(), we can't quit charge() loop without
1558          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1559          * under OOM is always welcomed, use TASK_KILLABLE here.
1560          */
1561         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1562         if (!locked || mem->oom_kill_disable)
1563                 need_to_kill = false;
1564         if (locked)
1565                 mem_cgroup_oom_notify(mem);
1566         mutex_unlock(&memcg_oom_mutex);
1567
1568         if (need_to_kill) {
1569                 finish_wait(&memcg_oom_waitq, &owait.wait);
1570                 mem_cgroup_out_of_memory(mem, mask);
1571         } else {
1572                 schedule();
1573                 finish_wait(&memcg_oom_waitq, &owait.wait);
1574         }
1575         mutex_lock(&memcg_oom_mutex);
1576         mem_cgroup_oom_unlock(mem);
1577         memcg_wakeup_oom(mem);
1578         mutex_unlock(&memcg_oom_mutex);
1579
1580         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1581                 return false;
1582         /* Give chance to dying process */
1583         schedule_timeout(1);
1584         return true;
1585 }
1586
1587 /*
1588  * Currently used to update mapped file statistics, but the routine can be
1589  * generalized to update other statistics as well.
1590  *
1591  * Notes: Race condition
1592  *
1593  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1594  * it tends to be costly. But considering some conditions, we doesn't need
1595  * to do so _always_.
1596  *
1597  * Considering "charge", lock_page_cgroup() is not required because all
1598  * file-stat operations happen after a page is attached to radix-tree. There
1599  * are no race with "charge".
1600  *
1601  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1602  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1603  * if there are race with "uncharge". Statistics itself is properly handled
1604  * by flags.
1605  *
1606  * Considering "move", this is an only case we see a race. To make the race
1607  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1608  * possibility of race condition. If there is, we take a lock.
1609  */
1610
1611 void mem_cgroup_update_page_stat(struct page *page,
1612                                  enum mem_cgroup_page_stat_item idx, int val)
1613 {
1614         struct mem_cgroup *mem;
1615         struct page_cgroup *pc = lookup_page_cgroup(page);
1616         bool need_unlock = false;
1617         unsigned long uninitialized_var(flags);
1618
1619         if (unlikely(!pc))
1620                 return;
1621
1622         rcu_read_lock();
1623         mem = pc->mem_cgroup;
1624         if (unlikely(!mem || !PageCgroupUsed(pc)))
1625                 goto out;
1626         /* pc->mem_cgroup is unstable ? */
1627         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1628                 /* take a lock against to access pc->mem_cgroup */
1629                 move_lock_page_cgroup(pc, &flags);
1630                 need_unlock = true;
1631                 mem = pc->mem_cgroup;
1632                 if (!mem || !PageCgroupUsed(pc))
1633                         goto out;
1634         }
1635
1636         switch (idx) {
1637         case MEMCG_NR_FILE_MAPPED:
1638                 if (val > 0)
1639                         SetPageCgroupFileMapped(pc);
1640                 else if (!page_mapped(page))
1641                         ClearPageCgroupFileMapped(pc);
1642                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1643                 break;
1644         default:
1645                 BUG();
1646         }
1647
1648         this_cpu_add(mem->stat->count[idx], val);
1649
1650 out:
1651         if (unlikely(need_unlock))
1652                 move_unlock_page_cgroup(pc, &flags);
1653         rcu_read_unlock();
1654         return;
1655 }
1656 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1657
1658 /*
1659  * size of first charge trial. "32" comes from vmscan.c's magic value.
1660  * TODO: maybe necessary to use big numbers in big irons.
1661  */
1662 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1663 struct memcg_stock_pcp {
1664         struct mem_cgroup *cached; /* this never be root cgroup */
1665         int charge;
1666         struct work_struct work;
1667 };
1668 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1669 static atomic_t memcg_drain_count;
1670
1671 /*
1672  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1673  * from local stock and true is returned. If the stock is 0 or charges from a
1674  * cgroup which is not current target, returns false. This stock will be
1675  * refilled.
1676  */
1677 static bool consume_stock(struct mem_cgroup *mem)
1678 {
1679         struct memcg_stock_pcp *stock;
1680         bool ret = true;
1681
1682         stock = &get_cpu_var(memcg_stock);
1683         if (mem == stock->cached && stock->charge)
1684                 stock->charge -= PAGE_SIZE;
1685         else /* need to call res_counter_charge */
1686                 ret = false;
1687         put_cpu_var(memcg_stock);
1688         return ret;
1689 }
1690
1691 /*
1692  * Returns stocks cached in percpu to res_counter and reset cached information.
1693  */
1694 static void drain_stock(struct memcg_stock_pcp *stock)
1695 {
1696         struct mem_cgroup *old = stock->cached;
1697
1698         if (stock->charge) {
1699                 res_counter_uncharge(&old->res, stock->charge);
1700                 if (do_swap_account)
1701                         res_counter_uncharge(&old->memsw, stock->charge);
1702         }
1703         stock->cached = NULL;
1704         stock->charge = 0;
1705 }
1706
1707 /*
1708  * This must be called under preempt disabled or must be called by
1709  * a thread which is pinned to local cpu.
1710  */
1711 static void drain_local_stock(struct work_struct *dummy)
1712 {
1713         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1714         drain_stock(stock);
1715 }
1716
1717 /*
1718  * Cache charges(val) which is from res_counter, to local per_cpu area.
1719  * This will be consumed by consume_stock() function, later.
1720  */
1721 static void refill_stock(struct mem_cgroup *mem, int val)
1722 {
1723         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1724
1725         if (stock->cached != mem) { /* reset if necessary */
1726                 drain_stock(stock);
1727                 stock->cached = mem;
1728         }
1729         stock->charge += val;
1730         put_cpu_var(memcg_stock);
1731 }
1732
1733 /*
1734  * Tries to drain stocked charges in other cpus. This function is asynchronous
1735  * and just put a work per cpu for draining localy on each cpu. Caller can
1736  * expects some charges will be back to res_counter later but cannot wait for
1737  * it.
1738  */
1739 static void drain_all_stock_async(void)
1740 {
1741         int cpu;
1742         /* This function is for scheduling "drain" in asynchronous way.
1743          * The result of "drain" is not directly handled by callers. Then,
1744          * if someone is calling drain, we don't have to call drain more.
1745          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1746          * there is a race. We just do loose check here.
1747          */
1748         if (atomic_read(&memcg_drain_count))
1749                 return;
1750         /* Notify other cpus that system-wide "drain" is running */
1751         atomic_inc(&memcg_drain_count);
1752         get_online_cpus();
1753         for_each_online_cpu(cpu) {
1754                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1755                 schedule_work_on(cpu, &stock->work);
1756         }
1757         put_online_cpus();
1758         atomic_dec(&memcg_drain_count);
1759         /* We don't wait for flush_work */
1760 }
1761
1762 /* This is a synchronous drain interface. */
1763 static void drain_all_stock_sync(void)
1764 {
1765         /* called when force_empty is called */
1766         atomic_inc(&memcg_drain_count);
1767         schedule_on_each_cpu(drain_local_stock);
1768         atomic_dec(&memcg_drain_count);
1769 }
1770
1771 /*
1772  * This function drains percpu counter value from DEAD cpu and
1773  * move it to local cpu. Note that this function can be preempted.
1774  */
1775 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1776 {
1777         int i;
1778
1779         spin_lock(&mem->pcp_counter_lock);
1780         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1781                 s64 x = per_cpu(mem->stat->count[i], cpu);
1782
1783                 per_cpu(mem->stat->count[i], cpu) = 0;
1784                 mem->nocpu_base.count[i] += x;
1785         }
1786         /* need to clear ON_MOVE value, works as a kind of lock. */
1787         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1788         spin_unlock(&mem->pcp_counter_lock);
1789 }
1790
1791 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1792 {
1793         int idx = MEM_CGROUP_ON_MOVE;
1794
1795         spin_lock(&mem->pcp_counter_lock);
1796         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1797         spin_unlock(&mem->pcp_counter_lock);
1798 }
1799
1800 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1801                                         unsigned long action,
1802                                         void *hcpu)
1803 {
1804         int cpu = (unsigned long)hcpu;
1805         struct memcg_stock_pcp *stock;
1806         struct mem_cgroup *iter;
1807
1808         if ((action == CPU_ONLINE)) {
1809                 for_each_mem_cgroup_all(iter)
1810                         synchronize_mem_cgroup_on_move(iter, cpu);
1811                 return NOTIFY_OK;
1812         }
1813
1814         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1815                 return NOTIFY_OK;
1816
1817         for_each_mem_cgroup_all(iter)
1818                 mem_cgroup_drain_pcp_counter(iter, cpu);
1819
1820         stock = &per_cpu(memcg_stock, cpu);
1821         drain_stock(stock);
1822         return NOTIFY_OK;
1823 }
1824
1825
1826 /* See __mem_cgroup_try_charge() for details */
1827 enum {
1828         CHARGE_OK,              /* success */
1829         CHARGE_RETRY,           /* need to retry but retry is not bad */
1830         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1831         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1832         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1833 };
1834
1835 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1836                                 int csize, bool oom_check)
1837 {
1838         struct mem_cgroup *mem_over_limit;
1839         struct res_counter *fail_res;
1840         unsigned long flags = 0;
1841         int ret;
1842
1843         ret = res_counter_charge(&mem->res, csize, &fail_res);
1844
1845         if (likely(!ret)) {
1846                 if (!do_swap_account)
1847                         return CHARGE_OK;
1848                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1849                 if (likely(!ret))
1850                         return CHARGE_OK;
1851
1852                 res_counter_uncharge(&mem->res, csize);
1853                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1854                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1855         } else
1856                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1857         /*
1858          * csize can be either a huge page (HPAGE_SIZE), a batch of
1859          * regular pages (CHARGE_SIZE), or a single regular page
1860          * (PAGE_SIZE).
1861          *
1862          * Never reclaim on behalf of optional batching, retry with a
1863          * single page instead.
1864          */
1865         if (csize == CHARGE_SIZE)
1866                 return CHARGE_RETRY;
1867
1868         if (!(gfp_mask & __GFP_WAIT))
1869                 return CHARGE_WOULDBLOCK;
1870
1871         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1872                                               gfp_mask, flags);
1873         if (mem_cgroup_check_margin(mem_over_limit, csize))
1874                 return CHARGE_RETRY;
1875         /*
1876          * Even though the limit is exceeded at this point, reclaim
1877          * may have been able to free some pages.  Retry the charge
1878          * before killing the task.
1879          *
1880          * Only for regular pages, though: huge pages are rather
1881          * unlikely to succeed so close to the limit, and we fall back
1882          * to regular pages anyway in case of failure.
1883          */
1884         if (csize == PAGE_SIZE && ret)
1885                 return CHARGE_RETRY;
1886
1887         /*
1888          * At task move, charge accounts can be doubly counted. So, it's
1889          * better to wait until the end of task_move if something is going on.
1890          */
1891         if (mem_cgroup_wait_acct_move(mem_over_limit))
1892                 return CHARGE_RETRY;
1893
1894         /* If we don't need to call oom-killer at el, return immediately */
1895         if (!oom_check)
1896                 return CHARGE_NOMEM;
1897         /* check OOM */
1898         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1899                 return CHARGE_OOM_DIE;
1900
1901         return CHARGE_RETRY;
1902 }
1903
1904 /*
1905  * Unlike exported interface, "oom" parameter is added. if oom==true,
1906  * oom-killer can be invoked.
1907  */
1908 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1909                                    gfp_t gfp_mask,
1910                                    struct mem_cgroup **memcg, bool oom,
1911                                    int page_size)
1912 {
1913         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1914         struct mem_cgroup *mem = NULL;
1915         int ret;
1916         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1917
1918         /*
1919          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1920          * in system level. So, allow to go ahead dying process in addition to
1921          * MEMDIE process.
1922          */
1923         if (unlikely(test_thread_flag(TIF_MEMDIE)
1924                      || fatal_signal_pending(current)))
1925                 goto bypass;
1926
1927         /*
1928          * We always charge the cgroup the mm_struct belongs to.
1929          * The mm_struct's mem_cgroup changes on task migration if the
1930          * thread group leader migrates. It's possible that mm is not
1931          * set, if so charge the init_mm (happens for pagecache usage).
1932          */
1933         if (!*memcg && !mm)
1934                 goto bypass;
1935 again:
1936         if (*memcg) { /* css should be a valid one */
1937                 mem = *memcg;
1938                 VM_BUG_ON(css_is_removed(&mem->css));
1939                 if (mem_cgroup_is_root(mem))
1940                         goto done;
1941                 if (page_size == PAGE_SIZE && consume_stock(mem))
1942                         goto done;
1943                 css_get(&mem->css);
1944         } else {
1945                 struct task_struct *p;
1946
1947                 rcu_read_lock();
1948                 p = rcu_dereference(mm->owner);
1949                 /*
1950                  * Because we don't have task_lock(), "p" can exit.
1951                  * In that case, "mem" can point to root or p can be NULL with
1952                  * race with swapoff. Then, we have small risk of mis-accouning.
1953                  * But such kind of mis-account by race always happens because
1954                  * we don't have cgroup_mutex(). It's overkill and we allo that
1955                  * small race, here.
1956                  * (*) swapoff at el will charge against mm-struct not against
1957                  * task-struct. So, mm->owner can be NULL.
1958                  */
1959                 mem = mem_cgroup_from_task(p);
1960                 if (!mem || mem_cgroup_is_root(mem)) {
1961                         rcu_read_unlock();
1962                         goto done;
1963                 }
1964                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1965                         /*
1966                          * It seems dagerous to access memcg without css_get().
