slub: Fix slub_lock down/up imbalance
[linux-3.10.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 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
65 #else
66 #define do_swap_account         (0)
67 #endif
68
69 /*
70  * Per memcg event counter is incremented at every pagein/pageout. This counter
71  * is used for trigger some periodic events. This is straightforward and better
72  * than using jiffies etc. to handle periodic memcg event.
73  *
74  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
75  */
76 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
77 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
78
79 /*
80  * Statistics for memory cgroup.
81  */
82 enum mem_cgroup_stat_index {
83         /*
84          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85          */
86         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
87         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
88         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
89         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
90         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
91         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
92         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
93         /* incremented at every  pagein/pageout */
94         MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
95         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
96
97         MEM_CGROUP_STAT_NSTATS,
98 };
99
100 struct mem_cgroup_stat_cpu {
101         s64 count[MEM_CGROUP_STAT_NSTATS];
102 };
103
104 /*
105  * per-zone information in memory controller.
106  */
107 struct mem_cgroup_per_zone {
108         /*
109          * spin_lock to protect the per cgroup LRU
110          */
111         struct list_head        lists[NR_LRU_LISTS];
112         unsigned long           count[NR_LRU_LISTS];
113
114         struct zone_reclaim_stat reclaim_stat;
115         struct rb_node          tree_node;      /* RB tree node */
116         unsigned long long      usage_in_excess;/* Set to the value by which */
117                                                 /* the soft limit is exceeded*/
118         bool                    on_tree;
119         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
120                                                 /* use container_of        */
121 };
122 /* Macro for accessing counter */
123 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
124
125 struct mem_cgroup_per_node {
126         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
127 };
128
129 struct mem_cgroup_lru_info {
130         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
131 };
132
133 /*
134  * Cgroups above their limits are maintained in a RB-Tree, independent of
135  * their hierarchy representation
136  */
137
138 struct mem_cgroup_tree_per_zone {
139         struct rb_root rb_root;
140         spinlock_t lock;
141 };
142
143 struct mem_cgroup_tree_per_node {
144         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
145 };
146
147 struct mem_cgroup_tree {
148         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
149 };
150
151 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
152
153 struct mem_cgroup_threshold {
154         struct eventfd_ctx *eventfd;
155         u64 threshold;
156 };
157
158 /* For threshold */
159 struct mem_cgroup_threshold_ary {
160         /* An array index points to threshold just below usage. */
161         int current_threshold;
162         /* Size of entries[] */
163         unsigned int size;
164         /* Array of thresholds */
165         struct mem_cgroup_threshold entries[0];
166 };
167
168 struct mem_cgroup_thresholds {
169         /* Primary thresholds array */
170         struct mem_cgroup_threshold_ary *primary;
171         /*
172          * Spare threshold array.
173          * This is needed to make mem_cgroup_unregister_event() "never fail".
174          * It must be able to store at least primary->size - 1 entries.
175          */
176         struct mem_cgroup_threshold_ary *spare;
177 };
178
179 /* for OOM */
180 struct mem_cgroup_eventfd_list {
181         struct list_head list;
182         struct eventfd_ctx *eventfd;
183 };
184
185 static void mem_cgroup_threshold(struct mem_cgroup *mem);
186 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
187
188 /*
189  * The memory controller data structure. The memory controller controls both
190  * page cache and RSS per cgroup. We would eventually like to provide
191  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
192  * to help the administrator determine what knobs to tune.
193  *
194  * TODO: Add a water mark for the memory controller. Reclaim will begin when
195  * we hit the water mark. May be even add a low water mark, such that
196  * no reclaim occurs from a cgroup at it's low water mark, this is
197  * a feature that will be implemented much later in the future.
198  */
199 struct mem_cgroup {
200         struct cgroup_subsys_state css;
201         /*
202          * the counter to account for memory usage
203          */
204         struct res_counter res;
205         /*
206          * the counter to account for mem+swap usage.
207          */
208         struct res_counter memsw;
209         /*
210          * Per cgroup active and inactive list, similar to the
211          * per zone LRU lists.
212          */
213         struct mem_cgroup_lru_info info;
214
215         /*
216           protect against reclaim related member.
217         */
218         spinlock_t reclaim_param_lock;
219
220         /*
221          * While reclaiming in a hierarchy, we cache the last child we
222          * reclaimed from.
223          */
224         int last_scanned_child;
225         /*
226          * Should the accounting and control be hierarchical, per subtree?
227          */
228         bool use_hierarchy;
229         atomic_t        oom_lock;
230         atomic_t        refcnt;
231
232         unsigned int    swappiness;
233         /* OOM-Killer disable */
234         int             oom_kill_disable;
235
236         /* set when res.limit == memsw.limit */
237         bool            memsw_is_minimum;
238
239         /* protect arrays of thresholds */
240         struct mutex thresholds_lock;
241
242         /* thresholds for memory usage. RCU-protected */
243         struct mem_cgroup_thresholds thresholds;
244
245         /* thresholds for mem+swap usage. RCU-protected */
246         struct mem_cgroup_thresholds memsw_thresholds;
247
248         /* For oom notifier event fd */
249         struct list_head oom_notify;
250
251         /*
252          * Should we move charges of a task when a task is moved into this
253          * mem_cgroup ? And what type of charges should we move ?
254          */
255         unsigned long   move_charge_at_immigrate;
256         /*
257          * percpu counter.
258          */
259         struct mem_cgroup_stat_cpu *stat;
260         /*
261          * used when a cpu is offlined or other synchronizations
262          * See mem_cgroup_read_stat().
263          */
264         struct mem_cgroup_stat_cpu nocpu_base;
265         spinlock_t pcp_counter_lock;
266 };
267
268 /* Stuffs for move charges at task migration. */
269 /*
270  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
271  * left-shifted bitmap of these types.
272  */
273 enum move_type {
274         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
275         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
276         NR_MOVE_TYPE,
277 };
278
279 /* "mc" and its members are protected by cgroup_mutex */
280 static struct move_charge_struct {
281         spinlock_t        lock; /* for from, to, moving_task */
282         struct mem_cgroup *from;
283         struct mem_cgroup *to;
284         unsigned long precharge;
285         unsigned long moved_charge;
286         unsigned long moved_swap;
287         struct task_struct *moving_task;        /* a task moving charges */
288         wait_queue_head_t waitq;                /* a waitq for other context */
289 } mc = {
290         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
291         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
292 };
293
294 static bool move_anon(void)
295 {
296         return test_bit(MOVE_CHARGE_TYPE_ANON,
297                                         &mc.to->move_charge_at_immigrate);
298 }
299
300 static bool move_file(void)
301 {
302         return test_bit(MOVE_CHARGE_TYPE_FILE,
303                                         &mc.to->move_charge_at_immigrate);
304 }
305
306 /*
307  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
308  * limit reclaim to prevent infinite loops, if they ever occur.
309  */
310 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
311 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
312
313 enum charge_type {
314         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
315         MEM_CGROUP_CHARGE_TYPE_MAPPED,
316         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
317         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
318         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
319         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
320         NR_CHARGE_TYPE,
321 };
322
323 /* only for here (for easy reading.) */
324 #define PCGF_CACHE      (1UL << PCG_CACHE)
325 #define PCGF_USED       (1UL << PCG_USED)
326 #define PCGF_LOCK       (1UL << PCG_LOCK)
327 /* Not used, but added here for completeness */
328 #define PCGF_ACCT       (1UL << PCG_ACCT)
329
330 /* for encoding cft->private value on file */
331 #define _MEM                    (0)
332 #define _MEMSWAP                (1)
333 #define _OOM_TYPE               (2)
334 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
335 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
336 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
337 /* Used for OOM nofiier */
338 #define OOM_CONTROL             (0)
339
340 /*
341  * Reclaim flags for mem_cgroup_hierarchical_reclaim
342  */
343 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
344 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
345 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
346 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
347 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
348 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
349
350 static void mem_cgroup_get(struct mem_cgroup *mem);
351 static void mem_cgroup_put(struct mem_cgroup *mem);
352 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
353 static void drain_all_stock_async(void);
354
355 static struct mem_cgroup_per_zone *
356 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
357 {
358         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
359 }
360
361 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
362 {
363         return &mem->css;
364 }
365
366 static struct mem_cgroup_per_zone *
367 page_cgroup_zoneinfo(struct page_cgroup *pc)
368 {
369         struct mem_cgroup *mem = pc->mem_cgroup;
370         int nid = page_cgroup_nid(pc);
371         int zid = page_cgroup_zid(pc);
372
373         if (!mem)
374                 return NULL;
375
376         return mem_cgroup_zoneinfo(mem, nid, zid);
377 }
378
379 static struct mem_cgroup_tree_per_zone *
380 soft_limit_tree_node_zone(int nid, int zid)
381 {
382         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
383 }
384
385 static struct mem_cgroup_tree_per_zone *
386 soft_limit_tree_from_page(struct page *page)
387 {
388         int nid = page_to_nid(page);
389         int zid = page_zonenum(page);
390
391         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
392 }
393
394 static void
395 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
396                                 struct mem_cgroup_per_zone *mz,
397                                 struct mem_cgroup_tree_per_zone *mctz,
398                                 unsigned long long new_usage_in_excess)
399 {
400         struct rb_node **p = &mctz->rb_root.rb_node;
401         struct rb_node *parent = NULL;
402         struct mem_cgroup_per_zone *mz_node;
403
404         if (mz->on_tree)
405                 return;
406
407         mz->usage_in_excess = new_usage_in_excess;
408         if (!mz->usage_in_excess)
409                 return;
410         while (*p) {
411                 parent = *p;
412                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
413                                         tree_node);
414                 if (mz->usage_in_excess < mz_node->usage_in_excess)
415                         p = &(*p)->rb_left;
416                 /*
417                  * We can't avoid mem cgroups that are over their soft
418                  * limit by the same amount
419                  */
420                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
421                         p = &(*p)->rb_right;
422         }
423         rb_link_node(&mz->tree_node, parent, p);
424         rb_insert_color(&mz->tree_node, &mctz->rb_root);
425         mz->on_tree = true;
426 }
427
428 static void
429 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
430                                 struct mem_cgroup_per_zone *mz,
431                                 struct mem_cgroup_tree_per_zone *mctz)
432 {
433         if (!mz->on_tree)
434                 return;
435         rb_erase(&mz->tree_node, &mctz->rb_root);
436         mz->on_tree = false;
437 }
438
439 static void
440 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
441                                 struct mem_cgroup_per_zone *mz,
442                                 struct mem_cgroup_tree_per_zone *mctz)
443 {
444         spin_lock(&mctz->lock);
445         __mem_cgroup_remove_exceeded(mem, mz, mctz);
446         spin_unlock(&mctz->lock);
447 }
448
449
450 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
451 {
452         unsigned long long excess;
453         struct mem_cgroup_per_zone *mz;
454         struct mem_cgroup_tree_per_zone *mctz;
455         int nid = page_to_nid(page);
456         int zid = page_zonenum(page);
457         mctz = soft_limit_tree_from_page(page);
458
459         /*
460          * Necessary to update all ancestors when hierarchy is used.
461          * because their event counter is not touched.
462          */
463         for (; mem; mem = parent_mem_cgroup(mem)) {
464                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
465                 excess = res_counter_soft_limit_excess(&mem->res);
466                 /*
467                  * We have to update the tree if mz is on RB-tree or
468                  * mem is over its softlimit.
469                  */
470                 if (excess || mz->on_tree) {
471                         spin_lock(&mctz->lock);
472                         /* if on-tree, remove it */
473                         if (mz->on_tree)
474                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
475                         /*
476                          * Insert again. mz->usage_in_excess will be updated.
477                          * If excess is 0, no tree ops.
478                          */
479                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
480                         spin_unlock(&mctz->lock);
481                 }
482         }
483 }
484
485 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
486 {
487         int node, zone;
488         struct mem_cgroup_per_zone *mz;
489         struct mem_cgroup_tree_per_zone *mctz;
490
491         for_each_node_state(node, N_POSSIBLE) {
492                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
493                         mz = mem_cgroup_zoneinfo(mem, node, zone);
494                         mctz = soft_limit_tree_node_zone(node, zone);
495                         mem_cgroup_remove_exceeded(mem, mz, mctz);
496                 }
497         }
498 }
499
500 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
501 {
502         return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
503 }
504
505 static struct mem_cgroup_per_zone *
506 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
507 {
508         struct rb_node *rightmost = NULL;
509         struct mem_cgroup_per_zone *mz;
510
511 retry:
512         mz = NULL;
513         rightmost = rb_last(&mctz->rb_root);
514         if (!rightmost)
515                 goto done;              /* Nothing to reclaim from */
516
517         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
518         /*
519          * Remove the node now but someone else can add it back,
520          * we will to add it back at the end of reclaim to its correct
521          * position in the tree.
522          */
523         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
524         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
525                 !css_tryget(&mz->mem->css))
526                 goto retry;
527 done:
528         return mz;
529 }
530
531 static struct mem_cgroup_per_zone *
532 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
533 {
534         struct mem_cgroup_per_zone *mz;
535
536         spin_lock(&mctz->lock);
537         mz = __mem_cgroup_largest_soft_limit_node(mctz);
538         spin_unlock(&mctz->lock);
539         return mz;
540 }
541
542 /*
543  * Implementation Note: reading percpu statistics for memcg.
544  *
545  * Both of vmstat[] and percpu_counter has threshold and do periodic
546  * synchronization to implement "quick" read. There are trade-off between
547  * reading cost and precision of value. Then, we may have a chance to implement
548  * a periodic synchronizion of counter in memcg's counter.
549  *
550  * But this _read() function is used for user interface now. The user accounts
551  * memory usage by memory cgroup and he _always_ requires exact value because
552  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
553  * have to visit all online cpus and make sum. So, for now, unnecessary
554  * synchronization is not implemented. (just implemented for cpu hotplug)
555  *
556  * If there are kernel internal actions which can make use of some not-exact
557  * value, and reading all cpu value can be performance bottleneck in some
558  * common workload, threashold and synchonization as vmstat[] should be
559  * implemented.
560  */
561 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
562                 enum mem_cgroup_stat_index idx)
563 {
564         int cpu;
565         s64 val = 0;
566
567         get_online_cpus();
568         for_each_online_cpu(cpu)
569                 val += per_cpu(mem->stat->count[idx], cpu);
570 #ifdef CONFIG_HOTPLUG_CPU
571         spin_lock(&mem->pcp_counter_lock);
572         val += mem->nocpu_base.count[idx];
573         spin_unlock(&mem->pcp_counter_lock);
574 #endif
575         put_online_cpus();
576         return val;
577 }
578
579 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
580 {
581         s64 ret;
582
583         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
584         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
585         return ret;
586 }
587
588 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
589                                          bool charge)
590 {
591         int val = (charge) ? 1 : -1;
592         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
593 }
594
595 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
596                                          struct page_cgroup *pc,
597                                          bool charge)
598 {
599         int val = (charge) ? 1 : -1;
600
601         preempt_disable();
602
603         if (PageCgroupCache(pc))
604                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
605         else
606                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
607
608         if (charge)
609                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
610         else
611                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
612         __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
613
614         preempt_enable();
615 }
616
617 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
618                                         enum lru_list idx)
619 {
620         int nid, zid;
621         struct mem_cgroup_per_zone *mz;
622         u64 total = 0;
623
624         for_each_online_node(nid)
625                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
626                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
627                         total += MEM_CGROUP_ZSTAT(mz, idx);
628                 }
629         return total;
630 }
631
632 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
633 {
634         s64 val;
635
636         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
637
638         return !(val & ((1 << event_mask_shift) - 1));
639 }
640
641 /*
642  * Check events in order.
643  *
644  */
645 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
646 {
647         /* threshold event is triggered in finer grain than soft limit */
648         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
649                 mem_cgroup_threshold(mem);
650                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
651                         mem_cgroup_update_tree(mem, page);
652         }
653 }
654
655 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
656 {
657         return container_of(cgroup_subsys_state(cont,
658                                 mem_cgroup_subsys_id), struct mem_cgroup,
659                                 css);
660 }
661
662 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
663 {
664         /*
665          * mm_update_next_owner() may clear mm->owner to NULL
666          * if it races with swapoff, page migration, etc.
667          * So this can be called with p == NULL.
668          */
669         if (unlikely(!p))
670                 return NULL;
671
672         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
673                                 struct mem_cgroup, css);
674 }
675
676 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
677 {
678         struct mem_cgroup *mem = NULL;
679
680         if (!mm)
681                 return NULL;
682         /*
683          * Because we have no locks, mm->owner's may be being moved to other
684          * cgroup. We use css_tryget() here even if this looks
685          * pessimistic (rather than adding locks here).
686          */
687         rcu_read_lock();
688         do {
689                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
690                 if (unlikely(!mem))
691                         break;
692         } while (!css_tryget(&mem->css));
693         rcu_read_unlock();
694         return mem;
695 }
696
697 /* The caller has to guarantee "mem" exists before calling this */
698 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
699 {
700         struct cgroup_subsys_state *css;
701         int found;
702
703         if (!mem) /* ROOT cgroup has the smallest ID */
704                 return root_mem_cgroup; /*css_put/get against root is ignored*/
705         if (!mem->use_hierarchy) {
706                 if (css_tryget(&mem->css))
707                         return mem;
708                 return NULL;
709         }
710         rcu_read_lock();
711         /*
712          * searching a memory cgroup which has the smallest ID under given
713          * ROOT cgroup. (ID >= 1)
714          */
715         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
716         if (css && css_tryget(css))
717                 mem = container_of(css, struct mem_cgroup, css);
718         else
719                 mem = NULL;
720         rcu_read_unlock();
721         return mem;
722 }
723
724 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
725                                         struct mem_cgroup *root,
726                                         bool cond)
727 {
728         int nextid = css_id(&iter->css) + 1;
729         int found;
730         int hierarchy_used;
731         struct cgroup_subsys_state *css;
732
733         hierarchy_used = iter->use_hierarchy;
734
735         css_put(&iter->css);
736         /* If no ROOT, walk all, ignore hierarchy */
737         if (!cond || (root && !hierarchy_used))
738                 return NULL;
739
740         if (!root)
741                 root = root_mem_cgroup;
742
743         do {
744                 iter = NULL;
745                 rcu_read_lock();
746
747                 css = css_get_next(&mem_cgroup_subsys, nextid,
748                                 &root->css, &found);
749                 if (css && css_tryget(css))
750                         iter = container_of(css, struct mem_cgroup, css);
751                 rcu_read_unlock();
752                 /* If css is NULL, no more cgroups will be found */
753                 nextid = found + 1;
754         } while (css && !iter);
755
756         return iter;
757 }
758 /*
759  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
760  * be careful that "break" loop is not allowed. We have reference count.
