memcg: cpu hotplug aware percpu count updates
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2 of the License, or
16  * (at your option) any later version.
17  *
18  * This program is distributed in the hope that it will be useful,
19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
21  * GNU General Public License for more details.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
44 #include <linux/fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
51 #include "internal.h"
52
53 #include <asm/uaccess.h>
54
55 #include <trace/events/vmscan.h>
56
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES      5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
60
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
64 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         /* Because this is for moving account, reuse mc.lock */
1136         spin_lock(&mc.lock);
1137         for_each_possible_cpu(cpu)
1138                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1139         spin_unlock(&mc.lock);
1140
1141         synchronize_rcu();
1142 }
1143
1144 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1145 {
1146         int cpu;
1147
1148         if (!mem)
1149                 return;
1150         spin_lock(&mc.lock);
1151         for_each_possible_cpu(cpu)
1152                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1153         spin_unlock(&mc.lock);
1154 }
1155 /*
1156  * 2 routines for checking "mem" is under move_account() or not.
1157  *
1158  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1159  *                        for avoiding race in accounting. If true,
1160  *                        pc->mem_cgroup may be overwritten.
1161  *
1162  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1163  *                        under hierarchy of moving cgroups. This is for
1164  *                        waiting at hith-memory prressure caused by "move".
1165  */
1166
1167 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1168 {
1169         VM_BUG_ON(!rcu_read_lock_held());
1170         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1171 }
1172
1173 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1174 {
1175         struct mem_cgroup *from;
1176         struct mem_cgroup *to;
1177         bool ret = false;
1178         /*
1179          * Unlike task_move routines, we access mc.to, mc.from not under
1180          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1181          */
1182         spin_lock(&mc.lock);
1183         from = mc.from;
1184         to = mc.to;
1185         if (!from)
1186                 goto unlock;
1187         if (from == mem || to == mem
1188             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1189             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1190                 ret = true;
1191 unlock:
1192         spin_unlock(&mc.lock);
1193         return ret;
1194 }
1195
1196 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1197 {
1198         if (mc.moving_task && current != mc.moving_task) {
1199                 if (mem_cgroup_under_move(mem)) {
1200                         DEFINE_WAIT(wait);
1201                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1202                         /* moving charge context might have finished. */
1203                         if (mc.moving_task)
1204                                 schedule();
1205                         finish_wait(&mc.waitq, &wait);
1206                         return true;
1207                 }
1208         }
1209         return false;
1210 }
1211
1212 /**
1213  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1214  * @memcg: The memory cgroup that went over limit
1215  * @p: Task that is going to be killed
1216  *
1217  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1218  * enabled
1219  */
1220 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1221 {
1222         struct cgroup *task_cgrp;
1223         struct cgroup *mem_cgrp;
1224         /*
1225          * Need a buffer in BSS, can't rely on allocations. The code relies
1226          * on the assumption that OOM is serialized for memory controller.
1227          * If this assumption is broken, revisit this code.
1228          */
1229         static char memcg_name[PATH_MAX];
1230         int ret;
1231
1232         if (!memcg || !p)
1233                 return;
1234
1235
1236         rcu_read_lock();
1237
1238         mem_cgrp = memcg->css.cgroup;
1239         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1240
1241         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1242         if (ret < 0) {
1243                 /*
1244                  * Unfortunately, we are unable to convert to a useful name
1245                  * But we'll still print out the usage information
1246                  */
1247                 rcu_read_unlock();
1248                 goto done;
1249         }
1250         rcu_read_unlock();
1251
1252         printk(KERN_INFO "Task in %s killed", memcg_name);
1253
1254         rcu_read_lock();
1255         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1256         if (ret < 0) {
1257                 rcu_read_unlock();
1258                 goto done;
1259         }
1260         rcu_read_unlock();
1261
1262         /*
1263          * Continues from above, so we don't need an KERN_ level
1264          */
1265         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1266 done:
1267
1268         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1269                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1270                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1271                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1272         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1273                 "failcnt %llu\n",
1274                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1275                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1276                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1277 }
1278
1279 /*
1280  * This function returns the number of memcg under hierarchy tree. Returns
1281  * 1(self count) if no children.