1967                          * But considering how consume_stok works, it's not
1968                          * necessary. If consume_stock success, some charges
1969                          * from this memcg are cached on this cpu. So, we
1970                          * don't need to call css_get()/css_tryget() before
1971                          * calling consume_stock().
1972                          */
1973                         rcu_read_unlock();
1974                         goto done;
1975                 }
1976                 /* after here, we may be blocked. we need to get refcnt */
1977                 if (!css_tryget(&mem->css)) {
1978                         rcu_read_unlock();
1979                         goto again;
1980                 }
1981                 rcu_read_unlock();
1982         }
1983
1984         do {
1985                 bool oom_check;
1986
1987                 /* If killed, bypass charge */
1988                 if (fatal_signal_pending(current)) {
1989                         css_put(&mem->css);
1990                         goto bypass;
1991                 }
1992
1993                 oom_check = false;
1994                 if (oom && !nr_oom_retries) {
1995                         oom_check = true;
1996                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1997                 }
1998
1999                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
2000
2001                 switch (ret) {
2002                 case CHARGE_OK:
2003                         break;
2004                 case CHARGE_RETRY: /* not in OOM situation but retry */
2005                         csize = page_size;
2006                         css_put(&mem->css);
2007                         mem = NULL;
2008                         goto again;
2009                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2010                         css_put(&mem->css);
2011                         goto nomem;
2012                 case CHARGE_NOMEM: /* OOM routine works */
2013                         if (!oom) {
2014                                 css_put(&mem->css);
2015                                 goto nomem;
2016                         }
2017                         /* If oom, we never return -ENOMEM */
2018                         nr_oom_retries--;
2019                         break;
2020                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2021                         css_put(&mem->css);
2022                         goto bypass;
2023                 }
2024         } while (ret != CHARGE_OK);
2025
2026         if (csize > page_size)
2027                 refill_stock(mem, csize - page_size);
2028         css_put(&mem->css);
2029 done:
2030         *memcg = mem;
2031         return 0;
2032 nomem:
2033         *memcg = NULL;
2034         return -ENOMEM;
2035 bypass:
2036         *memcg = NULL;
2037         return 0;
2038 }
2039
2040 /*
2041  * Somemtimes we have to undo a charge we got by try_charge().
2042  * This function is for that and do uncharge, put css's refcnt.
2043  * gotten by try_charge().
2044  */
2045 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2046                                                         unsigned long count)
2047 {
2048         if (!mem_cgroup_is_root(mem)) {
2049                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2050                 if (do_swap_account)
2051                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2052         }
2053 }
2054
2055 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2056                                      int page_size)
2057 {
2058         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2059 }
2060
2061 /*
2062  * A helper function to get mem_cgroup from ID. must be called under
2063  * rcu_read_lock(). The caller must check css_is_removed() or some if
2064  * it's concern. (dropping refcnt from swap can be called against removed
2065  * memcg.)
2066  */
2067 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2068 {
2069         struct cgroup_subsys_state *css;
2070
2071         /* ID 0 is unused ID */
2072         if (!id)
2073                 return NULL;
2074         css = css_lookup(&mem_cgroup_subsys, id);
2075         if (!css)
2076                 return NULL;
2077         return container_of(css, struct mem_cgroup, css);
2078 }
2079
2080 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2081 {
2082         struct mem_cgroup *mem = NULL;
2083         struct page_cgroup *pc;
2084         unsigned short id;
2085         swp_entry_t ent;
2086
2087         VM_BUG_ON(!PageLocked(page));
2088
2089         pc = lookup_page_cgroup(page);
2090         lock_page_cgroup(pc);
2091         if (PageCgroupUsed(pc)) {
2092                 mem = pc->mem_cgroup;
2093                 if (mem && !css_tryget(&mem->css))
2094                         mem = NULL;
2095         } else if (PageSwapCache(page)) {
2096                 ent.val = page_private(page);
2097                 id = lookup_swap_cgroup(ent);
2098                 rcu_read_lock();
2099                 mem = mem_cgroup_lookup(id);
2100                 if (mem && !css_tryget(&mem->css))
2101                         mem = NULL;
2102                 rcu_read_unlock();
2103         }
2104         unlock_page_cgroup(pc);
2105         return mem;
2106 }
2107
2108 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2109                                        struct page_cgroup *pc,
2110                                        enum charge_type ctype,
2111                                        int page_size)
2112 {
2113         int nr_pages = page_size >> PAGE_SHIFT;
2114
2115         /* try_charge() can return NULL to *memcg, taking care of it. */
2116         if (!mem)
2117                 return;
2118
2119         lock_page_cgroup(pc);
2120         if (unlikely(PageCgroupUsed(pc))) {
2121                 unlock_page_cgroup(pc);
2122                 mem_cgroup_cancel_charge(mem, page_size);
2123                 return;
2124         }
2125         /*
2126          * we don't need page_cgroup_lock about tail pages, becase they are not
2127          * accessed by any other context at this point.
2128          */
2129         pc->mem_cgroup = mem;
2130         /*
2131          * We access a page_cgroup asynchronously without lock_page_cgroup().
2132          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2133          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2134          * before USED bit, we need memory barrier here.
2135          * See mem_cgroup_add_lru_list(), etc.
2136          */
2137         smp_wmb();
2138         switch (ctype) {
2139         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2140         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2141                 SetPageCgroupCache(pc);
2142                 SetPageCgroupUsed(pc);
2143                 break;
2144         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2145                 ClearPageCgroupCache(pc);
2146                 SetPageCgroupUsed(pc);
2147                 break;
2148         default:
2149                 break;
2150         }
2151
2152         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2153         unlock_page_cgroup(pc);
2154         /*
2155          * "charge_statistics" updated event counter. Then, check it.
2156          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2157          * if they exceeds softlimit.
2158          */
2159         memcg_check_events(mem, pc->page);
2160 }
2161
2162 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2163
2164 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2165                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2166 /*
2167  * Because tail pages are not marked as "used", set it. We're under
2168  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2169  */
2170 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2171 {
2172         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2173         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2174         unsigned long flags;
2175
2176         if (mem_cgroup_disabled())
2177                 return;
2178         /*
2179          * We have no races with charge/uncharge but will have races with
2180          * page state accounting.
2181          */
2182         move_lock_page_cgroup(head_pc, &flags);
2183
2184         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2185         smp_wmb(); /* see __commit_charge() */
2186         if (PageCgroupAcctLRU(head_pc)) {
2187                 enum lru_list lru;
2188                 struct mem_cgroup_per_zone *mz;
2189
2190                 /*
2191                  * LRU flags cannot be copied because we need to add tail
2192                  *.page to LRU by generic call and our hook will be called.
2193                  * We hold lru_lock, then, reduce counter directly.
2194                  */
2195                 lru = page_lru(head);
2196                 mz = page_cgroup_zoneinfo(head_pc);
2197                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2198         }
2199         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2200         move_unlock_page_cgroup(head_pc, &flags);
2201 }
2202 #endif
2203
2204 /**
2205  * __mem_cgroup_move_account - move account of the page
2206  * @pc: page_cgroup of the page.
2207  * @from: mem_cgroup which the page is moved from.
2208  * @to: mem_cgroup which the page is moved to. @from != @to.
2209  * @uncharge: whether we should call uncharge and css_put against @from.
2210  *
2211  * The caller must confirm following.
2212  * - page is not on LRU (isolate_page() is useful.)
2213  * - the pc is locked, used, and ->mem_cgroup points to @from.
2214  *
2215  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2216  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2217  * true, this function does "uncharge" from old cgroup, but it doesn't if
2218  * @uncharge is false, so a caller should do "uncharge".
2219  */
2220
2221 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2222         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge,
2223         int charge_size)
2224 {
2225         int nr_pages = charge_size >> PAGE_SHIFT;
2226
2227         VM_BUG_ON(from == to);
2228         VM_BUG_ON(PageLRU(pc->page));
2229         VM_BUG_ON(!page_is_cgroup_locked(pc));
2230         VM_BUG_ON(!PageCgroupUsed(pc));
2231         VM_BUG_ON(pc->mem_cgroup != from);
2232
2233         if (PageCgroupFileMapped(pc)) {
2234                 /* Update mapped_file data for mem_cgroup */
2235                 preempt_disable();
2236                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2237                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2238                 preempt_enable();
2239         }
2240         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2241         if (uncharge)
2242                 /* This is not "cancel", but cancel_charge does all we need. */
2243                 mem_cgroup_cancel_charge(from, charge_size);
2244
2245         /* caller should have done css_get */
2246         pc->mem_cgroup = to;
2247         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2248         /*
2249          * We charges against "to" which may not have any tasks. Then, "to"
2250          * can be under rmdir(). But in current implementation, caller of
2251          * this function is just force_empty() and move charge, so it's
2252          * garanteed that "to" is never removed. So, we don't check rmdir
2253          * status here.
2254          */
2255 }
2256
2257 /*
2258  * check whether the @pc is valid for moving account and call
2259  * __mem_cgroup_move_account()
2260  */
2261 static int mem_cgroup_move_account(struct page_cgroup *pc,
2262                 struct mem_cgroup *from, struct mem_cgroup *to,
2263                 bool uncharge, int charge_size)
2264 {
2265         int ret = -EINVAL;
2266         unsigned long flags;
2267         /*
2268          * The page is isolated from LRU. So, collapse function
2269          * will not handle this page. But page splitting can happen.
2270          * Do this check under compound_page_lock(). The caller should
2271          * hold it.
2272          */
2273         if ((charge_size > PAGE_SIZE) && !PageTransHuge(pc->page))
2274                 return -EBUSY;
2275
2276         lock_page_cgroup(pc);
2277         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2278                 move_lock_page_cgroup(pc, &flags);
2279                 __mem_cgroup_move_account(pc, from, to, uncharge, charge_size);
2280                 move_unlock_page_cgroup(pc, &flags);
2281                 ret = 0;
2282         }
2283         unlock_page_cgroup(pc);
2284         /*
2285          * check events
2286          */
2287         memcg_check_events(to, pc->page);
2288         memcg_check_events(from, pc->page);
2289         return ret;
2290 }
2291
2292 /*
2293  * move charges to its parent.
2294  */
2295
2296 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2297                                   struct mem_cgroup *child,
2298                                   gfp_t gfp_mask)
2299 {
2300         struct page *page = pc->page;
2301         struct cgroup *cg = child->css.cgroup;
2302         struct cgroup *pcg = cg->parent;
2303         struct mem_cgroup *parent;
2304         int page_size = PAGE_SIZE;
2305         unsigned long flags;
2306         int ret;
2307
2308         /* Is ROOT ? */
2309         if (!pcg)
2310                 return -EINVAL;
2311
2312         ret = -EBUSY;
2313         if (!get_page_unless_zero(page))
2314                 goto out;
2315         if (isolate_lru_page(page))
2316                 goto put;
2317
2318         if (PageTransHuge(page))
2319                 page_size = HPAGE_SIZE;
2320
2321         parent = mem_cgroup_from_cont(pcg);
2322         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2323                                 &parent, false, page_size);
2324         if (ret || !parent)
2325                 goto put_back;
2326
2327         if (page_size > PAGE_SIZE)
2328                 flags = compound_lock_irqsave(page);
2329
2330         ret = mem_cgroup_move_account(pc, child, parent, true, page_size);
2331         if (ret)
2332                 mem_cgroup_cancel_charge(parent, page_size);
2333
2334         if (page_size > PAGE_SIZE)
2335                 compound_unlock_irqrestore(page, flags);
2336 put_back:
2337         putback_lru_page(page);
2338 put:
2339         put_page(page);
2340 out:
2341         return ret;
2342 }
2343
2344 /*
2345  * Charge the memory controller for page usage.
2346  * Return
2347  * 0 if the charge was successful
2348  * < 0 if the cgroup is over its limit
2349  */
2350 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2351                                 gfp_t gfp_mask, enum charge_type ctype)
2352 {
2353         struct mem_cgroup *mem = NULL;
2354         struct page_cgroup *pc;
2355         int ret;
2356         int page_size = PAGE_SIZE;
2357
2358         if (PageTransHuge(page)) {
2359                 page_size <<= compound_order(page);
2360                 VM_BUG_ON(!PageTransHuge(page));
2361         }
2362
2363         pc = lookup_page_cgroup(page);
2364         /* can happen at boot */
2365         if (unlikely(!pc))
2366                 return 0;
2367         prefetchw(pc);
2368
2369         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page_size);
2370         if (ret || !mem)
2371                 return ret;
2372
2373         __mem_cgroup_commit_charge(mem, pc, ctype, page_size);
2374         return 0;
2375 }
2376
2377 int mem_cgroup_newpage_charge(struct page *page,
2378                               struct mm_struct *mm, gfp_t gfp_mask)
2379 {
2380         if (mem_cgroup_disabled())
2381                 return 0;
2382         /*
2383          * If already mapped, we don't have to account.
2384          * If page cache, page->mapping has address_space.
2385          * But page->mapping may have out-of-use anon_vma pointer,
2386          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2387          * is NULL.
2388          */
2389         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2390                 return 0;
2391         if (unlikely(!mm))
2392                 mm = &init_mm;
2393         return mem_cgroup_charge_common(page, mm, gfp_mask,
2394                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2395 }
2396
2397 static void
2398 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2399                                         enum charge_type ctype);
2400
2401 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2402                                 gfp_t gfp_mask)
2403 {
2404         int ret;
2405
2406         if (mem_cgroup_disabled())
2407                 return 0;
2408         if (PageCompound(page))
2409                 return 0;
2410         /*
2411          * Corner case handling. This is called from add_to_page_cache()
2412          * in usual. But some FS (shmem) precharges this page before calling it
2413          * and call add_to_page_cache() with GFP_NOWAIT.