761  * Instead of that modify "cond" to be false and "continue" to exit the loop.
762  */
763 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
764         for (iter = mem_cgroup_start_loop(root);\
765              iter != NULL;\
766              iter = mem_cgroup_get_next(iter, root, cond))
767
768 #define for_each_mem_cgroup_tree(iter, root) \
769         for_each_mem_cgroup_tree_cond(iter, root, true)
770
771 #define for_each_mem_cgroup_all(iter) \
772         for_each_mem_cgroup_tree_cond(iter, NULL, true)
773
774
775 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
776 {
777         return (mem == root_mem_cgroup);
778 }
779
780 /*
781  * Following LRU functions are allowed to be used without PCG_LOCK.
782  * Operations are called by routine of global LRU independently from memcg.
783  * What we have to take care of here is validness of pc->mem_cgroup.
784  *
785  * Changes to pc->mem_cgroup happens when
786  * 1. charge
787  * 2. moving account
788  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
789  * It is added to LRU before charge.
790  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
791  * When moving account, the page is not on LRU. It's isolated.
792  */
793
794 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
795 {
796         struct page_cgroup *pc;
797         struct mem_cgroup_per_zone *mz;
798
799         if (mem_cgroup_disabled())
800                 return;
801         pc = lookup_page_cgroup(page);
802         /* can happen while we handle swapcache. */
803         if (!TestClearPageCgroupAcctLRU(pc))
804                 return;
805         VM_BUG_ON(!pc->mem_cgroup);
806         /*
807          * We don't check PCG_USED bit. It's cleared when the "page" is finally
808          * removed from global LRU.
809          */
810         mz = page_cgroup_zoneinfo(pc);
811         MEM_CGROUP_ZSTAT(mz, lru) -= 1;
812         if (mem_cgroup_is_root(pc->mem_cgroup))
813                 return;
814         VM_BUG_ON(list_empty(&pc->lru));
815         list_del_init(&pc->lru);
816         return;
817 }
818
819 void mem_cgroup_del_lru(struct page *page)
820 {
821         mem_cgroup_del_lru_list(page, page_lru(page));
822 }
823
824 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
825 {
826         struct mem_cgroup_per_zone *mz;
827         struct page_cgroup *pc;
828
829         if (mem_cgroup_disabled())
830                 return;
831
832         pc = lookup_page_cgroup(page);
833         /*
834          * Used bit is set without atomic ops but after smp_wmb().
835          * For making pc->mem_cgroup visible, insert smp_rmb() here.
836          */
837         smp_rmb();
838         /* unused or root page is not rotated. */
839         if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
840                 return;
841         mz = page_cgroup_zoneinfo(pc);
842         list_move(&pc->lru, &mz->lists[lru]);
843 }
844
845 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
846 {
847         struct page_cgroup *pc;
848         struct mem_cgroup_per_zone *mz;
849
850         if (mem_cgroup_disabled())
851                 return;
852         pc = lookup_page_cgroup(page);
853         VM_BUG_ON(PageCgroupAcctLRU(pc));
854         /*
855          * Used bit is set without atomic ops but after smp_wmb().
856          * For making pc->mem_cgroup visible, insert smp_rmb() here.
857          */
858         smp_rmb();
859         if (!PageCgroupUsed(pc))
860                 return;
861
862         mz = page_cgroup_zoneinfo(pc);
863         MEM_CGROUP_ZSTAT(mz, lru) += 1;
864         SetPageCgroupAcctLRU(pc);
865         if (mem_cgroup_is_root(pc->mem_cgroup))
866                 return;
867         list_add(&pc->lru, &mz->lists[lru]);
868 }
869
870 /*
871  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
872  * lru because the page may.be reused after it's fully uncharged (because of
873  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
874  * it again. This function is only used to charge SwapCache. It's done under
875  * lock_page and expected that zone->lru_lock is never held.
876  */
877 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
878 {
879         unsigned long flags;
880         struct zone *zone = page_zone(page);
881         struct page_cgroup *pc = lookup_page_cgroup(page);
882
883         spin_lock_irqsave(&zone->lru_lock, flags);
884         /*
885          * Forget old LRU when this page_cgroup is *not* used. This Used bit
886          * is guarded by lock_page() because the page is SwapCache.
887          */
888         if (!PageCgroupUsed(pc))
889                 mem_cgroup_del_lru_list(page, page_lru(page));
890         spin_unlock_irqrestore(&zone->lru_lock, flags);
891 }
892
893 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
894 {
895         unsigned long flags;
896         struct zone *zone = page_zone(page);
897         struct page_cgroup *pc = lookup_page_cgroup(page);
898
899         spin_lock_irqsave(&zone->lru_lock, flags);
900         /* link when the page is linked to LRU but page_cgroup isn't */
901         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
902                 mem_cgroup_add_lru_list(page, page_lru(page));
903         spin_unlock_irqrestore(&zone->lru_lock, flags);
904 }
905
906
907 void mem_cgroup_move_lists(struct page *page,
908                            enum lru_list from, enum lru_list to)
909 {
910         if (mem_cgroup_disabled())
911                 return;
912         mem_cgroup_del_lru_list(page, from);
913         mem_cgroup_add_lru_list(page, to);
914 }
915
916 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
917 {
918         int ret;
919         struct mem_cgroup *curr = NULL;
920         struct task_struct *p;
921
922         p = find_lock_task_mm(task);
923         if (!p)
924                 return 0;
925         curr = try_get_mem_cgroup_from_mm(p->mm);
926         task_unlock(p);
927         if (!curr)
928                 return 0;
929         /*
930          * We should check use_hierarchy of "mem" not "curr". Because checking
931          * use_hierarchy of "curr" here make this function true if hierarchy is
932          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
933          * hierarchy(even if use_hierarchy is disabled in "mem").
934          */
935         if (mem->use_hierarchy)
936                 ret = css_is_ancestor(&curr->css, &mem->css);
937         else
938                 ret = (curr == mem);
939         css_put(&curr->css);
940         return ret;
941 }
942
943 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
944 {
945         unsigned long active;
946         unsigned long inactive;
947         unsigned long gb;
948         unsigned long inactive_ratio;
949
950         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
951         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
952
953         gb = (inactive + active) >> (30 - PAGE_SHIFT);
954         if (gb)
955                 inactive_ratio = int_sqrt(10 * gb);
956         else
957                 inactive_ratio = 1;
958
959         if (present_pages) {
960                 present_pages[0] = inactive;
961                 present_pages[1] = active;
962         }
963
964         return inactive_ratio;
965 }
966
967 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
968 {
969         unsigned long active;
970         unsigned long inactive;
971         unsigned long present_pages[2];
972         unsigned long inactive_ratio;
973
974         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
975
976         inactive = present_pages[0];
977         active = present_pages[1];
978
979         if (inactive * inactive_ratio < active)
980                 return 1;
981
982         return 0;
983 }
984
985 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
986 {
987         unsigned long active;
988         unsigned long inactive;
989
990         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
991         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
992
993         return (active > inactive);
994 }
995
996 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
997                                        struct zone *zone,
998                                        enum lru_list lru)
999 {
1000         int nid = zone_to_nid(zone);
1001         int zid = zone_idx(zone);
1002         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1003
1004         return MEM_CGROUP_ZSTAT(mz, lru);
1005 }
1006
1007 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1008                                                       struct zone *zone)
1009 {
1010         int nid = zone_to_nid(zone);
1011         int zid = zone_idx(zone);
1012         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1013
1014         return &mz->reclaim_stat;
1015 }
1016
1017 struct zone_reclaim_stat *
1018 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1019 {
1020         struct page_cgroup *pc;
1021         struct mem_cgroup_per_zone *mz;
1022
1023         if (mem_cgroup_disabled())
1024                 return NULL;
1025
1026         pc = lookup_page_cgroup(page);
1027         /*
1028          * Used bit is set without atomic ops but after smp_wmb().
1029          * For making pc->mem_cgroup visible, insert smp_rmb() here.
1030          */
1031         smp_rmb();
1032         if (!PageCgroupUsed(pc))
1033                 return NULL;
1034
1035         mz = page_cgroup_zoneinfo(pc);
1036         if (!mz)
1037                 return NULL;
1038
1039         return &mz->reclaim_stat;
1040 }
1041
1042 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1043                                         struct list_head *dst,
1044                                         unsigned long *scanned, int order,
1045                                         int mode, struct zone *z,
1046                                         struct mem_cgroup *mem_cont,
1047                                         int active, int file)
1048 {
1049         unsigned long nr_taken = 0;
1050         struct page *page;
1051         unsigned long scan;
1052         LIST_HEAD(pc_list);
1053         struct list_head *src;
1054         struct page_cgroup *pc, *tmp;
1055         int nid = zone_to_nid(z);
1056         int zid = zone_idx(z);
1057         struct mem_cgroup_per_zone *mz;
1058         int lru = LRU_FILE * file + active;
1059         int ret;
1060
1061         BUG_ON(!mem_cont);
1062         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1063         src = &mz->lists[lru];
1064
1065         scan = 0;
1066         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1067                 if (scan >= nr_to_scan)
1068                         break;
1069
1070                 page = pc->page;
1071                 if (unlikely(!PageCgroupUsed(pc)))
1072                         continue;
1073                 if (unlikely(!PageLRU(page)))
1074                         continue;
1075
1076                 scan++;
1077                 ret = __isolate_lru_page(page, mode, file);
1078                 switch (ret) {
1079                 case 0:
1080                         list_move(&page->lru, dst);
1081                         mem_cgroup_del_lru(page);
1082                         nr_taken++;
1083                         break;
1084                 case -EBUSY:
1085                         /* we don't affect global LRU but rotate in our LRU */
1086                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1087                         break;
1088                 default:
1089                         break;
1090                 }
1091         }
1092
1093         *scanned = scan;
1094
1095         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1096                                       0, 0, 0, mode);
1097
1098         return nr_taken;
1099 }
1100
1101 #define mem_cgroup_from_res_counter(counter, member)    \
1102         container_of(counter, struct mem_cgroup, member)
1103
1104 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1105 {
1106         if (do_swap_account) {
1107                 if (res_counter_check_under_limit(&mem->res) &&
1108                         res_counter_check_under_limit(&mem->memsw))
1109                         return true;
1110         } else
1111                 if (res_counter_check_under_limit(&mem->res))
1112                         return true;
1113         return false;
1114 }
1115
1116 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1117 {
1118         struct cgroup *cgrp = memcg->css.cgroup;
1119         unsigned int swappiness;
1120
1121         /* root ? */
1122         if (cgrp->parent == NULL)
1123                 return vm_swappiness;
1124
1125         spin_lock(&memcg->reclaim_param_lock);
1126         swappiness = memcg->swappiness;
1127         spin_unlock(&memcg->reclaim_param_lock);
1128
1129         return swappiness;
1130 }
1131
1132 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1133 {
1134         int cpu;
1135
1136         get_online_cpus();
1137         spin_lock(&mem->pcp_counter_lock);
1138         for_each_online_cpu(cpu)
1139                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1140         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1141         spin_unlock(&mem->pcp_counter_lock);
1142         put_online_cpus();
1143
1144         synchronize_rcu();
1145 }
1146
1147 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1148 {
1149         int cpu;
1150
1151         if (!mem)
1152                 return;
1153         get_online_cpus();
1154         spin_lock(&mem->pcp_counter_lock);
1155         for_each_online_cpu(cpu)
1156                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1157         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1158         spin_unlock(&mem->pcp_counter_lock);
1159         put_online_cpus();
1160 }
1161 /*
1162  * 2 routines for checking "mem" is under move_account() or not.
1163  *
1164  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1165  *                        for avoiding race in accounting. If true,
1166  *                        pc->mem_cgroup may be overwritten.
1167  *
1168  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1169  *                        under hierarchy of moving cgroups. This is for
1170  *                        waiting at hith-memory prressure caused by "move".
1171  */
1172
1173 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1174 {
1175         VM_BUG_ON(!rcu_read_lock_held());
1176         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1177 }
1178
1179 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1180 {
1181         struct mem_cgroup *from;
1182         struct mem_cgroup *to;
1183         bool ret = false;
1184         /*
1185          * Unlike task_move routines, we access mc.to, mc.from not under
1186          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1187          */
1188         spin_lock(&mc.lock);
1189         from = mc.from;
1190         to = mc.to;
1191         if (!from)
1192                 goto unlock;
1193         if (from == mem || to == mem
1194             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1195             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1196                 ret = true;
1197 unlock:
1198         spin_unlock(&mc.lock);
1199         return ret;
1200 }
1201
1202 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1203 {
1204         if (mc.moving_task && current != mc.moving_task) {
1205                 if (mem_cgroup_under_move(mem)) {
1206                         DEFINE_WAIT(wait);
1207                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1208                         /* moving charge context might have finished. */
1209                         if (mc.moving_task)
1210                                 schedule();
1211                         finish_wait(&mc.waitq, &wait);
1212                         return true;
1213                 }
1214         }
1215         return false;
1216 }
1217
1218 /**
1219  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1220  * @memcg: The memory cgroup that went over limit
1221  * @p: Task that is going to be killed
1222  *
1223  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1224  * enabled
1225  */
1226 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1227 {
1228         struct cgroup *task_cgrp;
1229         struct cgroup *mem_cgrp;
1230         /*
1231          * Need a buffer in BSS, can't rely on allocations. The code relies
1232          * on the assumption that OOM is serialized for memory controller.
1233          * If this assumption is broken, revisit this code.
1234          */
1235         static char memcg_name[PATH_MAX];
1236         int ret;
1237
1238         if (!memcg || !p)
1239                 return;
1240
1241
1242         rcu_read_lock();
1243
1244         mem_cgrp = memcg->css.cgroup;
1245         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1246
1247         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1248         if (ret < 0) {
1249                 /*
1250                  * Unfortunately, we are unable to convert to a useful name
1251                  * But we'll still print out the usage information
1252                  */
1253                 rcu_read_unlock();
1254                 goto done;
1255         }
1256         rcu_read_unlock();
1257
1258         printk(KERN_INFO "Task in %s killed", memcg_name);
1259
1260         rcu_read_lock();
1261         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1262         if (ret < 0) {
1263                 rcu_read_unlock();
1264                 goto done;
1265         }
1266         rcu_read_unlock();
1267
1268         /*
1269          * Continues from above, so we don't need an KERN_ level
1270          */
1271         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1272 done:
1273
1274         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1275                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1276                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1277                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1278         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1279                 "failcnt %llu\n",
1280                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1281                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1282                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1283 }
1284
1285 /*
1286  * This function returns the number of memcg under hierarchy tree. Returns
1287  * 1(self count) if no children.
1288  */
1289 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1290 {
1291         int num = 0;
1292         struct mem_cgroup *iter;
1293
1294         for_each_mem_cgroup_tree(iter, mem)
1295                 num++;
1296         return num;
1297 }
1298
1299 /*
1300  * Return the memory (and swap, if configured) limit for a memcg.
1301  */
1302 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1303 {
1304         u64 limit;
1305         u64 memsw;
1306
1307         limit = res_counter_read_u64(&memcg->res, RES_LIMIT) +
1308                         total_swap_pages;
1309         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1310         /*
1311          * If memsw is finite and limits the amount of swap space available
1312          * to this memcg, return that limit.
1313          */
1314         return min(limit, memsw);
1315 }
1316
1317 /*
1318  * Visit the first child (need not be the first child as per the ordering
1319  * of the cgroup list, since we track last_scanned_child) of @mem and use
1320  * that to reclaim free pages from.
1321  */
1322 static struct mem_cgroup *
1323 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1324 {
1325         struct mem_cgroup *ret = NULL;
1326         struct cgroup_subsys_state *css;
1327         int nextid, found;
1328
1329         if (!root_mem->use_hierarchy) {
1330                 css_get(&root_mem->css);
1331                 ret = root_mem;
1332         }
1333
1334         while (!ret) {
1335                 rcu_read_lock();
1336                 nextid = root_mem->last_scanned_child + 1;
1337                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1338                                    &found);
1339                 if (css && css_tryget(css))
1340                         ret = container_of(css, struct mem_cgroup, css);
1341
1342                 rcu_read_unlock();
1343                 /* Updates scanning parameter */
1344                 spin_lock(&root_mem->reclaim_param_lock);
1345                 if (!css) {
1346                         /* this means start scan from ID:1 */
1347                         root_mem->last_scanned_child = 0;
1348                 } else
1349                         root_mem->last_scanned_child = found;
1350                 spin_unlock(&root_mem->reclaim_param_lock);
1351         }
1352
1353         return ret;
1354 }
1355
1356 /*
1357  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1358  * we reclaimed from, so that we don't end up penalizing one child extensively
1359  * based on its position in the children list.
1360  *
1361  * root_mem is the original ancestor that we've been reclaim from.
1362  *
1363  * We give up and return to the caller when we visit root_mem twice.
1364  * (other groups can be removed while we're walking....)
1365  *
1366  * If shrink==true, for avoiding to free too much, this returns immedieately.