1282  */
1283 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1284 {
1285         int num = 0;
1286         struct mem_cgroup *iter;
1287
1288         for_each_mem_cgroup_tree(iter, mem)
1289                 num++;
1290         return num;
1291 }
1292
1293 /*
1294  * Return the memory (and swap, if configured) limit for a memcg.
1295  */
1296 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1297 {
1298         u64 limit;
1299         u64 memsw;
1300
1301         limit = res_counter_read_u64(&memcg->res, RES_LIMIT) +
1302                         total_swap_pages;
1303         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1304         /*
1305          * If memsw is finite and limits the amount of swap space available
1306          * to this memcg, return that limit.
1307          */
1308         return min(limit, memsw);
1309 }
1310
1311 /*
1312  * Visit the first child (need not be the first child as per the ordering
1313  * of the cgroup list, since we track last_scanned_child) of @mem and use
1314  * that to reclaim free pages from.
1315  */
1316 static struct mem_cgroup *
1317 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1318 {
1319         struct mem_cgroup *ret = NULL;
1320         struct cgroup_subsys_state *css;
1321         int nextid, found;
1322
1323         if (!root_mem->use_hierarchy) {
1324                 css_get(&root_mem->css);
1325                 ret = root_mem;
1326         }
1327
1328         while (!ret) {
1329                 rcu_read_lock();
1330                 nextid = root_mem->last_scanned_child + 1;
1331                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1332                                    &found);
1333                 if (css && css_tryget(css))
1334                         ret = container_of(css, struct mem_cgroup, css);
1335
1336                 rcu_read_unlock();
1337                 /* Updates scanning parameter */
1338                 spin_lock(&root_mem->reclaim_param_lock);
1339                 if (!css) {
1340                         /* this means start scan from ID:1 */
1341                         root_mem->last_scanned_child = 0;
1342                 } else
1343                         root_mem->last_scanned_child = found;
1344                 spin_unlock(&root_mem->reclaim_param_lock);
1345         }
1346
1347         return ret;
1348 }
1349
1350 /*
1351  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1352  * we reclaimed from, so that we don't end up penalizing one child extensively
1353  * based on its position in the children list.
1354  *
1355  * root_mem is the original ancestor that we've been reclaim from.
1356  *
1357  * We give up and return to the caller when we visit root_mem twice.
1358  * (other groups can be removed while we're walking....)
1359  *
1360  * If shrink==true, for avoiding to free too much, this returns immedieately.
1361  */
1362 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1363                                                 struct zone *zone,
1364                                                 gfp_t gfp_mask,
1365                                                 unsigned long reclaim_options)
1366 {
1367         struct mem_cgroup *victim;
1368         int ret, total = 0;
1369         int loop = 0;
1370         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1371         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1372         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1373         unsigned long excess = mem_cgroup_get_excess(root_mem);
1374
1375         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1376         if (root_mem->memsw_is_minimum)
1377                 noswap = true;
1378
1379         while (1) {
1380                 victim = mem_cgroup_select_victim(root_mem);
1381                 if (victim == root_mem) {
1382                         loop++;
1383                         if (loop >= 1)
1384                                 drain_all_stock_async();
1385                         if (loop >= 2) {
1386                                 /*
1387                                  * If we have not been able to reclaim
1388                                  * anything, it might because there are
1389                                  * no reclaimable pages under this hierarchy
1390                                  */
1391                                 if (!check_soft || !total) {
1392                                         css_put(&victim->css);
1393                                         break;
1394                                 }
1395                                 /*
1396                                  * We want to do more targetted reclaim.