2414          *
2415          * For GFP_NOWAIT case, the page may be pre-charged before calling
2416          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2417          * charge twice. (It works but has to pay a bit larger cost.)
2418          * And when the page is SwapCache, it should take swap information
2419          * into account. This is under lock_page() now.
2420          */
2421         if (!(gfp_mask & __GFP_WAIT)) {
2422                 struct page_cgroup *pc;
2423
2424                 pc = lookup_page_cgroup(page);
2425                 if (!pc)
2426                         return 0;
2427                 lock_page_cgroup(pc);
2428                 if (PageCgroupUsed(pc)) {
2429                         unlock_page_cgroup(pc);
2430                         return 0;
2431                 }
2432                 unlock_page_cgroup(pc);
2433         }
2434
2435         if (unlikely(!mm))
2436                 mm = &init_mm;
2437
2438         if (page_is_file_cache(page))
2439                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2440                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2441
2442         /* shmem */
2443         if (PageSwapCache(page)) {
2444                 struct mem_cgroup *mem = NULL;
2445
2446                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2447                 if (!ret)
2448                         __mem_cgroup_commit_charge_swapin(page, mem,
2449                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2450         } else
2451                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2452                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2453
2454         return ret;
2455 }
2456
2457 /*
2458  * While swap-in, try_charge -> commit or cancel, the page is locked.
2459  * And when try_charge() successfully returns, one refcnt to memcg without
2460  * struct page_cgroup is acquired. This refcnt will be consumed by
2461  * "commit()" or removed by "cancel()"
2462  */
2463 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2464                                  struct page *page,
2465                                  gfp_t mask, struct mem_cgroup **ptr)
2466 {
2467         struct mem_cgroup *mem;
2468         int ret;
2469
2470         if (mem_cgroup_disabled())
2471                 return 0;
2472
2473         if (!do_swap_account)
2474                 goto charge_cur_mm;
2475         /*
2476          * A racing thread's fault, or swapoff, may have already updated
2477          * the pte, and even removed page from swap cache: in those cases
2478          * do_swap_page()'s pte_same() test will fail; but there's also a
2479          * KSM case which does need to charge the page.
2480          */
2481         if (!PageSwapCache(page))
2482                 goto charge_cur_mm;
2483         mem = try_get_mem_cgroup_from_page(page);
2484         if (!mem)
2485                 goto charge_cur_mm;
2486         *ptr = mem;
2487         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2488         css_put(&mem->css);
2489         return ret;
2490 charge_cur_mm:
2491         if (unlikely(!mm))
2492                 mm = &init_mm;
2493         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2494 }
2495
2496 static void
2497 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2498                                         enum charge_type ctype)
2499 {
2500         struct page_cgroup *pc;
2501
2502         if (mem_cgroup_disabled())
2503                 return;
2504         if (!ptr)
2505                 return;
2506         cgroup_exclude_rmdir(&ptr->css);
2507         pc = lookup_page_cgroup(page);
2508         mem_cgroup_lru_del_before_commit_swapcache(page);
2509         __mem_cgroup_commit_charge(ptr, pc, ctype, PAGE_SIZE);
2510         mem_cgroup_lru_add_after_commit_swapcache(page);
2511         /*
2512          * Now swap is on-memory. This means this page may be
2513          * counted both as mem and swap....double count.
2514          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2515          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2516          * may call delete_from_swap_cache() before reach here.
2517          */
2518         if (do_swap_account && PageSwapCache(page)) {
2519                 swp_entry_t ent = {.val = page_private(page)};
2520                 unsigned short id;
2521                 struct mem_cgroup *memcg;
2522
2523                 id = swap_cgroup_record(ent, 0);
2524                 rcu_read_lock();
2525                 memcg = mem_cgroup_lookup(id);
2526                 if (memcg) {
2527                         /*
2528                          * This recorded memcg can be obsolete one. So, avoid
2529                          * calling css_tryget
2530                          */
2531                         if (!mem_cgroup_is_root(memcg))
2532                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2533                         mem_cgroup_swap_statistics(memcg, false);
2534                         mem_cgroup_put(memcg);
2535                 }
2536                 rcu_read_unlock();
2537         }
2538         /*
2539          * At swapin, we may charge account against cgroup which has no tasks.
2540          * So, rmdir()->pre_destroy() can be called while we do this charge.
2541          * In that case, we need to call pre_destroy() again. check it here.
2542          */
2543         cgroup_release_and_wakeup_rmdir(&ptr->css);
2544 }
2545
2546 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2547 {
2548         __mem_cgroup_commit_charge_swapin(page, ptr,
2549                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2550 }
2551
2552 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2553 {
2554         if (mem_cgroup_disabled())
2555                 return;
2556         if (!mem)
2557                 return;
2558         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2559 }
2560
2561 static void
2562 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2563               int page_size)
2564 {
2565         struct memcg_batch_info *batch = NULL;
2566         bool uncharge_memsw = true;
2567         /* If swapout, usage of swap doesn't decrease */
2568         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2569                 uncharge_memsw = false;
2570
2571         batch = &current->memcg_batch;
2572         /*
2573          * In usual, we do css_get() when we remember memcg pointer.
2574          * But in this case, we keep res->usage until end of a series of
2575          * uncharges. Then, it's ok to ignore memcg's refcnt.
2576          */
2577         if (!batch->memcg)
2578                 batch->memcg = mem;
2579         /*
2580          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2581          * In those cases, all pages freed continously can be expected to be in
2582          * the same cgroup and we have chance to coalesce uncharges.
2583          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2584          * because we want to do uncharge as soon as possible.
2585          */
2586
2587         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2588                 goto direct_uncharge;
2589
2590         if (page_size != PAGE_SIZE)
2591                 goto direct_uncharge;
2592
2593         /*
2594          * In typical case, batch->memcg == mem. This means we can
2595          * merge a series of uncharges to an uncharge of res_counter.
2596          * If not, we uncharge res_counter ony by one.
2597          */
2598         if (batch->memcg != mem)
2599                 goto direct_uncharge;
2600         /* remember freed charge and uncharge it later */
2601         batch->bytes += PAGE_SIZE;
2602         if (uncharge_memsw)
2603                 batch->memsw_bytes += PAGE_SIZE;
2604         return;
2605 direct_uncharge:
2606         res_counter_uncharge(&mem->res, page_size);
2607         if (uncharge_memsw)
2608                 res_counter_uncharge(&mem->memsw, page_size);
2609         if (unlikely(batch->memcg != mem))
2610                 memcg_oom_recover(mem);
2611         return;
2612 }
2613
2614 /*
2615  * uncharge if !page_mapped(page)
2616  */
2617 static struct mem_cgroup *
2618 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2619 {
2620         int count;
2621         struct page_cgroup *pc;
2622         struct mem_cgroup *mem = NULL;
2623         int page_size = PAGE_SIZE;
2624
2625         if (mem_cgroup_disabled())
2626                 return NULL;
2627
2628         if (PageSwapCache(page))
2629                 return NULL;
2630
2631         if (PageTransHuge(page)) {
2632                 page_size <<= compound_order(page);
2633                 VM_BUG_ON(!PageTransHuge(page));
2634         }
2635
2636         count = page_size >> PAGE_SHIFT;
2637         /*
2638          * Check if our page_cgroup is valid
2639          */
2640         pc = lookup_page_cgroup(page);
2641         if (unlikely(!pc || !PageCgroupUsed(pc)))
2642                 return NULL;
2643
2644         lock_page_cgroup(pc);
2645
2646         mem = pc->mem_cgroup;
2647
2648         if (!PageCgroupUsed(pc))
2649                 goto unlock_out;
2650
2651         switch (ctype) {
2652         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2653         case MEM_CGROUP_CHARGE_TYPE_DROP:
2654                 /* See mem_cgroup_prepare_migration() */
2655                 if (page_mapped(page) || PageCgroupMigration(pc))
2656                         goto unlock_out;
2657                 break;
2658         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2659                 if (!PageAnon(page)) {  /* Shared memory */
2660                         if (page->mapping && !page_is_file_cache(page))
2661                                 goto unlock_out;
2662                 } else if (page_mapped(page)) /* Anon */
2663                                 goto unlock_out;
2664                 break;
2665         default:
2666                 break;
2667         }
2668
2669         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2670
2671         ClearPageCgroupUsed(pc);
2672         /*
2673          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2674          * freed from LRU. This is safe because uncharged page is expected not
2675          * to be reused (freed soon). Exception is SwapCache, it's handled by
2676          * special functions.
2677          */
2678
2679         unlock_page_cgroup(pc);
2680         /*
2681          * even after unlock, we have mem->res.usage here and this memcg
2682          * will never be freed.
2683          */
2684         memcg_check_events(mem, page);
2685         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2686                 mem_cgroup_swap_statistics(mem, true);
2687                 mem_cgroup_get(mem);
2688         }
2689         if (!mem_cgroup_is_root(mem))
2690                 __do_uncharge(mem, ctype, page_size);
2691
2692         return mem;
2693
2694 unlock_out:
2695         unlock_page_cgroup(pc);
2696         return NULL;
2697 }
2698
2699 void mem_cgroup_uncharge_page(struct page *page)
2700 {
2701         /* early check. */
2702         if (page_mapped(page))
2703                 return;
2704         if (page->mapping && !PageAnon(page))
2705                 return;
2706         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2707 }
2708
2709 void mem_cgroup_uncharge_cache_page(struct page *page)
2710 {
2711         VM_BUG_ON(page_mapped(page));
2712         VM_BUG_ON(page->mapping);
2713         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2714 }
2715
2716 /*
2717  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2718  * In that cases, pages are freed continuously and we can expect pages
2719  * are in the same memcg. All these calls itself limits the number of
2720  * pages freed at once, then uncharge_start/end() is called properly.
2721  * This may be called prural(2) times in a context,
2722  */
2723
2724 void mem_cgroup_uncharge_start(void)
2725 {
2726         current->memcg_batch.do_batch++;
2727         /* We can do nest. */
2728         if (current->memcg_batch.do_batch == 1) {
2729                 current->memcg_batch.memcg = NULL;
2730                 current->memcg_batch.bytes = 0;
2731                 current->memcg_batch.memsw_bytes = 0;
2732         }
2733 }
2734
2735 void mem_cgroup_uncharge_end(void)
2736 {
2737         struct memcg_batch_info *batch = &current->memcg_batch;
2738
2739         if (!batch->do_batch)
2740                 return;
2741
2742         batch->do_batch--;
2743         if (batch->do_batch) /* If stacked, do nothing. */
2744                 return;
2745
2746         if (!batch->memcg)
2747                 return;
2748         /*
2749          * This "batch->memcg" is valid without any css_get/put etc...
2750          * bacause we hide charges behind us.
2751          */
2752         if (batch->bytes)
2753                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2754         if (batch->memsw_bytes)
2755                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2756         memcg_oom_recover(batch->memcg);
2757         /* forget this pointer (for sanity check) */
2758         batch->memcg = NULL;
2759 }
2760
2761 #ifdef CONFIG_SWAP
2762 /*
2763  * called after __delete_from_swap_cache() and drop "page" account.
2764  * memcg information is recorded to swap_cgroup of "ent"
2765  */
2766 void
2767 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2768 {
2769         struct mem_cgroup *memcg;
2770         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2771
2772         if (!swapout) /* this was a swap cache but the swap is unused ! */
2773                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2774
2775         memcg = __mem_cgroup_uncharge_common(page, ctype);
2776
2777         /*
2778          * record memcg information,  if swapout && memcg != NULL,
2779          * mem_cgroup_get() was called in uncharge().
2780          */
2781         if (do_swap_account && swapout && memcg)
2782                 swap_cgroup_record(ent, css_id(&memcg->css));
2783 }
2784 #endif
2785
2786 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2787 /*
2788  * called from swap_entry_free(). remove record in swap_cgroup and
2789  * uncharge "memsw" account.
2790  */
2791 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2792 {
2793         struct mem_cgroup *memcg;
2794         unsigned short id;
2795
2796         if (!do_swap_account)
2797                 return;
2798
2799         id = swap_cgroup_record(ent, 0);
2800         rcu_read_lock();
2801         memcg = mem_cgroup_lookup(id);
2802         if (memcg) {
2803                 /*
2804                  * We uncharge this because swap is freed.
2805                  * This memcg can be obsolete one. We avoid calling css_tryget
2806                  */
2807                 if (!mem_cgroup_is_root(memcg))
2808                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2809                 mem_cgroup_swap_statistics(memcg, false);
2810                 mem_cgroup_put(memcg);
2811         }
2812         rcu_read_unlock();
2813 }
2814
2815 /**
2816  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2817  * @entry: swap entry to be moved
2818  * @from:  mem_cgroup which the entry is moved from
2819  * @to:  mem_cgroup which the entry is moved to
2820  * @need_fixup: whether we should fixup res_counters and refcounts.
2821  *
2822  * It succeeds only when the swap_cgroup's record for this entry is the same
2823  * as the mem_cgroup's id of @from.
2824  *
2825  * Returns 0 on success, -EINVAL on failure.
2826  *
2827  * The caller must have charged to @to, IOW, called res_counter_charge() about
2828  * both res and memsw, and called css_get().