1367  */
1368 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1369                                                 struct zone *zone,
1370                                                 gfp_t gfp_mask,
1371                                                 unsigned long reclaim_options)
1372 {
1373         struct mem_cgroup *victim;
1374         int ret, total = 0;
1375         int loop = 0;
1376         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1377         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1378         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1379         unsigned long excess = mem_cgroup_get_excess(root_mem);
1380
1381         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1382         if (root_mem->memsw_is_minimum)
1383                 noswap = true;
1384
1385         while (1) {
1386                 victim = mem_cgroup_select_victim(root_mem);
1387                 if (victim == root_mem) {
1388                         loop++;
1389                         if (loop >= 1)
1390                                 drain_all_stock_async();
1391                         if (loop >= 2) {
1392                                 /*
1393                                  * If we have not been able to reclaim
1394                                  * anything, it might because there are
1395                                  * no reclaimable pages under this hierarchy
1396                                  */
1397                                 if (!check_soft || !total) {
1398                                         css_put(&victim->css);
1399                                         break;
1400                                 }
1401                                 /*
1402                                  * We want to do more targetted reclaim.
1403                                  * excess >> 2 is not to excessive so as to
1404                                  * reclaim too much, nor too less that we keep
1405                                  * coming back to reclaim from this cgroup
1406                                  */
1407                                 if (total >= (excess >> 2) ||
1408                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1409                                         css_put(&victim->css);
1410                                         break;
1411                                 }
1412                         }
1413                 }
1414                 if (!mem_cgroup_local_usage(victim)) {
1415                         /* this cgroup's local usage == 0 */
1416                         css_put(&victim->css);
1417                         continue;
1418                 }
1419                 /* we use swappiness of local cgroup */
1420                 if (check_soft)
1421                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1422                                 noswap, get_swappiness(victim), zone);
1423                 else
1424                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1425                                                 noswap, get_swappiness(victim));
1426                 css_put(&victim->css);
1427                 /*
1428                  * At shrinking usage, we can't check we should stop here or
1429                  * reclaim more. It's depends on callers. last_scanned_child
1430                  * will work enough for keeping fairness under tree.
1431                  */
1432                 if (shrink)
1433                         return ret;
1434                 total += ret;
1435                 if (check_soft) {
1436                         if (res_counter_check_under_soft_limit(&root_mem->res))
1437                                 return total;
1438                 } else if (mem_cgroup_check_under_limit(root_mem))
1439                         return 1 + total;
1440         }
1441         return total;
1442 }
1443
1444 /*
1445  * Check OOM-Killer is already running under our hierarchy.
1446  * If someone is running, return false.
1447  */
1448 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1449 {
1450         int x, lock_count = 0;
1451         struct mem_cgroup *iter;
1452
1453         for_each_mem_cgroup_tree(iter, mem) {
1454                 x = atomic_inc_return(&iter->oom_lock);
1455                 lock_count = max(x, lock_count);
1456         }
1457
1458         if (lock_count == 1)
1459                 return true;
1460         return false;
1461 }
1462
1463 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1464 {
1465         struct mem_cgroup *iter;
1466
1467         /*
1468          * When a new child is created while the hierarchy is under oom,
1469          * mem_cgroup_oom_lock() may not be called. We have to use
1470          * atomic_add_unless() here.
1471          */
1472         for_each_mem_cgroup_tree(iter, mem)
1473                 atomic_add_unless(&iter->oom_lock, -1, 0);
1474         return 0;
1475 }
1476
1477
1478 static DEFINE_MUTEX(memcg_oom_mutex);
1479 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1480
1481 struct oom_wait_info {
1482         struct mem_cgroup *mem;
1483         wait_queue_t    wait;
1484 };
1485
1486 static int memcg_oom_wake_function(wait_queue_t *wait,
1487         unsigned mode, int sync, void *arg)
1488 {
1489         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1490         struct oom_wait_info *oom_wait_info;
1491
1492         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1493
1494         if (oom_wait_info->mem == wake_mem)
1495                 goto wakeup;
1496         /* if no hierarchy, no match */
1497         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1498                 return 0;
1499         /*
1500          * Both of oom_wait_info->mem and wake_mem are stable under us.
1501          * Then we can use css_is_ancestor without taking care of RCU.
1502          */
1503         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1504             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1505                 return 0;
1506
1507 wakeup:
1508         return autoremove_wake_function(wait, mode, sync, arg);
1509 }
1510
1511 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1512 {
1513         /* for filtering, pass "mem" as argument. */
1514         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1515 }
1516
1517 static void memcg_oom_recover(struct mem_cgroup *mem)
1518 {
1519         if (mem && atomic_read(&mem->oom_lock))
1520                 memcg_wakeup_oom(mem);
1521 }
1522
1523 /*
1524  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1525  */
1526 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1527 {
1528         struct oom_wait_info owait;
1529         bool locked, need_to_kill;
1530
1531         owait.mem = mem;
1532         owait.wait.flags = 0;
1533         owait.wait.func = memcg_oom_wake_function;
1534         owait.wait.private = current;
1535         INIT_LIST_HEAD(&owait.wait.task_list);
1536         need_to_kill = true;
1537         /* At first, try to OOM lock hierarchy under mem.*/
1538         mutex_lock(&memcg_oom_mutex);
1539         locked = mem_cgroup_oom_lock(mem);
1540         /*
1541          * Even if signal_pending(), we can't quit charge() loop without
1542          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1543          * under OOM is always welcomed, use TASK_KILLABLE here.
1544          */
1545         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1546         if (!locked || mem->oom_kill_disable)
1547                 need_to_kill = false;
1548         if (locked)
1549                 mem_cgroup_oom_notify(mem);
1550         mutex_unlock(&memcg_oom_mutex);
1551
1552         if (need_to_kill) {
1553                 finish_wait(&memcg_oom_waitq, &owait.wait);
1554                 mem_cgroup_out_of_memory(mem, mask);
1555         } else {
1556                 schedule();
1557                 finish_wait(&memcg_oom_waitq, &owait.wait);
1558         }
1559         mutex_lock(&memcg_oom_mutex);
1560         mem_cgroup_oom_unlock(mem);
1561         memcg_wakeup_oom(mem);
1562         mutex_unlock(&memcg_oom_mutex);
1563
1564         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1565                 return false;
1566         /* Give chance to dying process */
1567         schedule_timeout(1);
1568         return true;
1569 }
1570
1571 /*
1572  * Currently used to update mapped file statistics, but the routine can be
1573  * generalized to update other statistics as well.
1574  *
1575  * Notes: Race condition
1576  *
1577  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1578  * it tends to be costly. But considering some conditions, we doesn't need
1579  * to do so _always_.
1580  *
1581  * Considering "charge", lock_page_cgroup() is not required because all
1582  * file-stat operations happen after a page is attached to radix-tree. There
1583  * are no race with "charge".
1584  *
1585  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1586  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1587  * if there are race with "uncharge". Statistics itself is properly handled
1588  * by flags.
1589  *
1590  * Considering "move", this is an only case we see a race. To make the race
1591  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1592  * possibility of race condition. If there is, we take a lock.
1593  */
1594
1595 static void mem_cgroup_update_file_stat(struct page *page, int idx, int val)
1596 {
1597         struct mem_cgroup *mem;
1598         struct page_cgroup *pc = lookup_page_cgroup(page);
1599         bool need_unlock = false;
1600
1601         if (unlikely(!pc))
1602                 return;
1603
1604         rcu_read_lock();
1605         mem = pc->mem_cgroup;
1606         if (unlikely(!mem || !PageCgroupUsed(pc)))
1607                 goto out;
1608         /* pc->mem_cgroup is unstable ? */
1609         if (unlikely(mem_cgroup_stealed(mem))) {
1610                 /* take a lock against to access pc->mem_cgroup */
1611                 lock_page_cgroup(pc);
1612                 need_unlock = true;
1613                 mem = pc->mem_cgroup;
1614                 if (!mem || !PageCgroupUsed(pc))
1615                         goto out;
1616         }
1617
1618         this_cpu_add(mem->stat->count[idx], val);
1619
1620         switch (idx) {
1621         case MEM_CGROUP_STAT_FILE_MAPPED:
1622                 if (val > 0)
1623                         SetPageCgroupFileMapped(pc);
1624                 else if (!page_mapped(page))
1625                         ClearPageCgroupFileMapped(pc);
1626                 break;
1627         default:
1628                 BUG();
1629         }
1630
1631 out:
1632         if (unlikely(need_unlock))
1633                 unlock_page_cgroup(pc);
1634         rcu_read_unlock();
1635         return;
1636 }
1637
1638 void mem_cgroup_update_file_mapped(struct page *page, int val)
1639 {
1640         mem_cgroup_update_file_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, val);
1641 }
1642
1643 /*
1644  * size of first charge trial. "32" comes from vmscan.c's magic value.
1645  * TODO: maybe necessary to use big numbers in big irons.
1646  */
1647 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1648 struct memcg_stock_pcp {
1649         struct mem_cgroup *cached; /* this never be root cgroup */
1650         int charge;
1651         struct work_struct work;
1652 };
1653 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1654 static atomic_t memcg_drain_count;
1655
1656 /*
1657  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1658  * from local stock and true is returned. If the stock is 0 or charges from a
1659  * cgroup which is not current target, returns false. This stock will be
1660  * refilled.
1661  */
1662 static bool consume_stock(struct mem_cgroup *mem)
1663 {
1664         struct memcg_stock_pcp *stock;
1665         bool ret = true;
1666
1667         stock = &get_cpu_var(memcg_stock);
1668         if (mem == stock->cached && stock->charge)
1669                 stock->charge -= PAGE_SIZE;
1670         else /* need to call res_counter_charge */
1671                 ret = false;
1672         put_cpu_var(memcg_stock);
1673         return ret;
1674 }
1675
1676 /*
1677  * Returns stocks cached in percpu to res_counter and reset cached information.
1678  */
1679 static void drain_stock(struct memcg_stock_pcp *stock)
1680 {
1681         struct mem_cgroup *old = stock->cached;
1682
1683         if (stock->charge) {
1684                 res_counter_uncharge(&old->res, stock->charge);
1685                 if (do_swap_account)
1686                         res_counter_uncharge(&old->memsw, stock->charge);
1687         }
1688         stock->cached = NULL;
1689         stock->charge = 0;
1690 }
1691
1692 /*
1693  * This must be called under preempt disabled or must be called by
1694  * a thread which is pinned to local cpu.
1695  */
1696 static void drain_local_stock(struct work_struct *dummy)
1697 {
1698         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1699         drain_stock(stock);
1700 }
1701
1702 /*
1703  * Cache charges(val) which is from res_counter, to local per_cpu area.
1704  * This will be consumed by consume_stock() function, later.
1705  */
1706 static void refill_stock(struct mem_cgroup *mem, int val)
1707 {
1708         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1709
1710         if (stock->cached != mem) { /* reset if necessary */
1711                 drain_stock(stock);
1712                 stock->cached = mem;
1713         }
1714         stock->charge += val;
1715         put_cpu_var(memcg_stock);
1716 }
1717
1718 /*
1719  * Tries to drain stocked charges in other cpus. This function is asynchronous
1720  * and just put a work per cpu for draining localy on each cpu. Caller can
1721  * expects some charges will be back to res_counter later but cannot wait for
1722  * it.
1723  */
1724 static void drain_all_stock_async(void)
1725 {
1726         int cpu;
1727         /* This function is for scheduling "drain" in asynchronous way.
1728          * The result of "drain" is not directly handled by callers. Then,
1729          * if someone is calling drain, we don't have to call drain more.
1730          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1731          * there is a race. We just do loose check here.
1732          */
1733         if (atomic_read(&memcg_drain_count))
1734                 return;
1735         /* Notify other cpus that system-wide "drain" is running */
1736         atomic_inc(&memcg_drain_count);
1737         get_online_cpus();
1738         for_each_online_cpu(cpu) {
1739                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1740                 schedule_work_on(cpu, &stock->work);
1741         }
1742         put_online_cpus();
1743         atomic_dec(&memcg_drain_count);
1744         /* We don't wait for flush_work */
1745 }
1746
1747 /* This is a synchronous drain interface. */
1748 static void drain_all_stock_sync(void)
1749 {
1750         /* called when force_empty is called */
1751         atomic_inc(&memcg_drain_count);
1752         schedule_on_each_cpu(drain_local_stock);
1753         atomic_dec(&memcg_drain_count);
1754 }
1755
1756 /*
1757  * This function drains percpu counter value from DEAD cpu and
1758  * move it to local cpu. Note that this function can be preempted.
1759  */
1760 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1761 {
1762         int i;
1763
1764         spin_lock(&mem->pcp_counter_lock);
1765         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1766                 s64 x = per_cpu(mem->stat->count[i], cpu);
1767
1768                 per_cpu(mem->stat->count[i], cpu) = 0;
1769                 mem->nocpu_base.count[i] += x;
1770         }
1771         /* need to clear ON_MOVE value, works as a kind of lock. */
1772         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1773         spin_unlock(&mem->pcp_counter_lock);
1774 }
1775
1776 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1777 {
1778         int idx = MEM_CGROUP_ON_MOVE;
1779
1780         spin_lock(&mem->pcp_counter_lock);
1781         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1782         spin_unlock(&mem->pcp_counter_lock);
1783 }
1784
1785 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1786                                         unsigned long action,
1787                                         void *hcpu)
1788 {
1789         int cpu = (unsigned long)hcpu;
1790         struct memcg_stock_pcp *stock;
1791         struct mem_cgroup *iter;
1792
1793         if ((action == CPU_ONLINE)) {
1794                 for_each_mem_cgroup_all(iter)
1795                         synchronize_mem_cgroup_on_move(iter, cpu);
1796                 return NOTIFY_OK;
1797         }
1798
1799         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1800                 return NOTIFY_OK;
1801
1802         for_each_mem_cgroup_all(iter)
1803                 mem_cgroup_drain_pcp_counter(iter, cpu);
1804
1805         stock = &per_cpu(memcg_stock, cpu);
1806         drain_stock(stock);
1807         return NOTIFY_OK;
1808 }
1809
1810
1811 /* See __mem_cgroup_try_charge() for details */
1812 enum {
1813         CHARGE_OK,              /* success */
1814         CHARGE_RETRY,           /* need to retry but retry is not bad */
1815         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1816         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1817         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1818 };
1819
1820 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1821                                 int csize, bool oom_check)
1822 {
1823         struct mem_cgroup *mem_over_limit;
1824         struct res_counter *fail_res;
1825         unsigned long flags = 0;
1826         int ret;
1827
1828         ret = res_counter_charge(&mem->res, csize, &fail_res);
1829
1830         if (likely(!ret)) {
1831                 if (!do_swap_account)
1832                         return CHARGE_OK;
1833                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1834                 if (likely(!ret))
1835                         return CHARGE_OK;
1836
1837                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1838                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1839         } else
1840                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1841
1842         if (csize > PAGE_SIZE) /* change csize and retry */
1843                 return CHARGE_RETRY;
1844
1845         if (!(gfp_mask & __GFP_WAIT))
1846                 return CHARGE_WOULDBLOCK;
1847
1848         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1849                                         gfp_mask, flags);
1850         /*
1851          * try_to_free_mem_cgroup_pages() might not give us a full
1852          * picture of reclaim. Some pages are reclaimed and might be
1853          * moved to swap cache or just unmapped from the cgroup.
1854          * Check the limit again to see if the reclaim reduced the
1855          * current usage of the cgroup before giving up
1856          */
1857         if (ret || mem_cgroup_check_under_limit(mem_over_limit))
1858                 return CHARGE_RETRY;
1859
1860         /*
1861          * At task move, charge accounts can be doubly counted. So, it's
1862          * better to wait until the end of task_move if something is going on.
1863          */
1864         if (mem_cgroup_wait_acct_move(mem_over_limit))
1865                 return CHARGE_RETRY;
1866
1867         /* If we don't need to call oom-killer at el, return immediately */
1868         if (!oom_check)
1869                 return CHARGE_NOMEM;
1870         /* check OOM */
1871         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1872                 return CHARGE_OOM_DIE;
1873
1874         return CHARGE_RETRY;
1875 }
1876
1877 /*
1878  * Unlike exported interface, "oom" parameter is added. if oom==true,
1879  * oom-killer can be invoked.
1880  */
1881 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1882                 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1883 {
1884         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1885         struct mem_cgroup *mem = NULL;
1886         int ret;
1887         int csize = CHARGE_SIZE;
1888
1889         /*
1890          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1891          * in system level. So, allow to go ahead dying process in addition to
1892          * MEMDIE process.
1893          */
1894         if (unlikely(test_thread_flag(TIF_MEMDIE)
1895                      || fatal_signal_pending(current)))
1896                 goto bypass;
1897
1898         /*
1899          * We always charge the cgroup the mm_struct belongs to.
1900          * The mm_struct's mem_cgroup changes on task migration if the
1901          * thread group leader migrates. It's possible that mm is not
1902          * set, if so charge the init_mm (happens for pagecache usage).
1903          */
1904         if (!*memcg && !mm)
1905                 goto bypass;
1906 again:
1907         if (*memcg) { /* css should be a valid one */
1908                 mem = *memcg;
1909                 VM_BUG_ON(css_is_removed(&mem->css));
1910                 if (mem_cgroup_is_root(mem))
1911                         goto done;
1912                 if (consume_stock(mem))
1913                         goto done;
1914                 css_get(&mem->css);
1915         } else {
1916                 struct task_struct *p;
1917
1918                 rcu_read_lock();
1919                 p = rcu_dereference(mm->owner);
1920                 VM_BUG_ON(!p);
1921                 /*
1922                  * because we don't have task_lock(), "p" can exit while
1923                  * we're here. In that case, "mem" can point to root
1924                  * cgroup but never be NULL. (and task_struct itself is freed
1925                  * by RCU, cgroup itself is RCU safe.) Then, we have small
1926                  * risk here to get wrong cgroup. But such kind of mis-account
1927                  * by race always happens because we don't have cgroup_mutex().
1928                  * It's overkill and we allow that small race, here.
1929                  */
1930                 mem = mem_cgroup_from_task(p);
1931                 VM_BUG_ON(!mem);
1932                 if (mem_cgroup_is_root(mem)) {
1933                         rcu_read_unlock();
1934                         goto done;
1935                 }
1936                 if (consume_stock(mem)) {
1937                         /*
1938                          * It seems dagerous to access memcg without css_get().