1397                                  * excess >> 2 is not to excessive so as to
1398                                  * reclaim too much, nor too less that we keep
1399                                  * coming back to reclaim from this cgroup
1400                                  */
1401                                 if (total >= (excess >> 2) ||
1402                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1403                                         css_put(&victim->css);
1404                                         break;
1405                                 }
1406                         }
1407                 }
1408                 if (!mem_cgroup_local_usage(victim)) {
1409                         /* this cgroup's local usage == 0 */
1410                         css_put(&victim->css);
1411                         continue;
1412                 }
1413                 /* we use swappiness of local cgroup */
1414                 if (check_soft)
1415                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1416                                 noswap, get_swappiness(victim), zone);
1417                 else
1418                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1419                                                 noswap, get_swappiness(victim));
1420                 css_put(&victim->css);
1421                 /*
1422                  * At shrinking usage, we can't check we should stop here or
1423                  * reclaim more. It's depends on callers. last_scanned_child
1424                  * will work enough for keeping fairness under tree.
1425                  */
1426                 if (shrink)
1427                         return ret;
1428                 total += ret;
1429                 if (check_soft) {
1430                         if (res_counter_check_under_soft_limit(&root_mem->res))
1431                                 return total;
1432                 } else if (mem_cgroup_check_under_limit(root_mem))
1433                         return 1 + total;
1434         }
1435         return total;
1436 }
1437
1438 /*
1439  * Check OOM-Killer is already running under our hierarchy.
1440  * If someone is running, return false.
1441  */
1442 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1443 {
1444         int x, lock_count = 0;
1445         struct mem_cgroup *iter;
1446
1447         for_each_mem_cgroup_tree(iter, mem) {
1448                 x = atomic_inc_return(&iter->oom_lock);
1449                 lock_count = max(x, lock_count);
1450         }
1451
1452         if (lock_count == 1)
1453                 return true;
1454         return false;
1455 }
1456
1457 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1458 {
1459         struct mem_cgroup *iter;
1460
1461         /*
1462          * When a new child is created while the hierarchy is under oom,
1463          * mem_cgroup_oom_lock() may not be called. We have to use
1464          * atomic_add_unless() here.
1465          */
1466         for_each_mem_cgroup_tree(iter, mem)
1467                 atomic_add_unless(&iter->oom_lock, -1, 0);
1468         return 0;
1469 }
1470
1471
1472 static DEFINE_MUTEX(memcg_oom_mutex);
1473 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1474
1475 struct oom_wait_info {
1476         struct mem_cgroup *mem;
1477         wait_queue_t    wait;
1478 };
1479
1480 static int memcg_oom_wake_function(wait_queue_t *wait,
1481         unsigned mode, int sync, void *arg)
1482 {
1483         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1484         struct oom_wait_info *oom_wait_info;
1485
1486         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1487
1488         if (oom_wait_info->mem == wake_mem)
1489                 goto wakeup;
1490         /* if no hierarchy, no match */
1491         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1492                 return 0;
1493         /*
1494          * Both of oom_wait_info->mem and wake_mem are stable under us.
1495          * Then we can use css_is_ancestor without taking care of RCU.
1496          */
1497         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1498             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1499                 return 0;
1500
1501 wakeup:
1502         return autoremove_wake_function(wait, mode, sync, arg);
1503 }
1504
1505 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1506 {
1507         /* for filtering, pass "mem" as argument. */
1508         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1509 }
1510
1511 static void memcg_oom_recover(struct mem_cgroup *mem)
1512 {
1513         if (mem && atomic_read(&mem->oom_lock))
1514                 memcg_wakeup_oom(mem);
1515 }
1516
1517 /*
1518  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1519  */
1520 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1521 {
1522         struct oom_wait_info owait;
1523         bool locked, need_to_kill;
1524
1525         owait.mem = mem;
1526         owait.wait.flags = 0;
1527         owait.wait.func = memcg_oom_wake_function;
1528         owait.wait.private = current;
1529         INIT_LIST_HEAD(&owait.wait.task_list);
1530         need_to_kill = true;
1531         /* At first, try to OOM lock hierarchy under mem.*/
1532         mutex_lock(&memcg_oom_mutex);
1533         locked = mem_cgroup_oom_lock(mem);
1534         /*
1535          * Even if signal_pending(), we can't quit charge() loop without
1536          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1537          * under OOM is always welcomed, use TASK_KILLABLE here.