2829  */
2830 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2831                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2832 {
2833         unsigned short old_id, new_id;
2834
2835         old_id = css_id(&from->css);
2836         new_id = css_id(&to->css);
2837
2838         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2839                 mem_cgroup_swap_statistics(from, false);
2840                 mem_cgroup_swap_statistics(to, true);
2841                 /*
2842                  * This function is only called from task migration context now.
2843                  * It postpones res_counter and refcount handling till the end
2844                  * of task migration(mem_cgroup_clear_mc()) for performance
2845                  * improvement. But we cannot postpone mem_cgroup_get(to)
2846                  * because if the process that has been moved to @to does
2847                  * swap-in, the refcount of @to might be decreased to 0.
2848                  */
2849                 mem_cgroup_get(to);
2850                 if (need_fixup) {
2851                         if (!mem_cgroup_is_root(from))
2852                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2853                         mem_cgroup_put(from);
2854                         /*
2855                          * we charged both to->res and to->memsw, so we should
2856                          * uncharge to->res.
2857                          */
2858                         if (!mem_cgroup_is_root(to))
2859                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2860                 }
2861                 return 0;
2862         }
2863         return -EINVAL;
2864 }
2865 #else
2866 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2867                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2868 {
2869         return -EINVAL;
2870 }
2871 #endif
2872
2873 /*
2874  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2875  * page belongs to.
2876  */
2877 int mem_cgroup_prepare_migration(struct page *page,
2878         struct page *newpage, struct mem_cgroup **ptr)
2879 {
2880         struct page_cgroup *pc;
2881         struct mem_cgroup *mem = NULL;
2882         enum charge_type ctype;
2883         int ret = 0;
2884
2885         VM_BUG_ON(PageTransHuge(page));
2886         if (mem_cgroup_disabled())
2887                 return 0;
2888
2889         pc = lookup_page_cgroup(page);
2890         lock_page_cgroup(pc);
2891         if (PageCgroupUsed(pc)) {
2892                 mem = pc->mem_cgroup;
2893                 css_get(&mem->css);
2894                 /*
2895                  * At migrating an anonymous page, its mapcount goes down
2896                  * to 0 and uncharge() will be called. But, even if it's fully
2897                  * unmapped, migration may fail and this page has to be
2898                  * charged again. We set MIGRATION flag here and delay uncharge
2899                  * until end_migration() is called
2900                  *
2901                  * Corner Case Thinking
2902                  * A)
2903                  * When the old page was mapped as Anon and it's unmap-and-freed
2904                  * while migration was ongoing.
2905                  * If unmap finds the old page, uncharge() of it will be delayed
2906                  * until end_migration(). If unmap finds a new page, it's
2907                  * uncharged when it make mapcount to be 1->0. If unmap code
2908                  * finds swap_migration_entry, the new page will not be mapped
2909                  * and end_migration() will find it(mapcount==0).
2910                  *
2911                  * B)
2912                  * When the old page was mapped but migraion fails, the kernel
2913                  * remaps it. A charge for it is kept by MIGRATION flag even
2914                  * if mapcount goes down to 0. We can do remap successfully
2915                  * without charging it again.
2916                  *
2917                  * C)
2918                  * The "old" page is under lock_page() until the end of
2919                  * migration, so, the old page itself will not be swapped-out.
2920                  * If the new page is swapped out before end_migraton, our
2921                  * hook to usual swap-out path will catch the event.
2922                  */
2923                 if (PageAnon(page))
2924                         SetPageCgroupMigration(pc);
2925         }
2926         unlock_page_cgroup(pc);
2927         /*
2928          * If the page is not charged at this point,
2929          * we return here.
2930          */
2931         if (!mem)
2932                 return 0;
2933
2934         *ptr = mem;
2935         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false, PAGE_SIZE);
2936         css_put(&mem->css);/* drop extra refcnt */
2937         if (ret || *ptr == NULL) {
2938                 if (PageAnon(page)) {
2939                         lock_page_cgroup(pc);
2940                         ClearPageCgroupMigration(pc);
2941                         unlock_page_cgroup(pc);
2942                         /*
2943                          * The old page may be fully unmapped while we kept it.
2944                          */
2945                         mem_cgroup_uncharge_page(page);
2946                 }
2947                 return -ENOMEM;
2948         }
2949         /*
2950          * We charge new page before it's used/mapped. So, even if unlock_page()
2951          * is called before end_migration, we can catch all events on this new
2952          * page. In the case new page is migrated but not remapped, new page's
2953          * mapcount will be finally 0 and we call uncharge in end_migration().
2954          */
2955         pc = lookup_page_cgroup(newpage);
2956         if (PageAnon(page))
2957                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2958         else if (page_is_file_cache(page))
2959                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2960         else
2961                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2962         __mem_cgroup_commit_charge(mem, pc, ctype, PAGE_SIZE);
2963         return ret;
2964 }
2965
2966 /* remove redundant charge if migration failed*/
2967 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2968         struct page *oldpage, struct page *newpage, bool migration_ok)
2969 {
2970         struct page *used, *unused;
2971         struct page_cgroup *pc;
2972
2973         if (!mem)
2974                 return;
2975         /* blocks rmdir() */
2976         cgroup_exclude_rmdir(&mem->css);
2977         if (!migration_ok) {
2978                 used = oldpage;
2979                 unused = newpage;
2980         } else {
2981                 used = newpage;
2982                 unused = oldpage;
2983         }
2984         /*
2985          * We disallowed uncharge of pages under migration because mapcount
2986          * of the page goes down to zero, temporarly.
2987          * Clear the flag and check the page should be charged.
2988          */
2989         pc = lookup_page_cgroup(oldpage);
2990         lock_page_cgroup(pc);
2991         ClearPageCgroupMigration(pc);
2992         unlock_page_cgroup(pc);
2993
2994         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2995
2996         /*
2997          * If a page is a file cache, radix-tree replacement is very atomic
2998          * and we can skip this check. When it was an Anon page, its mapcount
2999          * goes down to 0. But because we added MIGRATION flage, it's not
3000          * uncharged yet. There are several case but page->mapcount check
3001          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3002          * check. (see prepare_charge() also)
3003          */
3004         if (PageAnon(used))
3005                 mem_cgroup_uncharge_page(used);
3006         /*
3007          * At migration, we may charge account against cgroup which has no
3008          * tasks.
3009          * So, rmdir()->pre_destroy() can be called while we do this charge.
3010          * In that case, we need to call pre_destroy() again. check it here.
3011          */
3012         cgroup_release_and_wakeup_rmdir(&mem->css);
3013 }
3014
3015 /*
3016  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3017  * Calling hierarchical_reclaim is not enough because we should update
3018  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3019  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3020  * not from the memcg which this page would be charged to.
3021  * try_charge_swapin does all of these works properly.
3022  */
3023 int mem_cgroup_shmem_charge_fallback(struct page *page,
3024                             struct mm_struct *mm,
3025                             gfp_t gfp_mask)
3026 {
3027         struct mem_cgroup *mem = NULL;
3028         int ret;
3029
3030         if (mem_cgroup_disabled())
3031                 return 0;
3032
3033         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3034         if (!ret)
3035                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3036
3037         return ret;
3038 }
3039
3040 static DEFINE_MUTEX(set_limit_mutex);
3041
3042 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3043                                 unsigned long long val)
3044 {
3045         int retry_count;
3046         u64 memswlimit, memlimit;
3047         int ret = 0;
3048         int children = mem_cgroup_count_children(memcg);
3049         u64 curusage, oldusage;
3050         int enlarge;
3051
3052         /*
3053          * For keeping hierarchical_reclaim simple, how long we should retry
3054          * is depends on callers. We set our retry-count to be function
3055          * of # of children which we should visit in this loop.
3056          */
3057         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3058
3059         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3060
3061         enlarge = 0;
3062         while (retry_count) {
3063                 if (signal_pending(current)) {
3064                         ret = -EINTR;
3065                         break;
3066                 }
3067                 /*
3068                  * Rather than hide all in some function, I do this in
3069                  * open coded manner. You see what this really does.
3070                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3071                  */
3072                 mutex_lock(&set_limit_mutex);
3073                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3074                 if (memswlimit < val) {
3075                         ret = -EINVAL;
3076                         mutex_unlock(&set_limit_mutex);
3077                         break;
3078                 }
3079
3080                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3081                 if (memlimit < val)
3082                         enlarge = 1;
3083
3084                 ret = res_counter_set_limit(&memcg->res, val);
3085                 if (!ret) {
3086                         if (memswlimit == val)
3087                                 memcg->memsw_is_minimum = true;
3088                         else
3089                                 memcg->memsw_is_minimum = false;
3090                 }
3091                 mutex_unlock(&set_limit_mutex);
3092
3093                 if (!ret)
3094                         break;
3095
3096                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3097                                                 MEM_CGROUP_RECLAIM_SHRINK);
3098                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3099                 /* Usage is reduced ? */
3100                 if (curusage >= oldusage)
3101                         retry_count--;
3102                 else
3103                         oldusage = curusage;
3104         }
3105         if (!ret && enlarge)
3106                 memcg_oom_recover(memcg);
3107
3108         return ret;
3109 }
3110
3111 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3112                                         unsigned long long val)
3113 {
3114         int retry_count;
3115         u64 memlimit, memswlimit, oldusage, curusage;
3116         int children = mem_cgroup_count_children(memcg);
3117         int ret = -EBUSY;
3118         int enlarge = 0;
3119
3120         /* see mem_cgroup_resize_res_limit */
3121         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3122         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3123         while (retry_count) {
3124                 if (signal_pending(current)) {
3125                         ret = -EINTR;
3126                         break;
3127                 }
3128                 /*
3129                  * Rather than hide all in some function, I do this in
3130                  * open coded manner. You see what this really does.
3131                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3132                  */
3133                 mutex_lock(&set_limit_mutex);
3134                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3135                 if (memlimit > val) {
3136                         ret = -EINVAL;
3137                         mutex_unlock(&set_limit_mutex);
3138                         break;
3139                 }
3140                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3141                 if (memswlimit < val)
3142                         enlarge = 1;
3143                 ret = res_counter_set_limit(&memcg->memsw, val);
3144                 if (!ret) {
3145                         if (memlimit == val)
3146                                 memcg->memsw_is_minimum = true;
3147                         else
3148                                 memcg->memsw_is_minimum = false;
3149                 }
3150                 mutex_unlock(&set_limit_mutex);
3151
3152                 if (!ret)
3153                         break;
3154
3155                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3156                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3157                                                 MEM_CGROUP_RECLAIM_SHRINK);
3158                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3159                 /* Usage is reduced ? */
3160                 if (curusage >= oldusage)
3161                         retry_count--;
3162                 else
3163                         oldusage = curusage;
3164         }
3165         if (!ret && enlarge)
3166                 memcg_oom_recover(memcg);
3167         return ret;
3168 }
3169
3170 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3171                                             gfp_t gfp_mask)
3172 {
3173         unsigned long nr_reclaimed = 0;
3174         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3175         unsigned long reclaimed;
3176         int loop = 0;
3177         struct mem_cgroup_tree_per_zone *mctz;
3178         unsigned long long excess;
3179
3180         if (order > 0)
3181                 return 0;
3182
3183         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3184         /*
3185          * This loop can run a while, specially if mem_cgroup's continuously
3186          * keep exceeding their soft limit and putting the system under
3187          * pressure
3188          */
3189         do {
3190                 if (next_mz)
3191                         mz = next_mz;
3192                 else
3193                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3194                 if (!mz)
3195                         break;
3196
3197                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3198                                                 gfp_mask,
3199                                                 MEM_CGROUP_RECLAIM_SOFT);
3200                 nr_reclaimed += reclaimed;
3201                 spin_lock(&mctz->lock);
3202
3203                 /*
3204                  * If we failed to reclaim anything from this memory cgroup
3205                  * it is time to move on to the next cgroup
3206                  */
3207                 next_mz = NULL;
3208                 if (!reclaimed) {
3209                         do {
3210                                 /*
3211                                  * Loop until we find yet another one.
3212                                  *
3213                                  * By the time we get the soft_limit lock
3214                                  * again, someone might have aded the
3215                                  * group back on the RB tree. Iterate to
3216                                  * make sure we get a different mem.
3217                                  * mem_cgroup_largest_soft_limit_node returns
3218                                  * NULL if no other cgroup is present on
3219                                  * the tree
3220                                  */
3221                                 next_mz =
3222                                 __mem_cgroup_largest_soft_limit_node(mctz);
3223                                 if (next_mz == mz) {
3224                                         css_put(&next_mz->mem->css);
3225                                         next_mz = NULL;
3226                                 } else /* next_mz == NULL or other memcg */
3227                                         break;
3228                         } while (1);
3229                 }
3230                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3231                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3232                 /*
3233                  * One school of thought says that we should not add
3234                  * back the node to the tree if reclaim returns 0.
3235                  * But our reclaim could return 0, simply because due
3236                  * to priority we are exposing a smaller subset of
3237                  * memory to reclaim from. Consider this as a longer
3238                  * term TODO.
3239                  */
3240                 /* If excess == 0, no tree ops */
3241                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3242                 spin_unlock(&mctz->lock);
3243                 css_put(&mz->mem->css);
3244                 loop++;
3245                 /*
3246                  * Could not reclaim anything and there are no more
3247                  * mem cgroups to try or we seem to be looping without
3248                  * reclaiming anything.