1939                          * But considering how consume_stok works, it's not
1940                          * necessary. If consume_stock success, some charges
1941                          * from this memcg are cached on this cpu. So, we
1942                          * don't need to call css_get()/css_tryget() before
1943                          * calling consume_stock().
1944                          */
1945                         rcu_read_unlock();
1946                         goto done;
1947                 }
1948                 /* after here, we may be blocked. we need to get refcnt */
1949                 if (!css_tryget(&mem->css)) {
1950                         rcu_read_unlock();
1951                         goto again;
1952                 }
1953                 rcu_read_unlock();
1954         }
1955
1956         do {
1957                 bool oom_check;
1958
1959                 /* If killed, bypass charge */
1960                 if (fatal_signal_pending(current)) {
1961                         css_put(&mem->css);
1962                         goto bypass;
1963                 }
1964
1965                 oom_check = false;
1966                 if (oom && !nr_oom_retries) {
1967                         oom_check = true;
1968                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1969                 }
1970
1971                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1972
1973                 switch (ret) {
1974                 case CHARGE_OK:
1975                         break;
1976                 case CHARGE_RETRY: /* not in OOM situation but retry */
1977                         csize = PAGE_SIZE;
1978                         css_put(&mem->css);
1979                         mem = NULL;
1980                         goto again;
1981                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1982                         css_put(&mem->css);
1983                         goto nomem;
1984                 case CHARGE_NOMEM: /* OOM routine works */
1985                         if (!oom) {
1986                                 css_put(&mem->css);
1987                                 goto nomem;
1988                         }
1989                         /* If oom, we never return -ENOMEM */
1990                         nr_oom_retries--;
1991                         break;
1992                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
1993                         css_put(&mem->css);
1994                         goto bypass;
1995                 }
1996         } while (ret != CHARGE_OK);
1997
1998         if (csize > PAGE_SIZE)
1999                 refill_stock(mem, csize - PAGE_SIZE);
2000         css_put(&mem->css);
2001 done:
2002         *memcg = mem;
2003         return 0;
2004 nomem:
2005         *memcg = NULL;
2006         return -ENOMEM;
2007 bypass:
2008         *memcg = NULL;
2009         return 0;
2010 }
2011
2012 /*
2013  * Somemtimes we have to undo a charge we got by try_charge().
2014  * This function is for that and do uncharge, put css's refcnt.
2015  * gotten by try_charge().
2016  */
2017 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2018                                                         unsigned long count)
2019 {
2020         if (!mem_cgroup_is_root(mem)) {
2021                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2022                 if (do_swap_account)
2023                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2024         }
2025 }
2026
2027 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
2028 {
2029         __mem_cgroup_cancel_charge(mem, 1);
2030 }
2031
2032 /*
2033  * A helper function to get mem_cgroup from ID. must be called under
2034  * rcu_read_lock(). The caller must check css_is_removed() or some if
2035  * it's concern. (dropping refcnt from swap can be called against removed
2036  * memcg.)
2037  */
2038 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2039 {
2040         struct cgroup_subsys_state *css;
2041
2042         /* ID 0 is unused ID */
2043         if (!id)
2044                 return NULL;
2045         css = css_lookup(&mem_cgroup_subsys, id);
2046         if (!css)
2047                 return NULL;
2048         return container_of(css, struct mem_cgroup, css);
2049 }
2050
2051 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2052 {
2053         struct mem_cgroup *mem = NULL;
2054         struct page_cgroup *pc;
2055         unsigned short id;
2056         swp_entry_t ent;
2057
2058         VM_BUG_ON(!PageLocked(page));
2059
2060         pc = lookup_page_cgroup(page);
2061         lock_page_cgroup(pc);
2062         if (PageCgroupUsed(pc)) {
2063                 mem = pc->mem_cgroup;
2064                 if (mem && !css_tryget(&mem->css))
2065                         mem = NULL;
2066         } else if (PageSwapCache(page)) {
2067                 ent.val = page_private(page);
2068                 id = lookup_swap_cgroup(ent);
2069                 rcu_read_lock();
2070                 mem = mem_cgroup_lookup(id);
2071                 if (mem && !css_tryget(&mem->css))
2072                         mem = NULL;
2073                 rcu_read_unlock();
2074         }
2075         unlock_page_cgroup(pc);
2076         return mem;
2077 }
2078
2079 /*
2080  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
2081  * USED state. If already USED, uncharge and return.
2082  */
2083
2084 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2085                                      struct page_cgroup *pc,
2086                                      enum charge_type ctype)
2087 {
2088         /* try_charge() can return NULL to *memcg, taking care of it. */
2089         if (!mem)
2090                 return;
2091
2092         lock_page_cgroup(pc);
2093         if (unlikely(PageCgroupUsed(pc))) {
2094                 unlock_page_cgroup(pc);
2095                 mem_cgroup_cancel_charge(mem);
2096                 return;
2097         }
2098
2099         pc->mem_cgroup = mem;
2100         /*
2101          * We access a page_cgroup asynchronously without lock_page_cgroup().
2102          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2103          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2104          * before USED bit, we need memory barrier here.
2105          * See mem_cgroup_add_lru_list(), etc.
2106          */
2107         smp_wmb();
2108         switch (ctype) {
2109         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2110         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2111                 SetPageCgroupCache(pc);
2112                 SetPageCgroupUsed(pc);
2113                 break;
2114         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2115                 ClearPageCgroupCache(pc);
2116                 SetPageCgroupUsed(pc);
2117                 break;
2118         default:
2119                 break;
2120         }
2121
2122         mem_cgroup_charge_statistics(mem, pc, true);
2123
2124         unlock_page_cgroup(pc);
2125         /*
2126          * "charge_statistics" updated event counter. Then, check it.
2127          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2128          * if they exceeds softlimit.
2129          */
2130         memcg_check_events(mem, pc->page);
2131 }
2132
2133 /**
2134  * __mem_cgroup_move_account - move account of the page
2135  * @pc: page_cgroup of the page.
2136  * @from: mem_cgroup which the page is moved from.
2137  * @to: mem_cgroup which the page is moved to. @from != @to.
2138  * @uncharge: whether we should call uncharge and css_put against @from.
2139  *
2140  * The caller must confirm following.
2141  * - page is not on LRU (isolate_page() is useful.)
2142  * - the pc is locked, used, and ->mem_cgroup points to @from.
2143  *
2144  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2145  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2146  * true, this function does "uncharge" from old cgroup, but it doesn't if
2147  * @uncharge is false, so a caller should do "uncharge".
2148  */
2149
2150 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2151         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2152 {
2153         VM_BUG_ON(from == to);
2154         VM_BUG_ON(PageLRU(pc->page));
2155         VM_BUG_ON(!PageCgroupLocked(pc));
2156         VM_BUG_ON(!PageCgroupUsed(pc));
2157         VM_BUG_ON(pc->mem_cgroup != from);
2158
2159         if (PageCgroupFileMapped(pc)) {
2160                 /* Update mapped_file data for mem_cgroup */
2161                 preempt_disable();
2162                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2163                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2164                 preempt_enable();
2165         }
2166         mem_cgroup_charge_statistics(from, pc, false);
2167         if (uncharge)
2168                 /* This is not "cancel", but cancel_charge does all we need. */
2169                 mem_cgroup_cancel_charge(from);
2170
2171         /* caller should have done css_get */
2172         pc->mem_cgroup = to;
2173         mem_cgroup_charge_statistics(to, pc, true);
2174         /*
2175          * We charges against "to" which may not have any tasks. Then, "to"
2176          * can be under rmdir(). But in current implementation, caller of
2177          * this function is just force_empty() and move charge, so it's
2178          * garanteed that "to" is never removed. So, we don't check rmdir
2179          * status here.
2180          */
2181 }
2182
2183 /*
2184  * check whether the @pc is valid for moving account and call
2185  * __mem_cgroup_move_account()
2186  */
2187 static int mem_cgroup_move_account(struct page_cgroup *pc,
2188                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2189 {
2190         int ret = -EINVAL;
2191         lock_page_cgroup(pc);
2192         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2193                 __mem_cgroup_move_account(pc, from, to, uncharge);
2194                 ret = 0;
2195         }
2196         unlock_page_cgroup(pc);
2197         /*
2198          * check events
2199          */
2200         memcg_check_events(to, pc->page);
2201         memcg_check_events(from, pc->page);
2202         return ret;
2203 }
2204
2205 /*
2206  * move charges to its parent.
2207  */
2208
2209 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2210                                   struct mem_cgroup *child,
2211                                   gfp_t gfp_mask)
2212 {
2213         struct page *page = pc->page;
2214         struct cgroup *cg = child->css.cgroup;
2215         struct cgroup *pcg = cg->parent;
2216         struct mem_cgroup *parent;
2217         int ret;
2218
2219         /* Is ROOT ? */
2220         if (!pcg)
2221                 return -EINVAL;
2222
2223         ret = -EBUSY;
2224         if (!get_page_unless_zero(page))
2225                 goto out;
2226         if (isolate_lru_page(page))
2227                 goto put;
2228
2229         parent = mem_cgroup_from_cont(pcg);
2230         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
2231         if (ret || !parent)
2232                 goto put_back;
2233
2234         ret = mem_cgroup_move_account(pc, child, parent, true);
2235         if (ret)
2236                 mem_cgroup_cancel_charge(parent);
2237 put_back:
2238         putback_lru_page(page);
2239 put:
2240         put_page(page);
2241 out:
2242         return ret;
2243 }
2244
2245 /*
2246  * Charge the memory controller for page usage.
2247  * Return
2248  * 0 if the charge was successful
2249  * < 0 if the cgroup is over its limit
2250  */
2251 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2252                                 gfp_t gfp_mask, enum charge_type ctype)
2253 {
2254         struct mem_cgroup *mem = NULL;
2255         struct page_cgroup *pc;
2256         int ret;
2257
2258         pc = lookup_page_cgroup(page);
2259         /* can happen at boot */
2260         if (unlikely(!pc))
2261                 return 0;
2262         prefetchw(pc);
2263
2264         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
2265         if (ret || !mem)
2266                 return ret;
2267
2268         __mem_cgroup_commit_charge(mem, pc, ctype);
2269         return 0;
2270 }
2271
2272 int mem_cgroup_newpage_charge(struct page *page,
2273                               struct mm_struct *mm, gfp_t gfp_mask)
2274 {
2275         if (mem_cgroup_disabled())
2276                 return 0;
2277         if (PageCompound(page))
2278                 return 0;
2279         /*
2280          * If already mapped, we don't have to account.
2281          * If page cache, page->mapping has address_space.
2282          * But page->mapping may have out-of-use anon_vma pointer,
2283          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2284          * is NULL.
2285          */
2286         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2287                 return 0;
2288         if (unlikely(!mm))
2289                 mm = &init_mm;
2290         return mem_cgroup_charge_common(page, mm, gfp_mask,
2291                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2292 }
2293
2294 static void
2295 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2296                                         enum charge_type ctype);
2297
2298 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2299                                 gfp_t gfp_mask)
2300 {
2301         int ret;
2302
2303         if (mem_cgroup_disabled())
2304                 return 0;
2305         if (PageCompound(page))
2306                 return 0;
2307         /*
2308          * Corner case handling. This is called from add_to_page_cache()
2309          * in usual. But some FS (shmem) precharges this page before calling it
2310          * and call add_to_page_cache() with GFP_NOWAIT.
2311          *
2312          * For GFP_NOWAIT case, the page may be pre-charged before calling
2313          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2314          * charge twice. (It works but has to pay a bit larger cost.)
2315          * And when the page is SwapCache, it should take swap information
2316          * into account. This is under lock_page() now.
2317          */
2318         if (!(gfp_mask & __GFP_WAIT)) {
2319                 struct page_cgroup *pc;
2320
2321                 pc = lookup_page_cgroup(page);
2322                 if (!pc)
2323                         return 0;
2324                 lock_page_cgroup(pc);
2325                 if (PageCgroupUsed(pc)) {
2326                         unlock_page_cgroup(pc);
2327                         return 0;
2328                 }
2329                 unlock_page_cgroup(pc);
2330         }
2331
2332         if (unlikely(!mm))
2333                 mm = &init_mm;
2334
2335         if (page_is_file_cache(page))
2336                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2337                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2338
2339         /* shmem */
2340         if (PageSwapCache(page)) {
2341                 struct mem_cgroup *mem = NULL;
2342
2343                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2344                 if (!ret)
2345                         __mem_cgroup_commit_charge_swapin(page, mem,
2346                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2347         } else
2348                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2349                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2350
2351         return ret;
2352 }
2353
2354 /*
2355  * While swap-in, try_charge -> commit or cancel, the page is locked.
2356  * And when try_charge() successfully returns, one refcnt to memcg without
2357  * struct page_cgroup is acquired. This refcnt will be consumed by
2358  * "commit()" or removed by "cancel()"
2359  */
2360 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2361                                  struct page *page,
2362                                  gfp_t mask, struct mem_cgroup **ptr)
2363 {
2364         struct mem_cgroup *mem;
2365         int ret;
2366
2367         if (mem_cgroup_disabled())
2368                 return 0;
2369
2370         if (!do_swap_account)
2371                 goto charge_cur_mm;
2372         /*
2373          * A racing thread's fault, or swapoff, may have already updated
2374          * the pte, and even removed page from swap cache: in those cases
2375          * do_swap_page()'s pte_same() test will fail; but there's also a
2376          * KSM case which does need to charge the page.
2377          */
2378         if (!PageSwapCache(page))
2379                 goto charge_cur_mm;
2380         mem = try_get_mem_cgroup_from_page(page);
2381         if (!mem)
2382                 goto charge_cur_mm;
2383         *ptr = mem;
2384         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2385         css_put(&mem->css);
2386         return ret;
2387 charge_cur_mm:
2388         if (unlikely(!mm))
2389                 mm = &init_mm;
2390         return __mem_cgroup_try_charge(mm, mask, ptr, true);
2391 }
2392
2393 static void
2394 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2395                                         enum charge_type ctype)
2396 {
2397         struct page_cgroup *pc;
2398
2399         if (mem_cgroup_disabled())
2400                 return;
2401         if (!ptr)
2402                 return;
2403         cgroup_exclude_rmdir(&ptr->css);
2404         pc = lookup_page_cgroup(page);
2405         mem_cgroup_lru_del_before_commit_swapcache(page);
2406         __mem_cgroup_commit_charge(ptr, pc, ctype);
2407         mem_cgroup_lru_add_after_commit_swapcache(page);
2408         /*
2409          * Now swap is on-memory. This means this page may be
2410          * counted both as mem and swap....double count.
2411          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2412          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2413          * may call delete_from_swap_cache() before reach here.
2414          */
2415         if (do_swap_account && PageSwapCache(page)) {
2416                 swp_entry_t ent = {.val = page_private(page)};
2417                 unsigned short id;
2418                 struct mem_cgroup *memcg;
2419
2420                 id = swap_cgroup_record(ent, 0);
2421                 rcu_read_lock();
2422                 memcg = mem_cgroup_lookup(id);
2423                 if (memcg) {
2424                         /*
2425                          * This recorded memcg can be obsolete one. So, avoid
2426                          * calling css_tryget
2427                          */
2428                         if (!mem_cgroup_is_root(memcg))
2429                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2430                         mem_cgroup_swap_statistics(memcg, false);
2431                         mem_cgroup_put(memcg);
2432                 }
2433                 rcu_read_unlock();
2434         }
2435         /*
2436          * At swapin, we may charge account against cgroup which has no tasks.
2437          * So, rmdir()->pre_destroy() can be called while we do this charge.
2438          * In that case, we need to call pre_destroy() again. check it here.
2439          */
2440         cgroup_release_and_wakeup_rmdir(&ptr->css);
2441 }
2442
2443 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2444 {
2445         __mem_cgroup_commit_charge_swapin(page, ptr,
2446                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2447 }
2448
2449 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2450 {
2451         if (mem_cgroup_disabled())
2452                 return;
2453         if (!mem)
2454                 return;
2455         mem_cgroup_cancel_charge(mem);
2456 }
2457
2458 static void
2459 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2460 {
2461         struct memcg_batch_info *batch = NULL;
2462         bool uncharge_memsw = true;
2463         /* If swapout, usage of swap doesn't decrease */
2464         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2465                 uncharge_memsw = false;
2466
2467         batch = &current->memcg_batch;
2468         /*
2469          * In usual, we do css_get() when we remember memcg pointer.
2470          * But in this case, we keep res->usage until end of a series of
2471          * uncharges. Then, it's ok to ignore memcg's refcnt.
2472          */
2473         if (!batch->memcg)
2474                 batch->memcg = mem;
2475         /*
2476          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2477          * In those cases, all pages freed continously can be expected to be in
2478          * the same cgroup and we have chance to coalesce uncharges.
2479          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2480          * because we want to do uncharge as soon as possible.
2481          */
2482
2483         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2484                 goto direct_uncharge;
2485
2486         /*
2487          * In typical case, batch->memcg == mem. This means we can
2488          * merge a series of uncharges to an uncharge of res_counter.
2489          * If not, we uncharge res_counter ony by one.
