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