3249                  */
3250                 if (!nr_reclaimed &&
3251                         (next_mz == NULL ||
3252                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3253                         break;
3254         } while (!nr_reclaimed);
3255         if (next_mz)
3256                 css_put(&next_mz->mem->css);
3257         return nr_reclaimed;
3258 }
3259
3260 /*
3261  * This routine traverse page_cgroup in given list and drop them all.
3262  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3263  */
3264 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3265                                 int node, int zid, enum lru_list lru)
3266 {
3267         struct zone *zone;
3268         struct mem_cgroup_per_zone *mz;
3269         struct page_cgroup *pc, *busy;
3270         unsigned long flags, loop;
3271         struct list_head *list;
3272         int ret = 0;
3273
3274         zone = &NODE_DATA(node)->node_zones[zid];
3275         mz = mem_cgroup_zoneinfo(mem, node, zid);
3276         list = &mz->lists[lru];
3277
3278         loop = MEM_CGROUP_ZSTAT(mz, lru);
3279         /* give some margin against EBUSY etc...*/
3280         loop += 256;
3281         busy = NULL;
3282         while (loop--) {
3283                 ret = 0;
3284                 spin_lock_irqsave(&zone->lru_lock, flags);
3285                 if (list_empty(list)) {
3286                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3287                         break;
3288                 }
3289                 pc = list_entry(list->prev, struct page_cgroup, lru);
3290                 if (busy == pc) {
3291                         list_move(&pc->lru, list);
3292                         busy = NULL;
3293                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3294                         continue;
3295                 }
3296                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3297
3298                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3299                 if (ret == -ENOMEM)
3300                         break;
3301
3302                 if (ret == -EBUSY || ret == -EINVAL) {
3303                         /* found lock contention or "pc" is obsolete. */
3304                         busy = pc;
3305                         cond_resched();
3306                 } else
3307                         busy = NULL;
3308         }
3309
3310         if (!ret && !list_empty(list))
3311                 return -EBUSY;
3312         return ret;
3313 }
3314
3315 /*
3316  * make mem_cgroup's charge to be 0 if there is no task.
3317  * This enables deleting this mem_cgroup.
3318  */
3319 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3320 {
3321         int ret;
3322         int node, zid, shrink;
3323         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3324         struct cgroup *cgrp = mem->css.cgroup;
3325
3326         css_get(&mem->css);
3327
3328         shrink = 0;
3329         /* should free all ? */
3330         if (free_all)
3331                 goto try_to_free;
3332 move_account:
3333         do {
3334                 ret = -EBUSY;
3335                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3336                         goto out;
3337                 ret = -EINTR;
3338                 if (signal_pending(current))
3339                         goto out;
3340                 /* This is for making all *used* pages to be on LRU. */
3341                 lru_add_drain_all();
3342                 drain_all_stock_sync();
3343                 ret = 0;
3344                 mem_cgroup_start_move(mem);
3345                 for_each_node_state(node, N_HIGH_MEMORY) {
3346                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3347                                 enum lru_list l;
3348                                 for_each_lru(l) {
3349                                         ret = mem_cgroup_force_empty_list(mem,
3350                                                         node, zid, l);
3351                                         if (ret)
3352                                                 break;
3353                                 }
3354                         }
3355                         if (ret)
3356                                 break;
3357                 }
3358                 mem_cgroup_end_move(mem);
3359                 memcg_oom_recover(mem);
3360                 /* it seems parent cgroup doesn't have enough mem */
3361                 if (ret == -ENOMEM)
3362                         goto try_to_free;
3363                 cond_resched();
3364         /* "ret" should also be checked to ensure all lists are empty. */
3365         } while (mem->res.usage > 0 || ret);
3366 out:
3367         css_put(&mem->css);
3368         return ret;
3369
3370 try_to_free:
3371         /* returns EBUSY if there is a task or if we come here twice. */
3372         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3373                 ret = -EBUSY;
3374                 goto out;
3375         }
3376         /* we call try-to-free pages for make this cgroup empty */
3377         lru_add_drain_all();
3378         /* try to free all pages in this cgroup */
3379         shrink = 1;
3380         while (nr_retries && mem->res.usage > 0) {
3381                 int progress;
3382
3383                 if (signal_pending(current)) {
3384                         ret = -EINTR;
3385                         goto out;
3386                 }
3387                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3388                                                 false, get_swappiness(mem));
3389                 if (!progress) {
3390                         nr_retries--;
3391                         /* maybe some writeback is necessary */
3392                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3393                 }
3394
3395         }
3396         lru_add_drain();
3397         /* try move_account...there may be some *locked* pages. */
3398         goto move_account;
3399 }
3400
3401 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3402 {
3403         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3404 }
3405
3406
3407 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3408 {
3409         return mem_cgroup_from_cont(cont)->use_hierarchy;
3410 }
3411
3412 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3413                                         u64 val)
3414 {
3415         int retval = 0;
3416         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3417         struct cgroup *parent = cont->parent;
3418         struct mem_cgroup *parent_mem = NULL;
3419
3420         if (parent)
3421                 parent_mem = mem_cgroup_from_cont(parent);
3422
3423         cgroup_lock();
3424         /*
3425          * If parent's use_hierarchy is set, we can't make any modifications
3426          * in the child subtrees. If it is unset, then the change can
3427          * occur, provided the current cgroup has no children.
3428          *
3429          * For the root cgroup, parent_mem is NULL, we allow value to be
3430          * set if there are no children.
3431          */
3432         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3433                                 (val == 1 || val == 0)) {
3434                 if (list_empty(&cont->children))
3435                         mem->use_hierarchy = val;
3436                 else
3437                         retval = -EBUSY;
3438         } else
3439                 retval = -EINVAL;
3440         cgroup_unlock();
3441
3442         return retval;
3443 }
3444
3445
3446 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3447                                 enum mem_cgroup_stat_index idx)
3448 {
3449         struct mem_cgroup *iter;
3450         s64 val = 0;
3451
3452         /* each per cpu's value can be minus.Then, use s64 */
3453         for_each_mem_cgroup_tree(iter, mem)
3454                 val += mem_cgroup_read_stat(iter, idx);
3455
3456         if (val < 0) /* race ? */
3457                 val = 0;
3458         return val;
3459 }
3460
3461 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3462 {
3463         u64 val;
3464
3465         if (!mem_cgroup_is_root(mem)) {
3466                 if (!swap)
3467                         return res_counter_read_u64(&mem->res, RES_USAGE);
3468                 else
3469                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3470         }
3471
3472         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3473         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3474
3475         if (swap)
3476                 val += mem_cgroup_get_recursive_idx_stat(mem,
3477                                 MEM_CGROUP_STAT_SWAPOUT);
3478
3479         return val << PAGE_SHIFT;
3480 }
3481
3482 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3483 {
3484         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3485         u64 val;
3486         int type, name;
3487
3488         type = MEMFILE_TYPE(cft->private);
3489         name = MEMFILE_ATTR(cft->private);
3490         switch (type) {
3491         case _MEM:
3492                 if (name == RES_USAGE)
3493                         val = mem_cgroup_usage(mem, false);
3494                 else
3495                         val = res_counter_read_u64(&mem->res, name);
3496                 break;
3497         case _MEMSWAP:
3498                 if (name == RES_USAGE)
3499                         val = mem_cgroup_usage(mem, true);
3500                 else
3501                         val = res_counter_read_u64(&mem->memsw, name);
3502                 break;
3503         default:
3504                 BUG();
3505                 break;
3506         }
3507         return val;
3508 }
3509 /*
3510  * The user of this function is...
3511  * RES_LIMIT.
3512  */
3513 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3514                             const char *buffer)
3515 {
3516         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3517         int type, name;
3518         unsigned long long val;
3519         int ret;
3520
3521         type = MEMFILE_TYPE(cft->private);
3522         name = MEMFILE_ATTR(cft->private);
3523         switch (name) {
3524         case RES_LIMIT:
3525                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3526                         ret = -EINVAL;
3527                         break;
3528                 }
3529                 /* This function does all necessary parse...reuse it */
3530                 ret = res_counter_memparse_write_strategy(buffer, &val);
3531                 if (ret)
3532                         break;
3533                 if (type == _MEM)
3534                         ret = mem_cgroup_resize_limit(memcg, val);
3535                 else
3536                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3537                 break;
3538         case RES_SOFT_LIMIT:
3539                 ret = res_counter_memparse_write_strategy(buffer, &val);
3540                 if (ret)
3541                         break;
3542                 /*
3543                  * For memsw, soft limits are hard to implement in terms
3544                  * of semantics, for now, we support soft limits for
3545                  * control without swap
3546                  */
3547                 if (type == _MEM)
3548                         ret = res_counter_set_soft_limit(&memcg->res, val);
3549                 else
3550                         ret = -EINVAL;
3551                 break;
3552         default:
3553                 ret = -EINVAL; /* should be BUG() ? */
3554                 break;
3555         }
3556         return ret;
3557 }
3558
3559 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3560                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3561 {
3562         struct cgroup *cgroup;
3563         unsigned long long min_limit, min_memsw_limit, tmp;
3564
3565         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3566         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3567         cgroup = memcg->css.cgroup;
3568         if (!memcg->use_hierarchy)
3569                 goto out;
3570
3571         while (cgroup->parent) {
3572                 cgroup = cgroup->parent;
3573                 memcg = mem_cgroup_from_cont(cgroup);
3574                 if (!memcg->use_hierarchy)
3575                         break;
3576                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3577                 min_limit = min(min_limit, tmp);
3578                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3579                 min_memsw_limit = min(min_memsw_limit, tmp);
3580         }
3581 out:
3582         *mem_limit = min_limit;
3583         *memsw_limit = min_memsw_limit;
3584         return;
3585 }
3586
3587 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3588 {
3589         struct mem_cgroup *mem;
3590         int type, name;
3591
3592         mem = mem_cgroup_from_cont(cont);
3593         type = MEMFILE_TYPE(event);
3594         name = MEMFILE_ATTR(event);
3595         switch (name) {
3596         case RES_MAX_USAGE:
3597                 if (type == _MEM)
3598                         res_counter_reset_max(&mem->res);
3599                 else
3600                         res_counter_reset_max(&mem->memsw);
3601                 break;
3602         case RES_FAILCNT:
3603                 if (type == _MEM)
3604                         res_counter_reset_failcnt(&mem->res);
3605                 else
3606                         res_counter_reset_failcnt(&mem->memsw);
3607                 break;
3608         }
3609
3610         return 0;
3611 }
3612
3613 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3614                                         struct cftype *cft)
3615 {
3616         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3617 }
3618
3619 #ifdef CONFIG_MMU
3620 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3621                                         struct cftype *cft, u64 val)
3622 {
3623         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3624
3625         if (val >= (1 << NR_MOVE_TYPE))
3626                 return -EINVAL;
3627         /*
3628          * We check this value several times in both in can_attach() and
3629          * attach(), so we need cgroup lock to prevent this value from being
3630          * inconsistent.