2490          */
2491         if (batch->memcg != mem)
2492                 goto direct_uncharge;
2493         /* remember freed charge and uncharge it later */
2494         batch->bytes += PAGE_SIZE;
2495         if (uncharge_memsw)
2496                 batch->memsw_bytes += PAGE_SIZE;
2497         return;
2498 direct_uncharge:
2499         res_counter_uncharge(&mem->res, PAGE_SIZE);
2500         if (uncharge_memsw)
2501                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2502         if (unlikely(batch->memcg != mem))
2503                 memcg_oom_recover(mem);
2504         return;
2505 }
2506
2507 /*
2508  * uncharge if !page_mapped(page)
2509  */
2510 static struct mem_cgroup *
2511 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2512 {
2513         struct page_cgroup *pc;
2514         struct mem_cgroup *mem = NULL;
2515
2516         if (mem_cgroup_disabled())
2517                 return NULL;
2518
2519         if (PageSwapCache(page))
2520                 return NULL;
2521
2522         /*
2523          * Check if our page_cgroup is valid
2524          */
2525         pc = lookup_page_cgroup(page);
2526         if (unlikely(!pc || !PageCgroupUsed(pc)))
2527                 return NULL;
2528
2529         lock_page_cgroup(pc);
2530
2531         mem = pc->mem_cgroup;
2532
2533         if (!PageCgroupUsed(pc))
2534                 goto unlock_out;
2535
2536         switch (ctype) {
2537         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2538         case MEM_CGROUP_CHARGE_TYPE_DROP:
2539                 /* See mem_cgroup_prepare_migration() */
2540                 if (page_mapped(page) || PageCgroupMigration(pc))
2541                         goto unlock_out;
2542                 break;
2543         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2544                 if (!PageAnon(page)) {  /* Shared memory */
2545                         if (page->mapping && !page_is_file_cache(page))
2546                                 goto unlock_out;
2547                 } else if (page_mapped(page)) /* Anon */
2548                                 goto unlock_out;
2549                 break;
2550         default:
2551                 break;
2552         }
2553
2554         mem_cgroup_charge_statistics(mem, pc, false);
2555
2556         ClearPageCgroupUsed(pc);
2557         /*
2558          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2559          * freed from LRU. This is safe because uncharged page is expected not
2560          * to be reused (freed soon). Exception is SwapCache, it's handled by
2561          * special functions.
2562          */
2563
2564         unlock_page_cgroup(pc);
2565         /*
2566          * even after unlock, we have mem->res.usage here and this memcg
2567          * will never be freed.
2568          */
2569         memcg_check_events(mem, page);
2570         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2571                 mem_cgroup_swap_statistics(mem, true);
2572                 mem_cgroup_get(mem);
2573         }
2574         if (!mem_cgroup_is_root(mem))
2575                 __do_uncharge(mem, ctype);
2576
2577         return mem;
2578
2579 unlock_out:
2580         unlock_page_cgroup(pc);
2581         return NULL;
2582 }
2583
2584 void mem_cgroup_uncharge_page(struct page *page)
2585 {
2586         /* early check. */
2587         if (page_mapped(page))
2588                 return;
2589         if (page->mapping && !PageAnon(page))
2590                 return;
2591         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2592 }
2593
2594 void mem_cgroup_uncharge_cache_page(struct page *page)
2595 {
2596         VM_BUG_ON(page_mapped(page));
2597         VM_BUG_ON(page->mapping);
2598         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2599 }
2600
2601 /*
2602  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2603  * In that cases, pages are freed continuously and we can expect pages
2604  * are in the same memcg. All these calls itself limits the number of
2605  * pages freed at once, then uncharge_start/end() is called properly.
2606  * This may be called prural(2) times in a context,
2607  */
2608
2609 void mem_cgroup_uncharge_start(void)
2610 {
2611         current->memcg_batch.do_batch++;
2612         /* We can do nest. */
2613         if (current->memcg_batch.do_batch == 1) {
2614                 current->memcg_batch.memcg = NULL;
2615                 current->memcg_batch.bytes = 0;
2616                 current->memcg_batch.memsw_bytes = 0;
2617         }
2618 }
2619
2620 void mem_cgroup_uncharge_end(void)
2621 {
2622         struct memcg_batch_info *batch = &current->memcg_batch;
2623
2624         if (!batch->do_batch)
2625                 return;
2626
2627         batch->do_batch--;
2628         if (batch->do_batch) /* If stacked, do nothing. */
2629                 return;
2630
2631         if (!batch->memcg)
2632                 return;
2633         /*
2634          * This "batch->memcg" is valid without any css_get/put etc...
2635          * bacause we hide charges behind us.
2636          */
2637         if (batch->bytes)
2638                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2639         if (batch->memsw_bytes)
2640                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2641         memcg_oom_recover(batch->memcg);
2642         /* forget this pointer (for sanity check) */
2643         batch->memcg = NULL;
2644 }
2645
2646 #ifdef CONFIG_SWAP
2647 /*
2648  * called after __delete_from_swap_cache() and drop "page" account.
2649  * memcg information is recorded to swap_cgroup of "ent"
2650  */
2651 void
2652 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2653 {
2654         struct mem_cgroup *memcg;
2655         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2656
2657         if (!swapout) /* this was a swap cache but the swap is unused ! */
2658                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2659
2660         memcg = __mem_cgroup_uncharge_common(page, ctype);
2661
2662         /*
2663          * record memcg information,  if swapout && memcg != NULL,
2664          * mem_cgroup_get() was called in uncharge().
2665          */
2666         if (do_swap_account && swapout && memcg)
2667                 swap_cgroup_record(ent, css_id(&memcg->css));
2668 }
2669 #endif
2670
2671 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2672 /*
2673  * called from swap_entry_free(). remove record in swap_cgroup and
2674  * uncharge "memsw" account.
2675  */
2676 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2677 {
2678         struct mem_cgroup *memcg;
2679         unsigned short id;
2680
2681         if (!do_swap_account)
2682                 return;
2683
2684         id = swap_cgroup_record(ent, 0);
2685         rcu_read_lock();
2686         memcg = mem_cgroup_lookup(id);
2687         if (memcg) {
2688                 /*
2689                  * We uncharge this because swap is freed.
2690                  * This memcg can be obsolete one. We avoid calling css_tryget
2691                  */
2692                 if (!mem_cgroup_is_root(memcg))
2693                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2694                 mem_cgroup_swap_statistics(memcg, false);
2695                 mem_cgroup_put(memcg);
2696         }
2697         rcu_read_unlock();
2698 }
2699
2700 /**
2701  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2702  * @entry: swap entry to be moved
2703  * @from:  mem_cgroup which the entry is moved from
2704  * @to:  mem_cgroup which the entry is moved to
2705  * @need_fixup: whether we should fixup res_counters and refcounts.
2706  *
2707  * It succeeds only when the swap_cgroup's record for this entry is the same
2708  * as the mem_cgroup's id of @from.
2709  *
2710  * Returns 0 on success, -EINVAL on failure.
2711  *
2712  * The caller must have charged to @to, IOW, called res_counter_charge() about
2713  * both res and memsw, and called css_get().
2714  */
2715 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2716                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2717 {
2718         unsigned short old_id, new_id;
2719
2720         old_id = css_id(&from->css);
2721         new_id = css_id(&to->css);
2722
2723         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2724                 mem_cgroup_swap_statistics(from, false);
2725                 mem_cgroup_swap_statistics(to, true);
2726                 /*
2727                  * This function is only called from task migration context now.
2728                  * It postpones res_counter and refcount handling till the end
2729                  * of task migration(mem_cgroup_clear_mc()) for performance
2730                  * improvement. But we cannot postpone mem_cgroup_get(to)
2731                  * because if the process that has been moved to @to does
2732                  * swap-in, the refcount of @to might be decreased to 0.
2733                  */
2734                 mem_cgroup_get(to);
2735                 if (need_fixup) {
2736                         if (!mem_cgroup_is_root(from))
2737                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2738                         mem_cgroup_put(from);
2739                         /*
2740                          * we charged both to->res and to->memsw, so we should
2741                          * uncharge to->res.
2742                          */
2743                         if (!mem_cgroup_is_root(to))
2744                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2745                 }
2746                 return 0;
2747         }
2748         return -EINVAL;
2749 }
2750 #else
2751 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2752                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2753 {
2754         return -EINVAL;
2755 }
2756 #endif
2757
2758 /*
2759  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2760  * page belongs to.
2761  */
2762 int mem_cgroup_prepare_migration(struct page *page,
2763         struct page *newpage, struct mem_cgroup **ptr)
2764 {
2765         struct page_cgroup *pc;
2766         struct mem_cgroup *mem = NULL;
2767         enum charge_type ctype;
2768         int ret = 0;
2769
2770         if (mem_cgroup_disabled())
2771                 return 0;
2772
2773         pc = lookup_page_cgroup(page);
2774         lock_page_cgroup(pc);
2775         if (PageCgroupUsed(pc)) {
2776                 mem = pc->mem_cgroup;
2777                 css_get(&mem->css);
2778                 /*
2779                  * At migrating an anonymous page, its mapcount goes down
2780                  * to 0 and uncharge() will be called. But, even if it's fully
2781                  * unmapped, migration may fail and this page has to be
2782                  * charged again. We set MIGRATION flag here and delay uncharge
2783                  * until end_migration() is called
2784                  *
2785                  * Corner Case Thinking
2786                  * A)
2787                  * When the old page was mapped as Anon and it's unmap-and-freed
2788                  * while migration was ongoing.
2789                  * If unmap finds the old page, uncharge() of it will be delayed
2790                  * until end_migration(). If unmap finds a new page, it's
2791                  * uncharged when it make mapcount to be 1->0. If unmap code
2792                  * finds swap_migration_entry, the new page will not be mapped
2793                  * and end_migration() will find it(mapcount==0).
2794                  *
2795                  * B)
2796                  * When the old page was mapped but migraion fails, the kernel
2797                  * remaps it. A charge for it is kept by MIGRATION flag even
2798                  * if mapcount goes down to 0. We can do remap successfully
2799                  * without charging it again.
2800                  *
2801                  * C)
2802                  * The "old" page is under lock_page() until the end of
2803                  * migration, so, the old page itself will not be swapped-out.
2804                  * If the new page is swapped out before end_migraton, our
2805                  * hook to usual swap-out path will catch the event.
2806                  */
2807                 if (PageAnon(page))
2808                         SetPageCgroupMigration(pc);
2809         }
2810         unlock_page_cgroup(pc);
2811         /*
2812          * If the page is not charged at this point,
2813          * we return here.
2814          */
2815         if (!mem)
2816                 return 0;
2817
2818         *ptr = mem;
2819         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2820         css_put(&mem->css);/* drop extra refcnt */
2821         if (ret || *ptr == NULL) {
2822                 if (PageAnon(page)) {
2823                         lock_page_cgroup(pc);
2824                         ClearPageCgroupMigration(pc);
2825                         unlock_page_cgroup(pc);
2826                         /*
2827                          * The old page may be fully unmapped while we kept it.
2828                          */
2829                         mem_cgroup_uncharge_page(page);
2830                 }
2831                 return -ENOMEM;
2832         }
2833         /*
2834          * We charge new page before it's used/mapped. So, even if unlock_page()
2835          * is called before end_migration, we can catch all events on this new
2836          * page. In the case new page is migrated but not remapped, new page's
2837          * mapcount will be finally 0 and we call uncharge in end_migration().
2838          */
2839         pc = lookup_page_cgroup(newpage);
2840         if (PageAnon(page))
2841                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2842         else if (page_is_file_cache(page))
2843                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2844         else
2845                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2846         __mem_cgroup_commit_charge(mem, pc, ctype);
2847         return ret;
2848 }
2849
2850 /* remove redundant charge if migration failed*/
2851 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2852         struct page *oldpage, struct page *newpage)
2853 {
2854         struct page *used, *unused;
2855         struct page_cgroup *pc;
2856
2857         if (!mem)
2858                 return;
2859         /* blocks rmdir() */
2860         cgroup_exclude_rmdir(&mem->css);
2861         /* at migration success, oldpage->mapping is NULL. */
2862         if (oldpage->mapping) {
2863                 used = oldpage;
2864                 unused = newpage;
2865         } else {
2866                 used = newpage;
2867                 unused = oldpage;
2868         }
2869         /*
2870          * We disallowed uncharge of pages under migration because mapcount
2871          * of the page goes down to zero, temporarly.
2872          * Clear the flag and check the page should be charged.
2873          */
2874         pc = lookup_page_cgroup(oldpage);
2875         lock_page_cgroup(pc);
2876         ClearPageCgroupMigration(pc);
2877         unlock_page_cgroup(pc);
2878
2879         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2880
2881         /*
2882          * If a page is a file cache, radix-tree replacement is very atomic
2883          * and we can skip this check. When it was an Anon page, its mapcount
2884          * goes down to 0. But because we added MIGRATION flage, it's not
2885          * uncharged yet. There are several case but page->mapcount check
2886          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2887          * check. (see prepare_charge() also)
2888          */
2889         if (PageAnon(used))
2890                 mem_cgroup_uncharge_page(used);
2891         /*
2892          * At migration, we may charge account against cgroup which has no
2893          * tasks.
2894          * So, rmdir()->pre_destroy() can be called while we do this charge.
2895          * In that case, we need to call pre_destroy() again. check it here.
2896          */
2897         cgroup_release_and_wakeup_rmdir(&mem->css);
2898 }
2899
2900 /*
2901  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2902  * Calling hierarchical_reclaim is not enough because we should update
2903  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2904  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2905  * not from the memcg which this page would be charged to.
2906  * try_charge_swapin does all of these works properly.
2907  */
2908 int mem_cgroup_shmem_charge_fallback(struct page *page,
2909                             struct mm_struct *mm,
2910                             gfp_t gfp_mask)
2911 {
2912         struct mem_cgroup *mem = NULL;
2913         int ret;
2914
2915         if (mem_cgroup_disabled())
2916                 return 0;
2917
2918         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2919         if (!ret)
2920                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2921
2922         return ret;
2923 }
2924
2925 static DEFINE_MUTEX(set_limit_mutex);
2926
2927 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2928                                 unsigned long long val)
2929 {
2930         int retry_count;
2931         u64 memswlimit, memlimit;
2932         int ret = 0;
2933         int children = mem_cgroup_count_children(memcg);
2934         u64 curusage, oldusage;
2935         int enlarge;
2936
2937         /*
2938          * For keeping hierarchical_reclaim simple, how long we should retry
2939          * is depends on callers. We set our retry-count to be function
2940          * of # of children which we should visit in this loop.
2941          */
2942         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2943
2944         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2945
2946         enlarge = 0;
2947         while (retry_count) {
2948                 if (signal_pending(current)) {
2949                         ret = -EINTR;
2950                         break;
2951                 }
2952                 /*
2953                  * Rather than hide all in some function, I do this in
2954                  * open coded manner. You see what this really does.
2955                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2956                  */
2957                 mutex_lock(&set_limit_mutex);
2958                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2959                 if (memswlimit < val) {
2960                         ret = -EINVAL;
2961                         mutex_unlock(&set_limit_mutex);
2962                         break;
2963                 }
2964
2965                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2966                 if (memlimit < val)
2967                         enlarge = 1;
2968
2969                 ret = res_counter_set_limit(&memcg->res, val);
2970                 if (!ret) {
2971                         if (memswlimit == val)
2972                                 memcg->memsw_is_minimum = true;
2973                         else
2974                                 memcg->memsw_is_minimum = false;
2975                 }
2976                 mutex_unlock(&set_limit_mutex);
2977
2978                 if (!ret)
2979                         break;
2980
2981                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2982                                                 MEM_CGROUP_RECLAIM_SHRINK);
2983                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2984                 /* Usage is reduced ? */
2985                 if (curusage >= oldusage)
2986                         retry_count--;
2987                 else
2988                         oldusage = curusage;
2989         }
2990         if (!ret && enlarge)
2991                 memcg_oom_recover(memcg);
2992
2993         return ret;
2994 }
2995
2996 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2997                                         unsigned long long val)
2998 {
2999         int retry_count;
3000         u64 memlimit, memswlimit, oldusage, curusage;
3001         int children = mem_cgroup_count_children(memcg);
3002         int ret = -EBUSY;
3003         int enlarge = 0;
3004
3005         /* see mem_cgroup_resize_res_limit */
3006         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3007         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3008         while (retry_count) {
3009                 if (signal_pending(current)) {
3010                         ret = -EINTR;
3011                         break;
3012                 }
3013                 /*
3014                  * Rather than hide all in some function, I do this in
3015                  * open coded manner. You see what this really does.
3016                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3017                  */
3018                 mutex_lock(&set_limit_mutex);
3019                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3020                 if (memlimit > val) {
3021                         ret = -EINVAL;
3022                         mutex_unlock(&set_limit_mutex);
3023                         break;
3024                 }
3025                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3026                 if (memswlimit < val)
3027                         enlarge = 1;
3028                 ret = res_counter_set_limit(&memcg->memsw, val);
3029                 if (!ret) {
3030                         if (memlimit == val)
3031                                 memcg->memsw_is_minimum = true;
3032                         else
3033                                 memcg->memsw_is_minimum = false;
3034                 }
3035                 mutex_unlock(&set_limit_mutex);
3036
3037                 if (!ret)
3038                         break;
3039
3040                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3041                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3042                                                 MEM_CGROUP_RECLAIM_SHRINK);
3043                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3044                 /* Usage is reduced ? */
3045                 if (curusage >= oldusage)
3046                         retry_count--;
3047                 else
3048                         oldusage = curusage;
3049         }
3050         if (!ret && enlarge)
3051                 memcg_oom_recover(memcg);
3052         return ret;
3053 }
3054
3055 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3056                                             gfp_t gfp_mask)
3057 {
3058         unsigned long nr_reclaimed = 0;
3059         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3060         unsigned long reclaimed;
3061         int loop = 0;
3062         struct mem_cgroup_tree_per_zone *mctz;
3063         unsigned long long excess;
3064
3065         if (order > 0)
3066                 return 0;
3067
3068         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3069         /*
3070          * This loop can run a while, specially if mem_cgroup's continuously
3071          * keep exceeding their soft limit and putting the system under
3072          * pressure
3073          */
3074         do {
3075                 if (next_mz)
3076                         mz = next_mz;
3077                 else
3078                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3079                 if (!mz)
3080                         break;
3081
3082                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3083                                                 gfp_mask,
3084                                                 MEM_CGROUP_RECLAIM_SOFT);
3085                 nr_reclaimed += reclaimed;
3086                 spin_lock(&mctz->lock);
3087
3088                 /*
3089                  * If we failed to reclaim anything from this memory cgroup
3090                  * it is time to move on to the next cgroup
3091                  */
3092                 next_mz = NULL;
3093                 if (!reclaimed) {
3094                         do {
3095                                 /*
3096                                  * Loop until we find yet another one.