3631          */
3632         cgroup_lock();
3633         mem->move_charge_at_immigrate = val;
3634         cgroup_unlock();
3635
3636         return 0;
3637 }
3638 #else
3639 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3640                                         struct cftype *cft, u64 val)
3641 {
3642         return -ENOSYS;
3643 }
3644 #endif
3645
3646
3647 /* For read statistics */
3648 enum {
3649         MCS_CACHE,
3650         MCS_RSS,
3651         MCS_FILE_MAPPED,
3652         MCS_PGPGIN,
3653         MCS_PGPGOUT,
3654         MCS_SWAP,
3655         MCS_INACTIVE_ANON,
3656         MCS_ACTIVE_ANON,
3657         MCS_INACTIVE_FILE,
3658         MCS_ACTIVE_FILE,
3659         MCS_UNEVICTABLE,
3660         NR_MCS_STAT,
3661 };
3662
3663 struct mcs_total_stat {
3664         s64 stat[NR_MCS_STAT];
3665 };
3666
3667 struct {
3668         char *local_name;
3669         char *total_name;
3670 } memcg_stat_strings[NR_MCS_STAT] = {
3671         {"cache", "total_cache"},
3672         {"rss", "total_rss"},
3673         {"mapped_file", "total_mapped_file"},
3674         {"pgpgin", "total_pgpgin"},
3675         {"pgpgout", "total_pgpgout"},
3676         {"swap", "total_swap"},
3677         {"inactive_anon", "total_inactive_anon"},
3678         {"active_anon", "total_active_anon"},
3679         {"inactive_file", "total_inactive_file"},
3680         {"active_file", "total_active_file"},
3681         {"unevictable", "total_unevictable"}
3682 };
3683
3684
3685 static void
3686 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3687 {
3688         s64 val;
3689
3690         /* per cpu stat */
3691         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3692         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3693         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3694         s->stat[MCS_RSS] += val * PAGE_SIZE;
3695         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3696         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3697         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3698         s->stat[MCS_PGPGIN] += val;
3699         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3700         s->stat[MCS_PGPGOUT] += val;
3701         if (do_swap_account) {
3702                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3703                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3704         }
3705
3706         /* per zone stat */
3707         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3708         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3709         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3710         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3711         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3712         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3713         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3714         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3715         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3716         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3717 }
3718
3719 static void
3720 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3721 {
3722         struct mem_cgroup *iter;
3723
3724         for_each_mem_cgroup_tree(iter, mem)
3725                 mem_cgroup_get_local_stat(iter, s);
3726 }
3727
3728 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3729                                  struct cgroup_map_cb *cb)
3730 {
3731         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3732         struct mcs_total_stat mystat;
3733         int i;
3734
3735         memset(&mystat, 0, sizeof(mystat));
3736         mem_cgroup_get_local_stat(mem_cont, &mystat);
3737
3738         for (i = 0; i < NR_MCS_STAT; i++) {
3739                 if (i == MCS_SWAP && !do_swap_account)
3740                         continue;
3741                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3742         }
3743
3744         /* Hierarchical information */
3745         {
3746                 unsigned long long limit, memsw_limit;
3747                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3748                 cb->fill(cb, "hierarchical_memory_limit", limit);
3749                 if (do_swap_account)
3750                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3751         }
3752
3753         memset(&mystat, 0, sizeof(mystat));
3754         mem_cgroup_get_total_stat(mem_cont, &mystat);
3755         for (i = 0; i < NR_MCS_STAT; i++) {
3756                 if (i == MCS_SWAP && !do_swap_account)
3757                         continue;
3758                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3759         }
3760
3761 #ifdef CONFIG_DEBUG_VM
3762         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3763
3764         {
3765                 int nid, zid;
3766                 struct mem_cgroup_per_zone *mz;
3767                 unsigned long recent_rotated[2] = {0, 0};
3768                 unsigned long recent_scanned[2] = {0, 0};
3769
3770                 for_each_online_node(nid)
3771                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3772                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3773
3774                                 recent_rotated[0] +=
3775                                         mz->reclaim_stat.recent_rotated[0];
3776                                 recent_rotated[1] +=
3777                                         mz->reclaim_stat.recent_rotated[1];
3778                                 recent_scanned[0] +=
3779                                         mz->reclaim_stat.recent_scanned[0];
3780                                 recent_scanned[1] +=
3781                                         mz->reclaim_stat.recent_scanned[1];
3782                         }
3783                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3784                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3785                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3786                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3787         }
3788 #endif
3789
3790         return 0;
3791 }
3792
3793 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3794 {
3795         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3796
3797         return get_swappiness(memcg);
3798 }
3799
3800 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3801                                        u64 val)
3802 {
3803         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3804         struct mem_cgroup *parent;
3805
3806         if (val > 100)
3807                 return -EINVAL;
3808
3809         if (cgrp->parent == NULL)
3810                 return -EINVAL;
3811
3812         parent = mem_cgroup_from_cont(cgrp->parent);
3813
3814         cgroup_lock();
3815
3816         /* If under hierarchy, only empty-root can set this value */
3817         if ((parent->use_hierarchy) ||
3818             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3819                 cgroup_unlock();
3820                 return -EINVAL;
3821         }
3822
3823         spin_lock(&memcg->reclaim_param_lock);
3824         memcg->swappiness = val;
3825         spin_unlock(&memcg->reclaim_param_lock);
3826
3827         cgroup_unlock();
3828
3829         return 0;
3830 }
3831
3832 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3833 {
3834         struct mem_cgroup_threshold_ary *t;
3835         u64 usage;
3836         int i;
3837
3838         rcu_read_lock();
3839         if (!swap)
3840                 t = rcu_dereference(memcg->thresholds.primary);
3841         else
3842                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3843
3844         if (!t)
3845                 goto unlock;
3846
3847         usage = mem_cgroup_usage(memcg, swap);
3848
3849         /*
3850          * current_threshold points to threshold just below usage.
3851          * If it's not true, a threshold was crossed after last
3852          * call of __mem_cgroup_threshold().
3853          */
3854         i = t->current_threshold;
3855
3856         /*
3857          * Iterate backward over array of thresholds starting from
3858          * current_threshold and check if a threshold is crossed.
3859          * If none of thresholds below usage is crossed, we read
3860          * only one element of the array here.
3861          */
3862         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3863                 eventfd_signal(t->entries[i].eventfd, 1);
3864
3865         /* i = current_threshold + 1 */
3866         i++;
3867
3868         /*
3869          * Iterate forward over array of thresholds starting from
3870          * current_threshold+1 and check if a threshold is crossed.
3871          * If none of thresholds above usage is crossed, we read
3872          * only one element of the array here.
3873          */
3874         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3875                 eventfd_signal(t->entries[i].eventfd, 1);
3876
3877         /* Update current_threshold */
3878         t->current_threshold = i - 1;
3879 unlock:
3880         rcu_read_unlock();
3881 }
3882
3883 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3884 {
3885         while (memcg) {
3886                 __mem_cgroup_threshold(memcg, false);
3887                 if (do_swap_account)
3888                         __mem_cgroup_threshold(memcg, true);
3889
3890                 memcg = parent_mem_cgroup(memcg);
3891         }
3892 }
3893
3894 static int compare_thresholds(const void *a, const void *b)
3895 {
3896         const struct mem_cgroup_threshold *_a = a;
3897         const struct mem_cgroup_threshold *_b = b;
3898
3899         return _a->threshold - _b->threshold;
3900 }
3901
3902 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3903 {
3904         struct mem_cgroup_eventfd_list *ev;
3905
3906         list_for_each_entry(ev, &mem->oom_notify, list)
3907                 eventfd_signal(ev->eventfd, 1);
3908         return 0;
3909 }
3910
3911 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3912 {
3913         struct mem_cgroup *iter;
3914
3915         for_each_mem_cgroup_tree(iter, mem)
3916                 mem_cgroup_oom_notify_cb(iter);
3917 }
3918
3919 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3920         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3921 {
3922         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3923         struct mem_cgroup_thresholds *thresholds;
3924         struct mem_cgroup_threshold_ary *new;
3925         int type = MEMFILE_TYPE(cft->private);
3926         u64 threshold, usage;
3927         int i, size, ret;
3928
3929         ret = res_counter_memparse_write_strategy(args, &threshold);
3930         if (ret)
3931                 return ret;
3932
3933         mutex_lock(&memcg->thresholds_lock);
3934
3935         if (type == _MEM)
3936                 thresholds = &memcg->thresholds;
3937         else if (type == _MEMSWAP)
3938                 thresholds = &memcg->memsw_thresholds;
3939         else
3940                 BUG();
3941
3942         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3943
3944         /* Check if a threshold crossed before adding a new one */
3945         if (thresholds->primary)
3946                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3947
3948         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3949
3950         /* Allocate memory for new array of thresholds */
3951         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3952                         GFP_KERNEL);
3953         if (!new) {
3954                 ret = -ENOMEM;
3955                 goto unlock;
3956         }
3957         new->size = size;
3958
3959         /* Copy thresholds (if any) to new array */
3960         if (thresholds->primary) {
3961                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3962                                 sizeof(struct mem_cgroup_threshold));
3963         }
3964
3965         /* Add new threshold */
3966         new->entries[size - 1].eventfd = eventfd;
3967         new->entries[size - 1].threshold = threshold;
3968
3969         /* Sort thresholds. Registering of new threshold isn't time-critical */
3970         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3971                         compare_thresholds, NULL);
3972
3973         /* Find current threshold */
3974         new->current_threshold = -1;
3975         for (i = 0; i < size; i++) {
3976                 if (new->entries[i].threshold < usage) {
3977                         /*
3978                          * new->current_threshold will not be used until
3979                          * rcu_assign_pointer(), so it's safe to increment
3980                          * it here.
3981                          */
3982                         ++new->current_threshold;
3983                 }
3984         }
3985
3986         /* Free old spare buffer and save old primary buffer as spare */
3987         kfree(thresholds->spare);
3988         thresholds->spare = thresholds->primary;
3989
3990         rcu_assign_pointer(thresholds->primary, new);
3991
3992         /* To be sure that nobody uses thresholds */
3993         synchronize_rcu();
3994
3995 unlock:
3996         mutex_unlock(&memcg->thresholds_lock);
3997
3998         return ret;
3999 }
4000
4001 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4002         struct cftype *cft, struct eventfd_ctx *eventfd)
4003 {
4004         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4005         struct mem_cgroup_thresholds *thresholds;
4006         struct mem_cgroup_threshold_ary *new;
4007         int type = MEMFILE_TYPE(cft->private);
4008         u64 usage;
4009         int i, j, size;
4010
4011         mutex_lock(&memcg->thresholds_lock);
4012         if (type == _MEM)
4013                 thresholds = &memcg->thresholds;
4014         else if (type == _MEMSWAP)
4015                 thresholds = &memcg->memsw_thresholds;
4016         else
4017                 BUG();
4018
4019         /*
4020          * Something went wrong if we trying to unregister a threshold
4021          * if we don't have thresholds
4022          */
4023         BUG_ON(!thresholds);
4024
4025         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4026
4027         /* Check if a threshold crossed before removing */
4028         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4029
4030         /* Calculate new number of threshold */
4031         size = 0;
4032         for (i = 0; i < thresholds->primary->size; i++) {
4033                 if (thresholds->primary->entries[i].eventfd != eventfd)
4034                         size++;
4035         }
4036
4037         new = thresholds->spare;
4038
4039         /* Set thresholds array to NULL if we don't have thresholds */
4040         if (!size) {
4041                 kfree(new);
4042                 new = NULL;
4043                 goto swap_buffers;
4044         }
4045
4046         new->size = size;
4047
4048         /* Copy thresholds and find current threshold */
4049         new->current_threshold = -1;
4050         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4051                 if (thresholds->primary->entries[i].eventfd == eventfd)
4052                         continue;
4053
4054                 new->entries[j] = thresholds->primary->entries[i];
4055                 if (new->entries[j].threshold < usage) {
4056                         /*
4057                          * new->current_threshold will not be used
4058                          * until rcu_assign_pointer(), so it's safe to increment
4059                          * it here.
4060                          */
4061                         ++new->current_threshold;
4062                 }
4063                 j++;
4064         }
4065
4066 swap_buffers:
4067         /* Swap primary and spare array */
4068         thresholds->spare = thresholds->primary;
4069         rcu_assign_pointer(thresholds->primary, new);
4070
4071         /* To be sure that nobody uses thresholds */
4072         synchronize_rcu();
4073
4074         mutex_unlock(&memcg->thresholds_lock);
4075 }
4076
4077 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4078         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4079 {
4080         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4081         struct mem_cgroup_eventfd_list *event;
4082         int type = MEMFILE_TYPE(cft->private);
4083
4084         BUG_ON(type != _OOM_TYPE);
4085         event = kmalloc(sizeof(*event), GFP_KERNEL);
4086         if (!event)
4087                 return -ENOMEM;
4088
4089         mutex_lock(&memcg_oom_mutex);
4090
4091         event->eventfd = eventfd;
4092         list_add(&event->list, &memcg->oom_notify);
4093
4094         /* already in OOM ? */
4095         if (atomic_read(&memcg->oom_lock))
4096                 eventfd_signal(eventfd, 1);
4097         mutex_unlock(&memcg_oom_mutex);
4098
4099         return 0;
4100 }
4101
4102 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4103         struct cftype *cft, struct eventfd_ctx *eventfd)
4104 {
4105         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4106         struct mem_cgroup_eventfd_list *ev, *tmp;
4107         int type = MEMFILE_TYPE(cft->private);
4108
4109         BUG_ON(type != _OOM_TYPE);
4110
4111         mutex_lock(&memcg_oom_mutex);
4112
4113         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4114                 if (ev->eventfd == eventfd) {
4115                         list_del(&ev->list);
4116                         kfree(ev);
4117                 }
4118         }
4119
4120         mutex_unlock(&memcg_oom_mutex);
4121 }
4122
4123 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4124         struct cftype *cft,  struct cgroup_map_cb *cb)
4125 {
4126         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4127
4128         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4129
4130         if (atomic_read(&mem->oom_lock))
4131                 cb->fill(cb, "under_oom", 1);
4132         else
4133                 cb->fill(cb, "under_oom", 0);
4134         return 0;
4135 }
4136
4137 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4138         struct cftype *cft, u64 val)
4139 {
4140         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4141         struct mem_cgroup *parent;
4142
4143         /* cannot set to root cgroup and only 0 and 1 are allowed */
4144         if (!cgrp->parent || !((val == 0) || (val == 1)))
4145                 return -EINVAL;
4146
4147         parent = mem_cgroup_from_cont(cgrp->parent);
4148
4149         cgroup_lock();
4150         /* oom-kill-disable is a flag for subhierarchy. */
4151         if ((parent->use_hierarchy) ||
4152             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4153                 cgroup_unlock();
4154                 return -EINVAL;
4155         }
4156         mem->oom_kill_disable = val;
4157         if (!val)
4158                 memcg_oom_recover(mem);
4159         cgroup_unlock();
4160         return 0;
4161 }
4162
4163 static struct cftype mem_cgroup_files[] = {
4164         {
4165                 .name = "usage_in_bytes",
4166                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4167                 .read_u64 = mem_cgroup_read,
4168                 .register_event = mem_cgroup_usage_register_event,
4169                 .unregister_event = mem_cgroup_usage_unregister_event,
4170         },
4171         {
4172                 .name = "max_usage_in_bytes",
4173                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4174                 .trigger = mem_cgroup_reset,
4175                 .read_u64 = mem_cgroup_read,
4176         },
4177         {
4178                 .name = "limit_in_bytes",
4179                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4180                 .write_string = mem_cgroup_write,
4181                 .read_u64 = mem_cgroup_read,
4182         },
4183         {
4184                 .name = "soft_limit_in_bytes",
4185                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4186                 .write_string = mem_cgroup_write,
4187                 .read_u64 = mem_cgroup_read,
4188         },
4189         {
4190                 .name = "failcnt",
4191                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4192                 .trigger = mem_cgroup_reset,
4193                 .read_u64 = mem_cgroup_read,
4194         },
4195         {
4196                 .name = "stat",
4197                 .read_map = mem_control_stat_show,
4198         },
4199         {
4200                 .name = "force_empty",
4201                 .trigger = mem_cgroup_force_empty_write,
4202         },
4203         {
4204                 .name = "use_hierarchy",
4205                 .write_u64 = mem_cgroup_hierarchy_write,
4206                 .read_u64 = mem_cgroup_hierarchy_read,
4207         },
4208         {
4209                 .name = "swappiness",
4210                 .read_u64 = mem_cgroup_swappiness_read,
4211                 .write_u64 = mem_cgroup_swappiness_write,
4212         },
4213         {
4214                 .name = "move_charge_at_immigrate",
4215                 .read_u64 = mem_cgroup_move_charge_read,
4216                 .write_u64 = mem_cgroup_move_charge_write,
4217         },
4218         {
4219                 .name = "oom_control",
4220                 .read_map = mem_cgroup_oom_control_read,
4221                 .write_u64 = mem_cgroup_oom_control_write,
4222                 .register_event = mem_cgroup_oom_register_event,
4223                 .unregister_event = mem_cgroup_oom_unregister_event,
4224                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4225         },
4226 };
4227
4228 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4229 static struct cftype memsw_cgroup_files[] = {
4230         {
4231                 .name = "memsw.usage_in_bytes",
4232                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4233                 .read_u64 = mem_cgroup_read,
4234                 .register_event = mem_cgroup_usage_register_event,
4235                 .unregister_event = mem_cgroup_usage_unregister_event,
4236         },
4237         {
4238                 .name = "memsw.max_usage_in_bytes",
4239                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4240                 .trigger = mem_cgroup_reset,
4241                 .read_u64 = mem_cgroup_read,
4242         },
4243         {
4244                 .name = "memsw.limit_in_bytes",
4245                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4246                 .write_string = mem_cgroup_write,
4247                 .read_u64 = mem_cgroup_read,
4248         },
4249         {
4250                 .name = "memsw.failcnt",
4251                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4252                 .trigger = mem_cgroup_reset,
4253                 .read_u64 = mem_cgroup_read,
4254         },
4255 };
4256
4257 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4258 {
4259         if (!do_swap_account)
4260                 return 0;
4261         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4262                                 ARRAY_SIZE(memsw_cgroup_files));
4263 };
4264 #else
4265 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4266 {
4267         return 0;
4268 }
4269 #endif
4270
4271 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4272 {
4273         struct mem_cgroup_per_node *pn;
4274         struct mem_cgroup_per_zone *mz;
4275         enum lru_list l;
4276         int zone, tmp = node;
4277         /*
4278          * This routine is called against possible nodes.