3097                                  *
3098                                  * By the time we get the soft_limit lock
3099                                  * again, someone might have aded the
3100                                  * group back on the RB tree. Iterate to
3101                                  * make sure we get a different mem.
3102                                  * mem_cgroup_largest_soft_limit_node returns
3103                                  * NULL if no other cgroup is present on
3104                                  * the tree
3105                                  */
3106                                 next_mz =
3107                                 __mem_cgroup_largest_soft_limit_node(mctz);
3108                                 if (next_mz == mz) {
3109                                         css_put(&next_mz->mem->css);
3110                                         next_mz = NULL;
3111                                 } else /* next_mz == NULL or other memcg */
3112                                         break;
3113                         } while (1);
3114                 }
3115                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3116                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3117                 /*
3118                  * One school of thought says that we should not add
3119                  * back the node to the tree if reclaim returns 0.
3120                  * But our reclaim could return 0, simply because due
3121                  * to priority we are exposing a smaller subset of
3122                  * memory to reclaim from. Consider this as a longer
3123                  * term TODO.
3124                  */
3125                 /* If excess == 0, no tree ops */
3126                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3127                 spin_unlock(&mctz->lock);
3128                 css_put(&mz->mem->css);
3129                 loop++;
3130                 /*
3131                  * Could not reclaim anything and there are no more
3132                  * mem cgroups to try or we seem to be looping without
3133                  * reclaiming anything.
3134                  */
3135                 if (!nr_reclaimed &&
3136                         (next_mz == NULL ||
3137                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3138                         break;
3139         } while (!nr_reclaimed);
3140         if (next_mz)
3141                 css_put(&next_mz->mem->css);
3142         return nr_reclaimed;
3143 }
3144
3145 /*
3146  * This routine traverse page_cgroup in given list and drop them all.
3147  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3148  */
3149 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3150                                 int node, int zid, enum lru_list lru)
3151 {
3152         struct zone *zone;
3153         struct mem_cgroup_per_zone *mz;
3154         struct page_cgroup *pc, *busy;
3155         unsigned long flags, loop;
3156         struct list_head *list;
3157         int ret = 0;
3158
3159         zone = &NODE_DATA(node)->node_zones[zid];
3160         mz = mem_cgroup_zoneinfo(mem, node, zid);
3161         list = &mz->lists[lru];
3162
3163         loop = MEM_CGROUP_ZSTAT(mz, lru);
3164         /* give some margin against EBUSY etc...*/
3165         loop += 256;
3166         busy = NULL;
3167         while (loop--) {
3168                 ret = 0;
3169                 spin_lock_irqsave(&zone->lru_lock, flags);
3170                 if (list_empty(list)) {
3171                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3172                         break;
3173                 }
3174                 pc = list_entry(list->prev, struct page_cgroup, lru);
3175                 if (busy == pc) {
3176                         list_move(&pc->lru, list);
3177                         busy = NULL;
3178                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3179                         continue;
3180                 }
3181                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3182
3183                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3184                 if (ret == -ENOMEM)
3185                         break;
3186
3187                 if (ret == -EBUSY || ret == -EINVAL) {
3188                         /* found lock contention or "pc" is obsolete. */
3189                         busy = pc;
3190                         cond_resched();
3191                 } else
3192                         busy = NULL;
3193         }
3194
3195         if (!ret && !list_empty(list))
3196                 return -EBUSY;
3197         return ret;
3198 }
3199
3200 /*
3201  * make mem_cgroup's charge to be 0 if there is no task.
3202  * This enables deleting this mem_cgroup.
3203  */
3204 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3205 {
3206         int ret;
3207         int node, zid, shrink;
3208         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3209         struct cgroup *cgrp = mem->css.cgroup;
3210
3211         css_get(&mem->css);
3212
3213         shrink = 0;
3214         /* should free all ? */
3215         if (free_all)
3216                 goto try_to_free;
3217 move_account:
3218         do {
3219                 ret = -EBUSY;
3220                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3221                         goto out;
3222                 ret = -EINTR;
3223                 if (signal_pending(current))
3224                         goto out;
3225                 /* This is for making all *used* pages to be on LRU. */
3226                 lru_add_drain_all();
3227                 drain_all_stock_sync();
3228                 ret = 0;
3229                 mem_cgroup_start_move(mem);
3230                 for_each_node_state(node, N_HIGH_MEMORY) {
3231                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3232                                 enum lru_list l;
3233                                 for_each_lru(l) {
3234                                         ret = mem_cgroup_force_empty_list(mem,
3235                                                         node, zid, l);
3236                                         if (ret)
3237                                                 break;
3238                                 }
3239                         }
3240                         if (ret)
3241                                 break;
3242                 }
3243                 mem_cgroup_end_move(mem);
3244                 memcg_oom_recover(mem);
3245                 /* it seems parent cgroup doesn't have enough mem */
3246                 if (ret == -ENOMEM)
3247                         goto try_to_free;
3248                 cond_resched();
3249         /* "ret" should also be checked to ensure all lists are empty. */
3250         } while (mem->res.usage > 0 || ret);
3251 out:
3252         css_put(&mem->css);
3253         return ret;
3254
3255 try_to_free:
3256         /* returns EBUSY if there is a task or if we come here twice. */
3257         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3258                 ret = -EBUSY;
3259                 goto out;
3260         }
3261         /* we call try-to-free pages for make this cgroup empty */
3262         lru_add_drain_all();
3263         /* try to free all pages in this cgroup */
3264         shrink = 1;
3265         while (nr_retries && mem->res.usage > 0) {
3266                 int progress;
3267
3268                 if (signal_pending(current)) {
3269                         ret = -EINTR;
3270                         goto out;
3271                 }
3272                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3273                                                 false, get_swappiness(mem));
3274                 if (!progress) {
3275                         nr_retries--;
3276                         /* maybe some writeback is necessary */
3277                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3278                 }
3279
3280         }
3281         lru_add_drain();
3282         /* try move_account...there may be some *locked* pages. */
3283         goto move_account;
3284 }
3285
3286 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3287 {
3288         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3289 }
3290
3291
3292 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3293 {
3294         return mem_cgroup_from_cont(cont)->use_hierarchy;
3295 }
3296
3297 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3298                                         u64 val)
3299 {
3300         int retval = 0;
3301         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3302         struct cgroup *parent = cont->parent;
3303         struct mem_cgroup *parent_mem = NULL;
3304
3305         if (parent)
3306                 parent_mem = mem_cgroup_from_cont(parent);
3307
3308         cgroup_lock();
3309         /*
3310          * If parent's use_hierarchy is set, we can't make any modifications
3311          * in the child subtrees. If it is unset, then the change can
3312          * occur, provided the current cgroup has no children.
3313          *
3314          * For the root cgroup, parent_mem is NULL, we allow value to be
3315          * set if there are no children.
3316          */
3317         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3318                                 (val == 1 || val == 0)) {
3319                 if (list_empty(&cont->children))
3320                         mem->use_hierarchy = val;
3321                 else
3322                         retval = -EBUSY;
3323         } else
3324                 retval = -EINVAL;
3325         cgroup_unlock();
3326
3327         return retval;
3328 }
3329
3330
3331 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3332                                 enum mem_cgroup_stat_index idx)
3333 {
3334         struct mem_cgroup *iter;
3335         s64 val = 0;
3336
3337         /* each per cpu's value can be minus.Then, use s64 */
3338         for_each_mem_cgroup_tree(iter, mem)
3339                 val += mem_cgroup_read_stat(iter, idx);
3340
3341         if (val < 0) /* race ? */
3342                 val = 0;
3343         return val;
3344 }
3345
3346 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3347 {
3348         u64 val;
3349
3350         if (!mem_cgroup_is_root(mem)) {
3351                 if (!swap)
3352                         return res_counter_read_u64(&mem->res, RES_USAGE);
3353                 else
3354                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3355         }
3356
3357         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3358         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3359
3360         if (swap)
3361                 val += mem_cgroup_get_recursive_idx_stat(mem,
3362                                 MEM_CGROUP_STAT_SWAPOUT);
3363
3364         return val << PAGE_SHIFT;
3365 }
3366
3367 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3368 {
3369         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3370         u64 val;
3371         int type, name;
3372
3373         type = MEMFILE_TYPE(cft->private);
3374         name = MEMFILE_ATTR(cft->private);
3375         switch (type) {
3376         case _MEM:
3377                 if (name == RES_USAGE)
3378                         val = mem_cgroup_usage(mem, false);
3379                 else
3380                         val = res_counter_read_u64(&mem->res, name);
3381                 break;
3382         case _MEMSWAP:
3383                 if (name == RES_USAGE)
3384                         val = mem_cgroup_usage(mem, true);
3385                 else
3386                         val = res_counter_read_u64(&mem->memsw, name);
3387                 break;
3388         default:
3389                 BUG();
3390                 break;
3391         }
3392         return val;
3393 }
3394 /*
3395  * The user of this function is...
3396  * RES_LIMIT.
3397  */
3398 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3399                             const char *buffer)
3400 {
3401         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3402         int type, name;
3403         unsigned long long val;
3404         int ret;
3405
3406         type = MEMFILE_TYPE(cft->private);
3407         name = MEMFILE_ATTR(cft->private);
3408         switch (name) {
3409         case RES_LIMIT:
3410                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3411                         ret = -EINVAL;
3412                         break;
3413                 }
3414                 /* This function does all necessary parse...reuse it */
3415                 ret = res_counter_memparse_write_strategy(buffer, &val);
3416                 if (ret)
3417                         break;
3418                 if (type == _MEM)
3419                         ret = mem_cgroup_resize_limit(memcg, val);
3420                 else
3421                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3422                 break;
3423         case RES_SOFT_LIMIT:
3424                 ret = res_counter_memparse_write_strategy(buffer, &val);
3425                 if (ret)
3426                         break;
3427                 /*
3428                  * For memsw, soft limits are hard to implement in terms
3429                  * of semantics, for now, we support soft limits for
3430                  * control without swap
3431                  */
3432                 if (type == _MEM)
3433                         ret = res_counter_set_soft_limit(&memcg->res, val);
3434                 else
3435                         ret = -EINVAL;
3436                 break;
3437         default:
3438                 ret = -EINVAL; /* should be BUG() ? */
3439                 break;
3440         }
3441         return ret;
3442 }
3443
3444 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3445                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3446 {
3447         struct cgroup *cgroup;
3448         unsigned long long min_limit, min_memsw_limit, tmp;
3449
3450         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3451         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3452         cgroup = memcg->css.cgroup;
3453         if (!memcg->use_hierarchy)
3454                 goto out;
3455
3456         while (cgroup->parent) {
3457                 cgroup = cgroup->parent;
3458                 memcg = mem_cgroup_from_cont(cgroup);
3459                 if (!memcg->use_hierarchy)
3460                         break;
3461                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3462                 min_limit = min(min_limit, tmp);
3463                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3464                 min_memsw_limit = min(min_memsw_limit, tmp);
3465         }
3466 out:
3467         *mem_limit = min_limit;
3468         *memsw_limit = min_memsw_limit;
3469         return;
3470 }
3471
3472 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3473 {
3474         struct mem_cgroup *mem;
3475         int type, name;
3476
3477         mem = mem_cgroup_from_cont(cont);
3478         type = MEMFILE_TYPE(event);
3479         name = MEMFILE_ATTR(event);
3480         switch (name) {
3481         case RES_MAX_USAGE:
3482                 if (type == _MEM)
3483                         res_counter_reset_max(&mem->res);
3484                 else
3485                         res_counter_reset_max(&mem->memsw);
3486                 break;
3487         case RES_FAILCNT:
3488                 if (type == _MEM)
3489                         res_counter_reset_failcnt(&mem->res);
3490                 else
3491                         res_counter_reset_failcnt(&mem->memsw);
3492                 break;
3493         }
3494
3495         return 0;
3496 }
3497
3498 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3499                                         struct cftype *cft)
3500 {
3501         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3502 }
3503
3504 #ifdef CONFIG_MMU
3505 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3506                                         struct cftype *cft, u64 val)
3507 {
3508         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3509
3510         if (val >= (1 << NR_MOVE_TYPE))
3511                 return -EINVAL;
3512         /*
3513          * We check this value several times in both in can_attach() and
3514          * attach(), so we need cgroup lock to prevent this value from being
3515          * inconsistent.
3516          */
3517         cgroup_lock();
3518         mem->move_charge_at_immigrate = val;
3519         cgroup_unlock();
3520
3521         return 0;
3522 }
3523 #else
3524 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3525                                         struct cftype *cft, u64 val)
3526 {
3527         return -ENOSYS;
3528 }
3529 #endif
3530
3531
3532 /* For read statistics */
3533 enum {
3534         MCS_CACHE,
3535         MCS_RSS,
3536         MCS_FILE_MAPPED,
3537         MCS_PGPGIN,
3538         MCS_PGPGOUT,
3539         MCS_SWAP,
3540         MCS_INACTIVE_ANON,
3541         MCS_ACTIVE_ANON,
3542         MCS_INACTIVE_FILE,
3543         MCS_ACTIVE_FILE,
3544         MCS_UNEVICTABLE,
3545         NR_MCS_STAT,
3546 };
3547
3548 struct mcs_total_stat {
3549         s64 stat[NR_MCS_STAT];
3550 };
3551
3552 struct {
3553         char *local_name;
3554         char *total_name;
3555 } memcg_stat_strings[NR_MCS_STAT] = {
3556         {"cache", "total_cache"},
3557         {"rss", "total_rss"},
3558         {"mapped_file", "total_mapped_file"},
3559         {"pgpgin", "total_pgpgin"},
3560         {"pgpgout", "total_pgpgout"},
3561         {"swap", "total_swap"},
3562         {"inactive_anon", "total_inactive_anon"},
3563         {"active_anon", "total_active_anon"},
3564         {"inactive_file", "total_inactive_file"},
3565         {"active_file", "total_active_file"},
3566         {"unevictable", "total_unevictable"}
3567 };
3568
3569
3570 static void
3571 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3572 {
3573         s64 val;
3574
3575         /* per cpu stat */
3576         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3577         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3578         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3579         s->stat[MCS_RSS] += val * PAGE_SIZE;
3580         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3581         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3582         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3583         s->stat[MCS_PGPGIN] += val;
3584         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3585         s->stat[MCS_PGPGOUT] += val;
3586         if (do_swap_account) {
3587                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3588                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3589         }
3590
3591         /* per zone stat */
3592         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3593         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3594         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3595         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3596         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3597         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3598         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3599         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3600         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3601         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3602 }
3603
3604 static void
3605 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3606 {
3607         struct mem_cgroup *iter;
3608
3609         for_each_mem_cgroup_tree(iter, mem)
3610                 mem_cgroup_get_local_stat(iter, s);
3611 }
3612
3613 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3614                                  struct cgroup_map_cb *cb)
3615 {
3616         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3617         struct mcs_total_stat mystat;
3618         int i;
3619
3620         memset(&mystat, 0, sizeof(mystat));
3621         mem_cgroup_get_local_stat(mem_cont, &mystat);
3622
3623         for (i = 0; i < NR_MCS_STAT; i++) {
3624                 if (i == MCS_SWAP && !do_swap_account)
3625                         continue;
3626                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3627         }
3628
3629         /* Hierarchical information */
3630         {
3631                 unsigned long long limit, memsw_limit;
3632                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3633                 cb->fill(cb, "hierarchical_memory_limit", limit);
3634                 if (do_swap_account)
3635                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3636         }
3637
3638         memset(&mystat, 0, sizeof(mystat));
3639         mem_cgroup_get_total_stat(mem_cont, &mystat);
3640         for (i = 0; i < NR_MCS_STAT; i++) {
3641                 if (i == MCS_SWAP && !do_swap_account)
3642                         continue;
3643                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3644         }
3645
3646 #ifdef CONFIG_DEBUG_VM
3647         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3648
3649         {
3650                 int nid, zid;
3651                 struct mem_cgroup_per_zone *mz;
3652                 unsigned long recent_rotated[2] = {0, 0};
3653                 unsigned long recent_scanned[2] = {0, 0};
3654
3655                 for_each_online_node(nid)
3656                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3657                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3658
3659                                 recent_rotated[0] +=
3660                                         mz->reclaim_stat.recent_rotated[0];
3661                                 recent_rotated[1] +=
3662                                         mz->reclaim_stat.recent_rotated[1];
3663                                 recent_scanned[0] +=
3664                                         mz->reclaim_stat.recent_scanned[0];
3665                                 recent_scanned[1] +=
3666                                         mz->reclaim_stat.recent_scanned[1];
3667                         }
3668                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3669                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3670                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3671                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3672         }
3673 #endif
3674
3675         return 0;
3676 }
3677
3678 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3679 {
3680         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3681
3682         return get_swappiness(memcg);
3683 }
3684
3685 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3686                                        u64 val)
3687 {
3688         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3689         struct mem_cgroup *parent;
3690
3691         if (val > 100)
3692                 return -EINVAL;
3693
3694         if (cgrp->parent == NULL)
3695                 return -EINVAL;
3696
3697         parent = mem_cgroup_from_cont(cgrp->parent);
3698
3699         cgroup_lock();
3700
3701         /* If under hierarchy, only empty-root can set this value */
3702         if ((parent->use_hierarchy) ||
3703             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3704                 cgroup_unlock();
3705                 return -EINVAL;
3706         }
3707
3708         spin_lock(&memcg->reclaim_param_lock);
3709         memcg->swappiness = val;
3710         spin_unlock(&memcg->reclaim_param_lock);
3711
3712         cgroup_unlock();
3713
3714         return 0;
3715 }
3716
3717 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3718 {
3719         struct mem_cgroup_threshold_ary *t;
3720         u64 usage;
3721         int i;
3722
3723         rcu_read_lock();
3724         if (!swap)
3725                 t = rcu_dereference(memcg->thresholds.primary);
3726         else
3727                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3728
3729         if (!t)
3730                 goto unlock;
3731
3732         usage = mem_cgroup_usage(memcg, swap);
3733
3734         /*
3735          * current_threshold points to threshold just below usage.
3736          * If it's not true, a threshold was crossed after last
3737          * call of __mem_cgroup_threshold().
3738          */
3739         i = t->current_threshold;
3740
3741         /*
3742          * Iterate backward over array of thresholds starting from
3743          * current_threshold and check if a threshold is crossed.
3744          * If none of thresholds below usage is crossed, we read
3745          * only one element of the array here.
3746          */
3747         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3748                 eventfd_signal(t->entries[i].eventfd, 1);
3749
3750         /* i = current_threshold + 1 */
3751         i++;
3752
3753         /*
3754          * Iterate forward over array of thresholds starting from
3755          * current_threshold+1 and check if a threshold is crossed.
3756          * If none of thresholds above usage is crossed, we read
3757          * only one element of the array here.
3758          */
3759         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3760                 eventfd_signal(t->entries[i].eventfd, 1);
3761
3762         /* Update current_threshold */
3763         t->current_threshold = i - 1;
3764 unlock:
3765         rcu_read_unlock();
3766 }
3767
3768 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3769 {
3770         while (memcg) {
3771                 __mem_cgroup_threshold(memcg, false);
3772                 if (do_swap_account)
3773                         __mem_cgroup_threshold(memcg, true);
3774
3775                 memcg = parent_mem_cgroup(memcg);
3776         }
3777 }
3778
3779 static int compare_thresholds(const void *a, const void *b)
3780 {
3781         const struct mem_cgroup_threshold *_a = a;
3782         const struct mem_cgroup_threshold *_b = b;
3783
3784         return _a->threshold - _b->threshold;
3785 }
3786
3787 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3788 {
3789         struct mem_cgroup_eventfd_list *ev;
3790
3791         list_for_each_entry(ev, &mem->oom_notify, list)
3792                 eventfd_signal(ev->eventfd, 1);
3793         return 0;
3794 }
3795
3796 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3797 {
3798         struct mem_cgroup *iter;
3799
3800         for_each_mem_cgroup_tree(iter, mem)
3801                 mem_cgroup_oom_notify_cb(iter);
3802 }
3803
3804 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3805         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3806 {
3807         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3808         struct mem_cgroup_thresholds *thresholds;
3809         struct mem_cgroup_threshold_ary *new;
3810         int type = MEMFILE_TYPE(cft->private);
3811         u64 threshold, usage;
3812         int i, size, ret;
3813
3814         ret = res_counter_memparse_write_strategy(args, &threshold);
3815         if (ret)
3816                 return ret;
3817
3818         mutex_lock(&memcg->thresholds_lock);
3819
3820         if (type == _MEM)
3821                 thresholds = &memcg->thresholds;
3822         else if (type == _MEMSWAP)
3823                 thresholds = &memcg->memsw_thresholds;
3824         else
3825                 BUG();
3826
3827         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3828
3829         /* Check if a threshold crossed before adding a new one */
3830         if (thresholds->primary)
3831                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3832
3833         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3834
3835         /* Allocate memory for new array of thresholds */
3836         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3837                         GFP_KERNEL);
3838         if (!new) {
3839                 ret = -ENOMEM;
3840                 goto unlock;
3841         }
3842         new->size = size;
3843
3844         /* Copy thresholds (if any) to new array */
3845         if (thresholds->primary) {
3846                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3847                                 sizeof(struct mem_cgroup_threshold));
3848         }
3849
3850         /* Add new threshold */
3851         new->entries[size - 1].eventfd = eventfd;
3852         new->entries[size - 1].threshold = threshold;
3853
3854         /* Sort thresholds. Registering of new threshold isn't time-critical */
3855         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3856                         compare_thresholds, NULL);
3857
3858         /* Find current threshold */
3859         new->current_threshold = -1;
3860         for (i = 0; i < size; i++) {
3861                 if (new->entries[i].threshold < usage) {
3862                         /*
3863                          * new->current_threshold will not be used until
3864                          * rcu_assign_pointer(), so it's safe to increment
3865                          * it here.
3866                          */
3867                         ++new->current_threshold;
3868                 }
3869         }
3870
3871         /* Free old spare buffer and save old primary buffer as spare */
3872         kfree(thresholds->spare);
3873         thresholds->spare = thresholds->primary;
3874
3875         rcu_assign_pointer(thresholds->primary, new);
3876
3877         /* To be sure that nobody uses thresholds */
3878         synchronize_rcu();
3879
3880 unlock:
3881         mutex_unlock(&memcg->thresholds_lock);
3882
3883         return ret;
3884 }
3885
3886 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3887         struct cftype *cft, struct eventfd_ctx *eventfd)
3888 {
3889         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3890         struct mem_cgroup_thresholds *thresholds;
3891         struct mem_cgroup_threshold_ary *new;
3892         int type = MEMFILE_TYPE(cft->private);
3893         u64 usage;
3894         int i, j, size;
3895
3896         mutex_lock(&memcg->thresholds_lock);
3897         if (type == _MEM)
3898                 thresholds = &memcg->thresholds;
3899         else if (type == _MEMSWAP)
3900                 thresholds = &memcg->memsw_thresholds;
3901         else
3902                 BUG();
3903
3904         /*
3905          * Something went wrong if we trying to unregister a threshold
3906          * if we don't have thresholds
3907          */
3908         BUG_ON(!thresholds);
3909
3910         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3911
3912         /* Check if a threshold crossed before removing */
3913         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3914
3915         /* Calculate new number of threshold */
3916         size = 0;
3917         for (i = 0; i < thresholds->primary->size; i++) {
3918                 if (thresholds->primary->entries[i].eventfd != eventfd)
3919                         size++;
3920         }
3921
3922         new = thresholds->spare;
3923
3924         /* Set thresholds array to NULL if we don't have thresholds */
3925         if (!size) {
3926                 kfree(new);
3927                 new = NULL;
3928                 goto swap_buffers;
3929         }
3930
3931         new->size = size;
3932
3933         /* Copy thresholds and find current threshold */
3934         new->current_threshold = -1;
3935         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3936                 if (thresholds->primary->entries[i].eventfd == eventfd)
3937                         continue;
3938
3939                 new->entries[j] = thresholds->primary->entries[i];
3940                 if (new->entries[j].threshold < usage) {
3941                         /*
3942                          * new->current_threshold will not be used
3943                          * until rcu_assign_pointer(), so it's safe to increment
3944                          * it here.
3945                          */
3946                         ++new->current_threshold;
3947                 }
3948                 j++;
3949         }
3950
3951 swap_buffers:
3952         /* Swap primary and spare array */
3953         thresholds->spare = thresholds->primary;
3954         rcu_assign_pointer(thresholds->primary, new);
3955
3956         /* To be sure that nobody uses thresholds */
3957         synchronize_rcu();
3958
3959         mutex_unlock(&memcg->thresholds_lock);
3960 }
3961
3962 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3963         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3964 {
3965         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3966         struct mem_cgroup_eventfd_list *event;
3967         int type = MEMFILE_TYPE(cft->private);
3968
3969         BUG_ON(type != _OOM_TYPE);
3970         event = kmalloc(sizeof(*event), GFP_KERNEL);
3971         if (!event)
3972                 return -ENOMEM;
3973
3974         mutex_lock(&memcg_oom_mutex);
3975
3976         event->eventfd = eventfd;
3977         list_add(&event->list, &memcg->oom_notify);
3978
3979         /* already in OOM ? */
3980         if (atomic_read(&memcg->oom_lock))
3981                 eventfd_signal(eventfd, 1);
3982         mutex_unlock(&memcg_oom_mutex);
3983
3984         return 0;
3985 }
3986
3987 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3988         struct cftype *cft, struct eventfd_ctx *eventfd)
3989 {
3990         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3991         struct mem_cgroup_eventfd_list *ev, *tmp;
3992         int type = MEMFILE_TYPE(cft->private);
3993
3994         BUG_ON(type != _OOM_TYPE);
3995
3996         mutex_lock(&memcg_oom_mutex);
3997
3998         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3999                 if (ev->eventfd == eventfd) {
4000                         list_del(&ev->list);
4001                         kfree(ev);
4002                 }
4003         }
4004
4005         mutex_unlock(&memcg_oom_mutex);
4006 }
4007
4008 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4009         struct cftype *cft,  struct cgroup_map_cb *cb)
4010 {
4011         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4012
4013         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4014
4015         if (atomic_read(&mem->oom_lock))
4016                 cb->fill(cb, "under_oom", 1);
4017         else
4018                 cb->fill(cb, "under_oom", 0);
4019         return 0;
4020 }
4021
4022 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4023         struct cftype *cft, u64 val)
4024 {
4025         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4026         struct mem_cgroup *parent;
4027
4028         /* cannot set to root cgroup and only 0 and 1 are allowed */
4029         if (!cgrp->parent || !((val == 0) || (val == 1)))
4030                 return -EINVAL;
4031
4032         parent = mem_cgroup_from_cont(cgrp->parent);
4033
4034         cgroup_lock();
4035         /* oom-kill-disable is a flag for subhierarchy. */
4036         if ((parent->use_hierarchy) ||
4037             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4038                 cgroup_unlock();
4039                 return -EINVAL;
4040         }
4041         mem->oom_kill_disable = val;
4042         if (!val)
4043                 memcg_oom_recover(mem);
4044         cgroup_unlock();
4045         return 0;
4046 }
4047
4048 static struct cftype mem_cgroup_files[] = {
4049         {
4050                 .name = "usage_in_bytes",
4051                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4052                 .read_u64 = mem_cgroup_read,
4053                 .register_event = mem_cgroup_usage_register_event,
4054                 .unregister_event = mem_cgroup_usage_unregister_event,
4055         },
4056         {
4057                 .name = "max_usage_in_bytes",
4058                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4059                 .trigger = mem_cgroup_reset,
4060                 .read_u64 = mem_cgroup_read,
4061         },
4062         {
4063                 .name = "limit_in_bytes",
4064                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4065                 .write_string = mem_cgroup_write,
4066                 .read_u64 = mem_cgroup_read,
4067         },
4068         {
4069                 .name = "soft_limit_in_bytes",
4070                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4071                 .write_string = mem_cgroup_write,
4072                 .read_u64 = mem_cgroup_read,
4073         },
4074         {
4075                 .name = "failcnt",
4076                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4077                 .trigger = mem_cgroup_reset,
4078                 .read_u64 = mem_cgroup_read,
4079         },
4080         {
4081                 .name = "stat",
4082                 .read_map = mem_control_stat_show,
4083         },
4084         {
4085                 .name = "force_empty",
4086                 .trigger = mem_cgroup_force_empty_write,
4087         },
4088         {
4089                 .name = "use_hierarchy",
4090                 .write_u64 = mem_cgroup_hierarchy_write,
4091                 .read_u64 = mem_cgroup_hierarchy_read,
4092         },
4093         {
4094                 .name = "swappiness",
4095                 .read_u64 = mem_cgroup_swappiness_read,
4096                 .write_u64 = mem_cgroup_swappiness_write,
4097         },
4098         {
4099                 .name = "move_charge_at_immigrate",
4100                 .read_u64 = mem_cgroup_move_charge_read,
4101                 .write_u64 = mem_cgroup_move_charge_write,
4102         },
4103         {
4104                 .name = "oom_control",
4105                 .read_map = mem_cgroup_oom_control_read,
4106                 .write_u64 = mem_cgroup_oom_control_write,
4107                 .register_event = mem_cgroup_oom_register_event,
4108                 .unregister_event = mem_cgroup_oom_unregister_event,
4109                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4110         },
4111 };
4112
4113 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4114 static struct cftype memsw_cgroup_files[] = {
4115         {
4116                 .name = "memsw.usage_in_bytes",
4117                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4118                 .read_u64 = mem_cgroup_read,
4119                 .register_event = mem_cgroup_usage_register_event,
4120                 .unregister_event = mem_cgroup_usage_unregister_event,
4121         },
4122         {
4123                 .name = "memsw.max_usage_in_bytes",
4124                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4125                 .trigger = mem_cgroup_reset,
4126                 .read_u64 = mem_cgroup_read,
4127         },
4128         {
4129                 .name = "memsw.limit_in_bytes",
4130                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4131                 .write_string = mem_cgroup_write,
4132                 .read_u64 = mem_cgroup_read,
4133         },
4134         {
4135                 .name = "memsw.failcnt",
4136                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4137                 .trigger = mem_cgroup_reset,
4138                 .read_u64 = mem_cgroup_read,
4139         },
4140 };
4141
4142 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4143 {
4144         if (!do_swap_account)
4145                 return 0;
4146         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4147                                 ARRAY_SIZE(memsw_cgroup_files));
4148 };
4149 #else
4150 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4151 {
4152         return 0;
4153 }
4154 #endif
4155
4156 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4157 {
4158         struct mem_cgroup_per_node *pn;
4159         struct mem_cgroup_per_zone *mz;
4160         enum lru_list l;
4161         int zone, tmp = node;
4162         /*
4163          * This routine is called against possible nodes.
4164          * But it's BUG to call kmalloc() against offline node.
4165          *
4166          * TODO: this routine can waste much memory for nodes which will
4167          *       never be onlined. It's better to use memory hotplug callback
4168          *       function.
4169          */
4170         if (!node_state(node, N_NORMAL_MEMORY))
4171                 tmp = -1;
4172         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4173         if (!pn)
4174                 return 1;
4175
4176         mem->info.nodeinfo[node] = pn;
4177         memset(pn, 0, sizeof(*pn));
4178
4179         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4180                 mz = &pn->zoneinfo[zone];
4181                 for_each_lru(l)
4182                         INIT_LIST_HEAD(&mz->lists[l]);
4183                 mz->usage_in_excess = 0;
4184                 mz->on_tree = false;
4185                 mz->mem = mem;
4186         }
4187         return 0;
4188 }
4189
4190 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4191 {
4192         kfree(mem->info.nodeinfo[node]);
4193 }
4194
4195 static struct mem_cgroup *mem_cgroup_alloc(void)
4196 {
4197         struct mem_cgroup *mem;
4198         int size = sizeof(struct mem_cgroup);
4199
4200         /* Can be very big if MAX_NUMNODES is very big */
4201         if (size < PAGE_SIZE)
4202                 mem = kmalloc(size, GFP_KERNEL);
4203         else
4204                 mem = vmalloc(size);
4205
4206         if (!mem)
4207                 return NULL;
4208
4209         memset(mem, 0, size);
4210         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4211         if (!mem->stat) {
4212                 if (size < PAGE_SIZE)
4213                         kfree(mem);
4214                 else
4215                         vfree(mem);
4216                 mem = NULL;
4217         }
4218         spin_lock_init(&mem->pcp_counter_lock);
4219         return mem;
4220 }
4221
4222 /*
4223  * At destroying mem_cgroup, references from swap_cgroup can remain.
4224  * (scanning all at force_empty is too costly...)
4225  *
4226  * Instead of clearing all references at force_empty, we remember
4227  * the number of reference from swap_cgroup and free mem_cgroup when
4228  * it goes down to 0.
4229  *
4230  * Removal of cgroup itself succeeds regardless of refs from swap.
4231  */
4232
4233 static void __mem_cgroup_free(struct mem_cgroup *mem)
4234 {
4235         int node;
4236
4237         mem_cgroup_remove_from_trees(mem);
4238         free_css_id(&mem_cgroup_subsys, &mem->css);
4239
4240         for_each_node_state(node, N_POSSIBLE)
4241                 free_mem_cgroup_per_zone_info(mem, node);
4242
4243         free_percpu(mem->stat);
4244         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4245                 kfree(mem);
4246         else
4247                 vfree(mem);
4248 }
4249
4250 static void mem_cgroup_get(struct mem_cgroup *mem)
4251 {
4252         atomic_inc(&mem->refcnt);
4253 }
4254
4255 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4256 {
4257         if (atomic_sub_and_test(count, &mem->refcnt)) {
4258                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4259                 __mem_cgroup_free(mem);
4260                 if (parent)
4261                         mem_cgroup_put(parent);
4262         }
4263 }
4264
4265 static void mem_cgroup_put(struct mem_cgroup *mem)
4266 {
4267         __mem_cgroup_put(mem, 1);
4268 }
4269
4270 /*
4271  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4272  */
4273 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4274 {
4275         if (!mem->res.parent)
4276                 return NULL;
4277         return mem_cgroup_from_res_counter(mem->res.parent, res);
4278 }
4279
4280 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4281 static void __init enable_swap_cgroup(void)
4282 {
4283         if (!mem_cgroup_disabled() && really_do_swap_account)
4284                 do_swap_account = 1;
4285 }
4286 #else
4287 static void __init enable_swap_cgroup(void)
4288 {
4289 }
4290 #endif
4291
4292 static int mem_cgroup_soft_limit_tree_init(void)
4293 {
4294         struct mem_cgroup_tree_per_node *rtpn;
4295         struct mem_cgroup_tree_per_zone *rtpz;
4296         int tmp, node, zone;
4297
4298         for_each_node_state(node, N_POSSIBLE) {
4299                 tmp = node;
4300                 if (!node_state(node, N_NORMAL_MEMORY))
4301                         tmp = -1;
4302                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4303                 if (!rtpn)
4304                         return 1;
4305
4306                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4307
4308                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4309                         rtpz = &rtpn->rb_tree_per_zone[zone];
4310                         rtpz->rb_root = RB_ROOT;
4311                         spin_lock_init(&rtpz->lock);
4312                 }
4313         }
4314         return 0;
4315 }
4316
4317 static struct cgroup_subsys_state * __ref
4318 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4319 {
4320         struct mem_cgroup *mem, *parent;
4321         long error = -ENOMEM;
4322         int node;
4323
4324         mem = mem_cgroup_alloc();
4325         if (!mem)
4326                 return ERR_PTR(error);
4327
4328         for_each_node_state(node, N_POSSIBLE)
4329                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4330                         goto free_out;
4331
4332         /* root ? */
4333         if (cont->parent == NULL) {
4334                 int cpu;
4335                 enable_swap_cgroup();
4336                 parent = NULL;
4337                 root_mem_cgroup = mem;
4338                 if (mem_cgroup_soft_limit_tree_init())
4339                         goto free_out;
4340                 for_each_possible_cpu(cpu) {
4341                         struct memcg_stock_pcp *stock =
4342                                                 &per_cpu(memcg_stock, cpu);
4343                         INIT_WORK(&stock->work, drain_local_stock);
4344                 }
4345                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4346         } else {
4347                 parent = mem_cgroup_from_cont(cont->parent);
4348                 mem->use_hierarchy = parent->use_hierarchy;
4349                 mem->oom_kill_disable = parent->oom_kill_disable;
4350         }
4351
4352         if (parent && parent->use_hierarchy) {
4353                 res_counter_init(&mem->res, &parent->res);
4354                 res_counter_init(&mem->memsw, &parent->memsw);
4355                 /*
4356                  * We increment refcnt of the parent to ensure that we can
4357                  * safely access it on res_counter_charge/uncharge.