4279          * But it's BUG to call kmalloc() against offline node.
4280          *
4281          * TODO: this routine can waste much memory for nodes which will
4282          *       never be onlined. It's better to use memory hotplug callback
4283          *       function.
4284          */
4285         if (!node_state(node, N_NORMAL_MEMORY))
4286                 tmp = -1;
4287         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4288         if (!pn)
4289                 return 1;
4290
4291         mem->info.nodeinfo[node] = pn;
4292         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4293                 mz = &pn->zoneinfo[zone];
4294                 for_each_lru(l)
4295                         INIT_LIST_HEAD(&mz->lists[l]);
4296                 mz->usage_in_excess = 0;
4297                 mz->on_tree = false;
4298                 mz->mem = mem;
4299         }
4300         return 0;
4301 }
4302
4303 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4304 {
4305         kfree(mem->info.nodeinfo[node]);
4306 }
4307
4308 static struct mem_cgroup *mem_cgroup_alloc(void)
4309 {
4310         struct mem_cgroup *mem;
4311         int size = sizeof(struct mem_cgroup);
4312
4313         /* Can be very big if MAX_NUMNODES is very big */
4314         if (size < PAGE_SIZE)
4315                 mem = kzalloc(size, GFP_KERNEL);
4316         else
4317                 mem = vzalloc(size);
4318
4319         if (!mem)
4320                 return NULL;
4321
4322         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4323         if (!mem->stat)
4324                 goto out_free;
4325         spin_lock_init(&mem->pcp_counter_lock);
4326         return mem;
4327
4328 out_free:
4329         if (size < PAGE_SIZE)
4330                 kfree(mem);
4331         else
4332                 vfree(mem);
4333         return NULL;
4334 }
4335
4336 /*
4337  * At destroying mem_cgroup, references from swap_cgroup can remain.
4338  * (scanning all at force_empty is too costly...)
4339  *
4340  * Instead of clearing all references at force_empty, we remember
4341  * the number of reference from swap_cgroup and free mem_cgroup when
4342  * it goes down to 0.
4343  *
4344  * Removal of cgroup itself succeeds regardless of refs from swap.
4345  */
4346
4347 static void __mem_cgroup_free(struct mem_cgroup *mem)
4348 {
4349         int node;
4350
4351         mem_cgroup_remove_from_trees(mem);
4352         free_css_id(&mem_cgroup_subsys, &mem->css);
4353
4354         for_each_node_state(node, N_POSSIBLE)
4355                 free_mem_cgroup_per_zone_info(mem, node);
4356
4357         free_percpu(mem->stat);
4358         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4359                 kfree(mem);
4360         else
4361                 vfree(mem);
4362 }
4363
4364 static void mem_cgroup_get(struct mem_cgroup *mem)
4365 {
4366         atomic_inc(&mem->refcnt);
4367 }
4368
4369 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4370 {
4371         if (atomic_sub_and_test(count, &mem->refcnt)) {
4372                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4373                 __mem_cgroup_free(mem);
4374                 if (parent)
4375                         mem_cgroup_put(parent);
4376         }
4377 }
4378
4379 static void mem_cgroup_put(struct mem_cgroup *mem)
4380 {
4381         __mem_cgroup_put(mem, 1);
4382 }
4383
4384 /*
4385  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4386  */
4387 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4388 {
4389         if (!mem->res.parent)
4390                 return NULL;
4391         return mem_cgroup_from_res_counter(mem->res.parent, res);
4392 }
4393
4394 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4395 static void __init enable_swap_cgroup(void)
4396 {
4397         if (!mem_cgroup_disabled() && really_do_swap_account)
4398                 do_swap_account = 1;
4399 }
4400 #else
4401 static void __init enable_swap_cgroup(void)
4402 {
4403 }
4404 #endif
4405
4406 static int mem_cgroup_soft_limit_tree_init(void)
4407 {
4408         struct mem_cgroup_tree_per_node *rtpn;
4409         struct mem_cgroup_tree_per_zone *rtpz;
4410         int tmp, node, zone;
4411
4412         for_each_node_state(node, N_POSSIBLE) {
4413                 tmp = node;
4414                 if (!node_state(node, N_NORMAL_MEMORY))
4415                         tmp = -1;
4416                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4417                 if (!rtpn)
4418                         return 1;
4419
4420                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4421
4422                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4423                         rtpz = &rtpn->rb_tree_per_zone[zone];
4424                         rtpz->rb_root = RB_ROOT;
4425                         spin_lock_init(&rtpz->lock);
4426                 }
4427         }
4428         return 0;
4429 }
4430
4431 static struct cgroup_subsys_state * __ref
4432 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4433 {
4434         struct mem_cgroup *mem, *parent;
4435         long error = -ENOMEM;
4436         int node;
4437
4438         mem = mem_cgroup_alloc();
4439         if (!mem)
4440                 return ERR_PTR(error);
4441
4442         for_each_node_state(node, N_POSSIBLE)
4443                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4444                         goto free_out;
4445
4446         /* root ? */
4447         if (cont->parent == NULL) {
4448                 int cpu;
4449                 enable_swap_cgroup();
4450                 parent = NULL;
4451                 root_mem_cgroup = mem;
4452                 if (mem_cgroup_soft_limit_tree_init())
4453                         goto free_out;
4454                 for_each_possible_cpu(cpu) {
4455                         struct memcg_stock_pcp *stock =
4456                                                 &per_cpu(memcg_stock, cpu);
4457                         INIT_WORK(&stock->work, drain_local_stock);
4458                 }
4459                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4460         } else {
4461                 parent = mem_cgroup_from_cont(cont->parent);
4462                 mem->use_hierarchy = parent->use_hierarchy;
4463                 mem->oom_kill_disable = parent->oom_kill_disable;
4464         }
4465
4466         if (parent && parent->use_hierarchy) {
4467                 res_counter_init(&mem->res, &parent->res);
4468                 res_counter_init(&mem->memsw, &parent->memsw);
4469                 /*
4470                  * We increment refcnt of the parent to ensure that we can
4471                  * safely access it on res_counter_charge/uncharge.
4472                  * This refcnt will be decremented when freeing this
4473                  * mem_cgroup(see mem_cgroup_put).
4474                  */
4475                 mem_cgroup_get(parent);
4476         } else {
4477                 res_counter_init(&mem->res, NULL);
4478                 res_counter_init(&mem->memsw, NULL);
4479         }
4480         mem->last_scanned_child = 0;
4481         spin_lock_init(&mem->reclaim_param_lock);
4482         INIT_LIST_HEAD(&mem->oom_notify);
4483
4484         if (parent)
4485                 mem->swappiness = get_swappiness(parent);
4486         atomic_set(&mem->refcnt, 1);
4487         mem->move_charge_at_immigrate = 0;
4488         mutex_init(&mem->thresholds_lock);
4489         return &mem->css;
4490 free_out:
4491         __mem_cgroup_free(mem);
4492         root_mem_cgroup = NULL;
4493         return ERR_PTR(error);
4494 }
4495
4496 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4497                                         struct cgroup *cont)
4498 {
4499         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4500
4501         return mem_cgroup_force_empty(mem, false);
4502 }
4503
4504 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4505                                 struct cgroup *cont)
4506 {
4507         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4508
4509         mem_cgroup_put(mem);
4510 }
4511
4512 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4513                                 struct cgroup *cont)
4514 {
4515         int ret;
4516
4517         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4518                                 ARRAY_SIZE(mem_cgroup_files));
4519
4520         if (!ret)
4521                 ret = register_memsw_files(cont, ss);
4522         return ret;
4523 }
4524
4525 #ifdef CONFIG_MMU
4526 /* Handlers for move charge at task migration. */
4527 #define PRECHARGE_COUNT_AT_ONCE 256
4528 static int mem_cgroup_do_precharge(unsigned long count)
4529 {
4530         int ret = 0;
4531         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4532         struct mem_cgroup *mem = mc.to;
4533
4534         if (mem_cgroup_is_root(mem)) {
4535                 mc.precharge += count;
4536                 /* we don't need css_get for root */
4537                 return ret;
4538         }
4539         /* try to charge at once */
4540         if (count > 1) {
4541                 struct res_counter *dummy;
4542                 /*
4543                  * "mem" cannot be under rmdir() because we've already checked
4544                  * by cgroup_lock_live_cgroup() that it is not removed and we
4545                  * are still under the same cgroup_mutex. So we can postpone
4546                  * css_get().
4547                  */
4548                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4549                         goto one_by_one;
4550                 if (do_swap_account && res_counter_charge(&mem->memsw,
4551                                                 PAGE_SIZE * count, &dummy)) {
4552                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4553                         goto one_by_one;
4554                 }
4555                 mc.precharge += count;
4556                 return ret;
4557         }
4558 one_by_one:
4559         /* fall back to one by one charge */
4560         while (count--) {
4561                 if (signal_pending(current)) {
4562                         ret = -EINTR;
4563                         break;
4564                 }
4565                 if (!batch_count--) {
4566                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4567                         cond_resched();
4568                 }
4569                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4570                                               PAGE_SIZE);
4571                 if (ret || !mem)
4572                         /* mem_cgroup_clear_mc() will do uncharge later */
4573                         return -ENOMEM;
4574                 mc.precharge++;
4575         }
4576         return ret;
4577 }
4578
4579 /**
4580  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4581  * @vma: the vma the pte to be checked belongs
4582  * @addr: the address corresponding to the pte to be checked
4583  * @ptent: the pte to be checked
4584  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4585  *
4586  * Returns
4587  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4588  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4589  *     move charge. if @target is not NULL, the page is stored in target->page
4590  *     with extra refcnt got(Callers should handle it).
4591  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4592  *     target for charge migration. if @target is not NULL, the entry is stored
4593  *     in target->ent.
4594  *
4595  * Called with pte lock held.