4358                  * This refcnt will be decremented when freeing this
4359                  * mem_cgroup(see mem_cgroup_put).
4360                  */
4361                 mem_cgroup_get(parent);
4362         } else {
4363                 res_counter_init(&mem->res, NULL);
4364                 res_counter_init(&mem->memsw, NULL);
4365         }
4366         mem->last_scanned_child = 0;
4367         spin_lock_init(&mem->reclaim_param_lock);
4368         INIT_LIST_HEAD(&mem->oom_notify);
4369
4370         if (parent)
4371                 mem->swappiness = get_swappiness(parent);
4372         atomic_set(&mem->refcnt, 1);
4373         mem->move_charge_at_immigrate = 0;
4374         mutex_init(&mem->thresholds_lock);
4375         return &mem->css;
4376 free_out:
4377         __mem_cgroup_free(mem);
4378         root_mem_cgroup = NULL;
4379         return ERR_PTR(error);
4380 }
4381
4382 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4383                                         struct cgroup *cont)
4384 {
4385         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4386
4387         return mem_cgroup_force_empty(mem, false);
4388 }
4389
4390 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4391                                 struct cgroup *cont)
4392 {
4393         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4394
4395         mem_cgroup_put(mem);
4396 }
4397
4398 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4399                                 struct cgroup *cont)
4400 {
4401         int ret;
4402
4403         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4404                                 ARRAY_SIZE(mem_cgroup_files));
4405
4406         if (!ret)
4407                 ret = register_memsw_files(cont, ss);
4408         return ret;
4409 }
4410
4411 #ifdef CONFIG_MMU
4412 /* Handlers for move charge at task migration. */
4413 #define PRECHARGE_COUNT_AT_ONCE 256
4414 static int mem_cgroup_do_precharge(unsigned long count)
4415 {
4416         int ret = 0;
4417         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4418         struct mem_cgroup *mem = mc.to;
4419
4420         if (mem_cgroup_is_root(mem)) {
4421                 mc.precharge += count;
4422                 /* we don't need css_get for root */
4423                 return ret;
4424         }
4425         /* try to charge at once */
4426         if (count > 1) {
4427                 struct res_counter *dummy;
4428                 /*
4429                  * "mem" cannot be under rmdir() because we've already checked
4430                  * by cgroup_lock_live_cgroup() that it is not removed and we
4431                  * are still under the same cgroup_mutex. So we can postpone
4432                  * css_get().
4433                  */
4434                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4435                         goto one_by_one;
4436                 if (do_swap_account && res_counter_charge(&mem->memsw,
4437                                                 PAGE_SIZE * count, &dummy)) {
4438                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4439                         goto one_by_one;
4440                 }
4441                 mc.precharge += count;
4442                 return ret;
4443         }
4444 one_by_one:
4445         /* fall back to one by one charge */
4446         while (count--) {
4447                 if (signal_pending(current)) {
4448                         ret = -EINTR;
4449                         break;
4450                 }
4451                 if (!batch_count--) {
4452                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4453                         cond_resched();
4454                 }
4455                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4456                 if (ret || !mem)
4457                         /* mem_cgroup_clear_mc() will do uncharge later */
4458                         return -ENOMEM;
4459                 mc.precharge++;
4460         }
4461         return ret;
4462 }
4463
4464 /**
4465  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4466  * @vma: the vma the pte to be checked belongs
4467  * @addr: the address corresponding to the pte to be checked
4468  * @ptent: the pte to be checked
4469  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4470  *
4471  * Returns
4472  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4473  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4474  *     move charge. if @target is not NULL, the page is stored in target->page
4475  *     with extra refcnt got(Callers should handle it).
4476  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4477  *     target for charge migration. if @target is not NULL, the entry is stored
4478  *     in target->ent.
4479  *
4480  * Called with pte lock held.
4481  */
4482 union mc_target {
4483         struct page     *page;
4484         swp_entry_t     ent;
4485 };
4486
4487 enum mc_target_type {
4488         MC_TARGET_NONE, /* not used */
4489         MC_TARGET_PAGE,
4490         MC_TARGET_SWAP,
4491 };
4492
4493 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4494                                                 unsigned long addr, pte_t ptent)
4495 {
4496         struct page *page = vm_normal_page(vma, addr, ptent);
4497
4498         if (!page || !page_mapped(page))
4499                 return NULL;
4500         if (PageAnon(page)) {
4501                 /* we don't move shared anon */
4502                 if (!move_anon() || page_mapcount(page) > 2)
4503                         return NULL;
4504         } else if (!move_file())
4505                 /* we ignore mapcount for file pages */
4506                 return NULL;
4507         if (!get_page_unless_zero(page))
4508                 return NULL;
4509
4510         return page;
4511 }
4512
4513 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4514                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4515 {
4516         int usage_count;
4517         struct page *page = NULL;
4518         swp_entry_t ent = pte_to_swp_entry(ptent);
4519
4520         if (!move_anon() || non_swap_entry(ent))
4521                 return NULL;
4522         usage_count = mem_cgroup_count_swap_user(ent, &page);
4523         if (usage_count > 1) { /* we don't move shared anon */
4524                 if (page)
4525                         put_page(page);
4526                 return NULL;
4527         }
4528         if (do_swap_account)
4529                 entry->val = ent.val;
4530
4531         return page;
4532 }
4533
4534 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4535                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4536 {
4537         struct page *page = NULL;
4538         struct inode *inode;
4539         struct address_space *mapping;
4540         pgoff_t pgoff;
4541
4542         if (!vma->vm_file) /* anonymous vma */
4543                 return NULL;
4544         if (!move_file())
4545                 return NULL;
4546
4547         inode = vma->vm_file->f_path.dentry->d_inode;
4548         mapping = vma->vm_file->f_mapping;
4549         if (pte_none(ptent))
4550                 pgoff = linear_page_index(vma, addr);
4551         else /* pte_file(ptent) is true */
4552                 pgoff = pte_to_pgoff(ptent);
4553
4554         /* page is moved even if it's not RSS of this task(page-faulted). */
4555         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4556                 page = find_get_page(mapping, pgoff);
4557         } else { /* shmem/tmpfs file. we should take account of swap too. */
4558                 swp_entry_t ent;
4559                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4560                 if (do_swap_account)
4561                         entry->val = ent.val;
4562         }
4563
4564         return page;
4565 }
4566
4567 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4568                 unsigned long addr, pte_t ptent, union mc_target *target)
4569 {
4570         struct page *page = NULL;
4571         struct page_cgroup *pc;
4572         int ret = 0;
4573         swp_entry_t ent = { .val = 0 };
4574
4575         if (pte_present(ptent))
4576                 page = mc_handle_present_pte(vma, addr, ptent);
4577         else if (is_swap_pte(ptent))
4578                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4579         else if (pte_none(ptent) || pte_file(ptent))
4580                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4581
4582         if (!page && !ent.val)
4583                 return 0;
4584         if (page) {
4585                 pc = lookup_page_cgroup(page);
4586                 /*
4587                  * Do only loose check w/o page_cgroup lock.
4588                  * mem_cgroup_move_account() checks the pc is valid or not under
4589                  * the lock.
4590                  */
4591                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4592                         ret = MC_TARGET_PAGE;
4593                         if (target)
4594                                 target->page = page;
4595                 }
4596                 if (!ret || !target)
4597                         put_page(page);
4598         }
4599         /* There is a swap entry and a page doesn't exist or isn't charged */
4600         if (ent.val && !ret &&
4601                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4602                 ret = MC_TARGET_SWAP;
4603                 if (target)
4604                         target->ent = ent;
4605         }
4606         return ret;
4607 }
4608
4609 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4610                                         unsigned long addr, unsigned long end,
4611                                         struct mm_walk *walk)
4612 {
4613         struct vm_area_struct *vma = walk->private;
4614         pte_t *pte;
4615         spinlock_t *ptl;
4616
4617         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4618         for (; addr != end; pte++, addr += PAGE_SIZE)
4619                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4620                         mc.precharge++; /* increment precharge temporarily */
4621         pte_unmap_unlock(pte - 1, ptl);
4622         cond_resched();
4623
4624         return 0;
4625 }
4626
4627 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4628 {
4629         unsigned long precharge;
4630         struct vm_area_struct *vma;
4631
4632         down_read(&mm->mmap_sem);
4633         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4634                 struct mm_walk mem_cgroup_count_precharge_walk = {
4635                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4636                         .mm = mm,
4637                         .private = vma,
4638                 };
4639                 if (is_vm_hugetlb_page(vma))
4640                         continue;
4641                 walk_page_range(vma->vm_start, vma->vm_end,
4642                                         &mem_cgroup_count_precharge_walk);
4643         }
4644         up_read(&mm->mmap_sem);
4645
4646         precharge = mc.precharge;
4647         mc.precharge = 0;
4648
4649         return precharge;
4650 }
4651
4652 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4653 {
4654         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4655 }
4656
4657 static void mem_cgroup_clear_mc(void)
4658 {
4659         struct mem_cgroup *from = mc.from;
4660         struct mem_cgroup *to = mc.to;
4661
4662         /* we must uncharge all the leftover precharges from mc.to */
4663         if (mc.precharge) {
4664                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4665                 mc.precharge = 0;
4666         }
4667         /*
4668          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4669          * we must uncharge here.
4670          */
4671         if (mc.moved_charge) {
4672                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4673                 mc.moved_charge = 0;
4674         }
4675         /* we must fixup refcnts and charges */
4676         if (mc.moved_swap) {
4677                 /* uncharge swap account from the old cgroup */
4678                 if (!mem_cgroup_is_root(mc.from))
4679                         res_counter_uncharge(&mc.from->memsw,
4680                                                 PAGE_SIZE * mc.moved_swap);
4681                 __mem_cgroup_put(mc.from, mc.moved_swap);
4682
4683                 if (!mem_cgroup_is_root(mc.to)) {
4684                         /*
4685                          * we charged both to->res and to->memsw, so we should
4686                          * uncharge to->res.
4687                          */
4688                         res_counter_uncharge(&mc.to->res,
4689                                                 PAGE_SIZE * mc.moved_swap);
4690                 }
4691                 /* we've already done mem_cgroup_get(mc.to) */
4692
4693                 mc.moved_swap = 0;
4694         }
4695         spin_lock(&mc.lock);
4696         mc.from = NULL;
4697         mc.to = NULL;
4698         mc.moving_task = NULL;
4699         spin_unlock(&mc.lock);
4700         mem_cgroup_end_move(from);
4701         memcg_oom_recover(from);
4702         memcg_oom_recover(to);
4703         wake_up_all(&mc.waitq);
4704 }
4705
4706 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4707                                 struct cgroup *cgroup,
4708                                 struct task_struct *p,
4709                                 bool threadgroup)
4710 {
4711         int ret = 0;
4712         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4713
4714         if (mem->move_charge_at_immigrate) {
4715                 struct mm_struct *mm;
4716                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4717
4718                 VM_BUG_ON(from == mem);
4719
4720                 mm = get_task_mm(p);
4721                 if (!mm)
4722                         return 0;
4723                 /* We move charges only when we move a owner of the mm */
4724                 if (mm->owner == p) {
4725                         VM_BUG_ON(mc.from);
4726                         VM_BUG_ON(mc.to);
4727                         VM_BUG_ON(mc.precharge);
4728                         VM_BUG_ON(mc.moved_charge);
4729                         VM_BUG_ON(mc.moved_swap);
4730                         VM_BUG_ON(mc.moving_task);
4731                         mem_cgroup_start_move(from);
4732                         spin_lock(&mc.lock);
4733                         mc.from = from;
4734                         mc.to = mem;
4735                         mc.precharge = 0;
4736                         mc.moved_charge = 0;
4737                         mc.moved_swap = 0;
4738                         mc.moving_task = current;
4739                         spin_unlock(&mc.lock);
4740
4741                         ret = mem_cgroup_precharge_mc(mm);
4742                         if (ret)
4743                                 mem_cgroup_clear_mc();
4744                 }
4745                 mmput(mm);
4746         }
4747         return ret;
4748 }
4749
4750 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4751                                 struct cgroup *cgroup,
4752                                 struct task_struct *p,
4753                                 bool threadgroup)
4754 {
4755         mem_cgroup_clear_mc();
4756 }
4757
4758 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4759                                 unsigned long addr, unsigned long end,
4760                                 struct mm_walk *walk)
4761 {
4762         int ret = 0;
4763         struct vm_area_struct *vma = walk->private;
4764         pte_t *pte;
4765         spinlock_t *ptl;
4766
4767 retry:
4768         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4769         for (; addr != end; addr += PAGE_SIZE) {
4770                 pte_t ptent = *(pte++);
4771                 union mc_target target;
4772                 int type;
4773                 struct page *page;
4774                 struct page_cgroup *pc;
4775                 swp_entry_t ent;
4776
4777                 if (!mc.precharge)
4778                         break;
4779
4780                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4781                 switch (type) {
4782                 case MC_TARGET_PAGE:
4783                         page = target.page;
4784                         if (isolate_lru_page(page))
4785                                 goto put;
4786                         pc = lookup_page_cgroup(page);
4787                         if (!mem_cgroup_move_account(pc,
4788                                                 mc.from, mc.to, false)) {
4789                                 mc.precharge--;
4790                                 /* we uncharge from mc.from later. */
4791                                 mc.moved_charge++;
4792                         }
4793                         putback_lru_page(page);
4794 put:                    /* is_target_pte_for_mc() gets the page */
4795                         put_page(page);
4796                         break;
4797                 case MC_TARGET_SWAP:
4798                         ent = target.ent;
4799                         if (!mem_cgroup_move_swap_account(ent,
4800                                                 mc.from, mc.to, false)) {
4801                                 mc.precharge--;
4802                                 /* we fixup refcnts and charges later. */
4803                                 mc.moved_swap++;
4804                         }
4805                         break;
4806                 default:
4807                         break;
4808                 }
4809         }
4810         pte_unmap_unlock(pte - 1, ptl);
4811         cond_resched();
4812
4813         if (addr != end) {
4814                 /*
4815                  * We have consumed all precharges we got in can_attach().
4816                  * We try charge one by one, but don't do any additional
4817                  * charges to mc.to if we have failed in charge once in attach()
4818                  * phase.
4819                  */
4820                 ret = mem_cgroup_do_precharge(1);
4821                 if (!ret)
4822                         goto retry;
4823         }
4824
4825         return ret;
4826 }
4827
4828 static void mem_cgroup_move_charge(struct mm_struct *mm)
4829 {
4830         struct vm_area_struct *vma;
4831
4832         lru_add_drain_all();
4833         down_read(&mm->mmap_sem);
4834         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4835                 int ret;
4836                 struct mm_walk mem_cgroup_move_charge_walk = {
4837                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4838                         .mm = mm,
4839                         .private = vma,
4840                 };
4841                 if (is_vm_hugetlb_page(vma))
4842                         continue;
4843                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4844                                                 &mem_cgroup_move_charge_walk);
4845                 if (ret)
4846                         /*
4847                          * means we have consumed all precharges and failed in
4848                          * doing additional charge. Just abandon here.
4849                          */
4850                         break;
4851         }
4852         up_read(&mm->mmap_sem);
4853 }
4854
4855 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4856                                 struct cgroup *cont,
4857                                 struct cgroup *old_cont,
4858                                 struct task_struct *p,
4859                                 bool threadgroup)
4860 {
4861         struct mm_struct *mm;
4862
4863         if (!mc.to)
4864                 /* no need to move charge */
4865                 return;
4866
4867         mm = get_task_mm(p);
4868         if (mm) {
4869                 mem_cgroup_move_charge(mm);
4870                 mmput(mm);
4871         }
4872         mem_cgroup_clear_mc();
4873 }
4874 #else   /* !CONFIG_MMU */
4875 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4876                                 struct cgroup *cgroup,
4877                                 struct task_struct *p,
4878                                 bool threadgroup)
4879 {
4880         return 0;
4881 }
4882 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4883                                 struct cgroup *cgroup,
4884                                 struct task_struct *p,
4885                                 bool threadgroup)
4886 {
4887 }
4888 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4889                                 struct cgroup *cont,
4890                                 struct cgroup *old_cont,
4891                                 struct task_struct *p,
4892                                 bool threadgroup)
4893 {
4894 }
4895 #endif
4896
4897 struct cgroup_subsys mem_cgroup_subsys = {
4898         .name = "memory",
4899         .subsys_id = mem_cgroup_subsys_id,
4900         .create = mem_cgroup_create,
4901         .pre_destroy = mem_cgroup_pre_destroy,
4902         .destroy = mem_cgroup_destroy,
4903         .populate = mem_cgroup_populate,
4904         .can_attach = mem_cgroup_can_attach,
4905         .cancel_attach = mem_cgroup_cancel_attach,
4906         .attach = mem_cgroup_move_task,
4907         .early_init = 0,
4908         .use_id = 1,
4909 };
4910
4911 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4912
4913 static int __init disable_swap_account(char *s)
4914 {
4915         really_do_swap_account = 0;
4916         return 1;
4917 }
4918 __setup("noswapaccount", disable_swap_account);
4919 #endif