4596  */
4597 union mc_target {
4598         struct page     *page;
4599         swp_entry_t     ent;
4600 };
4601
4602 enum mc_target_type {
4603         MC_TARGET_NONE, /* not used */
4604         MC_TARGET_PAGE,
4605         MC_TARGET_SWAP,
4606 };
4607
4608 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4609                                                 unsigned long addr, pte_t ptent)
4610 {
4611         struct page *page = vm_normal_page(vma, addr, ptent);
4612
4613         if (!page || !page_mapped(page))
4614                 return NULL;
4615         if (PageAnon(page)) {
4616                 /* we don't move shared anon */
4617                 if (!move_anon() || page_mapcount(page) > 2)
4618                         return NULL;
4619         } else if (!move_file())
4620                 /* we ignore mapcount for file pages */
4621                 return NULL;
4622         if (!get_page_unless_zero(page))
4623                 return NULL;
4624
4625         return page;
4626 }
4627
4628 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4629                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4630 {
4631         int usage_count;
4632         struct page *page = NULL;
4633         swp_entry_t ent = pte_to_swp_entry(ptent);
4634
4635         if (!move_anon() || non_swap_entry(ent))
4636                 return NULL;
4637         usage_count = mem_cgroup_count_swap_user(ent, &page);
4638         if (usage_count > 1) { /* we don't move shared anon */
4639                 if (page)
4640                         put_page(page);
4641                 return NULL;
4642         }
4643         if (do_swap_account)
4644                 entry->val = ent.val;
4645
4646         return page;
4647 }
4648
4649 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4650                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4651 {
4652         struct page *page = NULL;
4653         struct inode *inode;
4654         struct address_space *mapping;
4655         pgoff_t pgoff;
4656
4657         if (!vma->vm_file) /* anonymous vma */
4658                 return NULL;
4659         if (!move_file())
4660                 return NULL;
4661
4662         inode = vma->vm_file->f_path.dentry->d_inode;
4663         mapping = vma->vm_file->f_mapping;
4664         if (pte_none(ptent))
4665                 pgoff = linear_page_index(vma, addr);
4666         else /* pte_file(ptent) is true */
4667                 pgoff = pte_to_pgoff(ptent);
4668
4669         /* page is moved even if it's not RSS of this task(page-faulted). */
4670         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4671                 page = find_get_page(mapping, pgoff);
4672         } else { /* shmem/tmpfs file. we should take account of swap too. */
4673                 swp_entry_t ent;
4674                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4675                 if (do_swap_account)
4676                         entry->val = ent.val;
4677         }
4678
4679         return page;
4680 }
4681
4682 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4683                 unsigned long addr, pte_t ptent, union mc_target *target)
4684 {
4685         struct page *page = NULL;
4686         struct page_cgroup *pc;
4687         int ret = 0;
4688         swp_entry_t ent = { .val = 0 };
4689
4690         if (pte_present(ptent))
4691                 page = mc_handle_present_pte(vma, addr, ptent);
4692         else if (is_swap_pte(ptent))
4693                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4694         else if (pte_none(ptent) || pte_file(ptent))
4695                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4696
4697         if (!page && !ent.val)
4698                 return 0;
4699         if (page) {
4700                 pc = lookup_page_cgroup(page);
4701                 /*
4702                  * Do only loose check w/o page_cgroup lock.
4703                  * mem_cgroup_move_account() checks the pc is valid or not under
4704                  * the lock.
4705                  */
4706                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4707                         ret = MC_TARGET_PAGE;
4708                         if (target)
4709                                 target->page = page;
4710                 }
4711                 if (!ret || !target)
4712                         put_page(page);
4713         }
4714         /* There is a swap entry and a page doesn't exist or isn't charged */
4715         if (ent.val && !ret &&
4716                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4717                 ret = MC_TARGET_SWAP;
4718                 if (target)
4719                         target->ent = ent;
4720         }
4721         return ret;
4722 }
4723
4724 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4725                                         unsigned long addr, unsigned long end,
4726                                         struct mm_walk *walk)
4727 {
4728         struct vm_area_struct *vma = walk->private;
4729         pte_t *pte;
4730         spinlock_t *ptl;
4731
4732         VM_BUG_ON(pmd_trans_huge(*pmd));
4733         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4734         for (; addr != end; pte++, addr += PAGE_SIZE)
4735                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4736                         mc.precharge++; /* increment precharge temporarily */
4737         pte_unmap_unlock(pte - 1, ptl);
4738         cond_resched();
4739
4740         return 0;
4741 }
4742
4743 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4744 {
4745         unsigned long precharge;
4746         struct vm_area_struct *vma;
4747
4748         down_read(&mm->mmap_sem);
4749         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4750                 struct mm_walk mem_cgroup_count_precharge_walk = {
4751                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4752                         .mm = mm,
4753                         .private = vma,
4754                 };
4755                 if (is_vm_hugetlb_page(vma))
4756                         continue;
4757                 walk_page_range(vma->vm_start, vma->vm_end,
4758                                         &mem_cgroup_count_precharge_walk);
4759         }
4760         up_read(&mm->mmap_sem);
4761
4762         precharge = mc.precharge;
4763         mc.precharge = 0;
4764
4765         return precharge;
4766 }
4767
4768 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4769 {
4770         unsigned long precharge = mem_cgroup_count_precharge(mm);
4771
4772         VM_BUG_ON(mc.moving_task);
4773         mc.moving_task = current;
4774         return mem_cgroup_do_precharge(precharge);
4775 }
4776
4777 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4778 static void __mem_cgroup_clear_mc(void)
4779 {
4780         struct mem_cgroup *from = mc.from;
4781         struct mem_cgroup *to = mc.to;
4782
4783         /* we must uncharge all the leftover precharges from mc.to */
4784         if (mc.precharge) {
4785                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4786                 mc.precharge = 0;
4787         }
4788         /*
4789          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4790          * we must uncharge here.
4791          */
4792         if (mc.moved_charge) {
4793                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4794                 mc.moved_charge = 0;
4795         }
4796         /* we must fixup refcnts and charges */
4797         if (mc.moved_swap) {
4798                 /* uncharge swap account from the old cgroup */
4799                 if (!mem_cgroup_is_root(mc.from))
4800                         res_counter_uncharge(&mc.from->memsw,
4801                                                 PAGE_SIZE * mc.moved_swap);
4802                 __mem_cgroup_put(mc.from, mc.moved_swap);
4803
4804                 if (!mem_cgroup_is_root(mc.to)) {
4805                         /*
4806                          * we charged both to->res and to->memsw, so we should
4807                          * uncharge to->res.
4808                          */
4809                         res_counter_uncharge(&mc.to->res,
4810                                                 PAGE_SIZE * mc.moved_swap);
4811                 }
4812                 /* we've already done mem_cgroup_get(mc.to) */
4813                 mc.moved_swap = 0;
4814         }
4815         memcg_oom_recover(from);
4816         memcg_oom_recover(to);
4817         wake_up_all(&mc.waitq);
4818 }
4819
4820 static void mem_cgroup_clear_mc(void)
4821 {
4822         struct mem_cgroup *from = mc.from;
4823
4824         /*
4825          * we must clear moving_task before waking up waiters at the end of
4826          * task migration.
4827          */
4828         mc.moving_task = NULL;
4829         __mem_cgroup_clear_mc();
4830         spin_lock(&mc.lock);
4831         mc.from = NULL;
4832         mc.to = NULL;
4833         spin_unlock(&mc.lock);
4834         mem_cgroup_end_move(from);
4835 }
4836
4837 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4838                                 struct cgroup *cgroup,
4839                                 struct task_struct *p,
4840                                 bool threadgroup)
4841 {
4842         int ret = 0;
4843         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4844
4845         if (mem->move_charge_at_immigrate) {
4846                 struct mm_struct *mm;
4847                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4848
4849                 VM_BUG_ON(from == mem);
4850
4851                 mm = get_task_mm(p);
4852                 if (!mm)
4853                         return 0;
4854                 /* We move charges only when we move a owner of the mm */
4855                 if (mm->owner == p) {
4856                         VM_BUG_ON(mc.from);
4857                         VM_BUG_ON(mc.to);
4858                         VM_BUG_ON(mc.precharge);
4859                         VM_BUG_ON(mc.moved_charge);
4860                         VM_BUG_ON(mc.moved_swap);
4861                         mem_cgroup_start_move(from);
4862                         spin_lock(&mc.lock);
4863                         mc.from = from;
4864                         mc.to = mem;
4865                         spin_unlock(&mc.lock);
4866                         /* We set mc.moving_task later */
4867
4868                         ret = mem_cgroup_precharge_mc(mm);
4869                         if (ret)
4870                                 mem_cgroup_clear_mc();
4871                 }
4872                 mmput(mm);
4873         }
4874         return ret;
4875 }
4876
4877 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4878                                 struct cgroup *cgroup,
4879                                 struct task_struct *p,
4880                                 bool threadgroup)
4881 {
4882         mem_cgroup_clear_mc();
4883 }
4884
4885 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4886                                 unsigned long addr, unsigned long end,
4887                                 struct mm_walk *walk)
4888 {
4889         int ret = 0;
4890         struct vm_area_struct *vma = walk->private;
4891         pte_t *pte;
4892         spinlock_t *ptl;
4893
4894 retry:
4895         VM_BUG_ON(pmd_trans_huge(*pmd));
4896         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4897         for (; addr != end; addr += PAGE_SIZE) {
4898                 pte_t ptent = *(pte++);
4899                 union mc_target target;
4900                 int type;
4901                 struct page *page;
4902                 struct page_cgroup *pc;
4903                 swp_entry_t ent;
4904
4905                 if (!mc.precharge)
4906                         break;
4907
4908                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4909                 switch (type) {
4910                 case MC_TARGET_PAGE:
4911                         page = target.page;
4912                         if (isolate_lru_page(page))
4913                                 goto put;
4914                         pc = lookup_page_cgroup(page);
4915                         if (!mem_cgroup_move_account(pc,
4916                                         mc.from, mc.to, false, PAGE_SIZE)) {
4917                                 mc.precharge--;
4918                                 /* we uncharge from mc.from later. */
4919                                 mc.moved_charge++;
4920                         }
4921                         putback_lru_page(page);
4922 put:                    /* is_target_pte_for_mc() gets the page */
4923                         put_page(page);
4924                         break;
4925                 case MC_TARGET_SWAP:
4926                         ent = target.ent;
4927                         if (!mem_cgroup_move_swap_account(ent,
4928                                                 mc.from, mc.to, false)) {
4929                                 mc.precharge--;
4930                                 /* we fixup refcnts and charges later. */
4931                                 mc.moved_swap++;
4932                         }
4933                         break;
4934                 default:
4935                         break;
4936                 }
4937         }
4938         pte_unmap_unlock(pte - 1, ptl);
4939         cond_resched();
4940
4941         if (addr != end) {
4942                 /*
4943                  * We have consumed all precharges we got in can_attach().
4944                  * We try charge one by one, but don't do any additional
4945                  * charges to mc.to if we have failed in charge once in attach()
4946                  * phase.
4947                  */
4948                 ret = mem_cgroup_do_precharge(1);
4949                 if (!ret)
4950                         goto retry;
4951         }
4952
4953         return ret;
4954 }
4955
4956 static void mem_cgroup_move_charge(struct mm_struct *mm)
4957 {
4958         struct vm_area_struct *vma;
4959
4960         lru_add_drain_all();
4961 retry:
4962         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4963                 /*
4964                  * Someone who are holding the mmap_sem might be waiting in
4965                  * waitq. So we cancel all extra charges, wake up all waiters,
4966                  * and retry. Because we cancel precharges, we might not be able
4967                  * to move enough charges, but moving charge is a best-effort
4968                  * feature anyway, so it wouldn't be a big problem.
4969                  */
4970                 __mem_cgroup_clear_mc();
4971                 cond_resched();
4972                 goto retry;
4973         }
4974         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4975                 int ret;
4976                 struct mm_walk mem_cgroup_move_charge_walk = {
4977                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4978                         .mm = mm,
4979                         .private = vma,
4980                 };
4981                 if (is_vm_hugetlb_page(vma))
4982                         continue;
4983                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4984                                                 &mem_cgroup_move_charge_walk);
4985                 if (ret)
4986                         /*
4987                          * means we have consumed all precharges and failed in
4988                          * doing additional charge. Just abandon here.
4989                          */
4990                         break;
4991         }
4992         up_read(&mm->mmap_sem);
4993 }
4994
4995 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4996                                 struct cgroup *cont,
4997                                 struct cgroup *old_cont,
4998                                 struct task_struct *p,
4999                                 bool threadgroup)
5000 {
5001         struct mm_struct *mm;
5002
5003         if (!mc.to)
5004                 /* no need to move charge */
5005                 return;
5006
5007         mm = get_task_mm(p);
5008         if (mm) {
5009                 mem_cgroup_move_charge(mm);
5010                 mmput(mm);
5011         }
5012         mem_cgroup_clear_mc();
5013 }
5014 #else   /* !CONFIG_MMU */
5015 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5016                                 struct cgroup *cgroup,
5017                                 struct task_struct *p,
5018                                 bool threadgroup)
5019 {
5020         return 0;
5021 }
5022 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5023                                 struct cgroup *cgroup,
5024                                 struct task_struct *p,
5025                                 bool threadgroup)
5026 {
5027 }
5028 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5029                                 struct cgroup *cont,
5030                                 struct cgroup *old_cont,
5031                                 struct task_struct *p,
5032                                 bool threadgroup)
5033 {
5034 }
5035 #endif
5036
5037 struct cgroup_subsys mem_cgroup_subsys = {
5038         .name = "memory",
5039         .subsys_id = mem_cgroup_subsys_id,
5040         .create = mem_cgroup_create,
5041         .pre_destroy = mem_cgroup_pre_destroy,
5042         .destroy = mem_cgroup_destroy,
5043         .populate = mem_cgroup_populate,
5044         .can_attach = mem_cgroup_can_attach,
5045         .cancel_attach = mem_cgroup_cancel_attach,
5046         .attach = mem_cgroup_move_task,
5047         .early_init = 0,
5048         .use_id = 1,
5049 };
5050
5051 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5052 static int __init enable_swap_account(char *s)
5053 {
5054         /* consider enabled if no parameter or 1 is given */
5055         if (!(*s) || !strcmp(s, "=1"))
5056                 really_do_swap_account = 1;
5057         else if (!strcmp(s, "=0"))
5058                 really_do_swap_account = 0;
5059         return 1;
5060 }
5061 __setup("swapaccount", enable_swap_account);
5062
5063 static int __init disable_swap_account(char *s)
5064 {
5065         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5066         enable_swap_account("=0");
5067         return 1;
5068 }
5069 __setup("noswapaccount", disable_swap_account);
5070 #endif