1538          */
1539         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1540         if (!locked || mem->oom_kill_disable)
1541                 need_to_kill = false;
1542         if (locked)
1543                 mem_cgroup_oom_notify(mem);
1544         mutex_unlock(&memcg_oom_mutex);
1545
1546         if (need_to_kill) {
1547                 finish_wait(&memcg_oom_waitq, &owait.wait);
1548                 mem_cgroup_out_of_memory(mem, mask);
1549         } else {
1550                 schedule();
1551                 finish_wait(&memcg_oom_waitq, &owait.wait);
1552         }
1553         mutex_lock(&memcg_oom_mutex);
1554         mem_cgroup_oom_unlock(mem);
1555         memcg_wakeup_oom(mem);
1556         mutex_unlock(&memcg_oom_mutex);
1557
1558         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1559                 return false;
1560         /* Give chance to dying process */
1561         schedule_timeout(1);
1562         return true;
1563 }
1564
1565 /*
1566  * Currently used to update mapped file statistics, but the routine can be
1567  * generalized to update other statistics as well.
1568  *
1569  * Notes: Race condition
1570  *
1571  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1572  * it tends to be costly. But considering some conditions, we doesn't need
1573  * to do so _always_.
1574  *
1575  * Considering "charge", lock_page_cgroup() is not required because all
1576  * file-stat operations happen after a page is attached to radix-tree. There
1577  * are no race with "charge".
1578  *
1579  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1580  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1581  * if there are race with "uncharge". Statistics itself is properly handled
1582  * by flags.
1583  *
1584  * Considering "move", this is an only case we see a race. To make the race
1585  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1586  * possibility of race condition. If there is, we take a lock.
1587  */
1588 void mem_cgroup_update_file_mapped(struct page *page, int val)
1589 {
1590         struct mem_cgroup *mem;
1591         struct page_cgroup *pc = lookup_page_cgroup(page);
1592         bool need_unlock = false;
1593
1594         if (unlikely(!pc))
1595                 return;
1596
1597         rcu_read_lock();
1598         mem = pc->mem_cgroup;
1599         if (unlikely(!mem || !PageCgroupUsed(pc)))
1600                 goto out;
1601         /* pc->mem_cgroup is unstable ? */
1602         if (unlikely(mem_cgroup_stealed(mem))) {
1603                 /* take a lock against to access pc->mem_cgroup */
1604                 lock_page_cgroup(pc);
1605                 need_unlock = true;
1606                 mem = pc->mem_cgroup;
1607                 if (!mem || !PageCgroupUsed(pc))
1608                         goto out;
1609         }
1610         if (val > 0) {
1611                 this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1612                 SetPageCgroupFileMapped(pc);
1613         } else {
1614                 this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1615                 if (!page_mapped(page)) /* for race between dec->inc counter */
1616                         ClearPageCgroupFileMapped(pc);
1617         }
1618
1619 out:
1620         if (unlikely(need_unlock))
1621                 unlock_page_cgroup(pc);
1622         rcu_read_unlock();
1623         return;
1624 }
1625
1626 /*
1627  * size of first charge trial. "32" comes from vmscan.c's magic value.
1628  * TODO: maybe necessary to use big numbers in big irons.
1629  */
1630 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1631 struct memcg_stock_pcp {
1632         struct mem_cgroup *cached; /* this never be root cgroup */
1633         int charge;
1634         struct work_struct work;
1635 };
1636 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1637 static atomic_t memcg_drain_count;
1638
1639 /*
1640  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1641  * from local stock and true is returned. If the stock is 0 or charges from a
1642  * cgroup which is not current target, returns false. This stock will be
1643  * refilled.
1644  */
1645 static bool consume_stock(struct mem_cgroup *mem)
1646 {
1647         struct memcg_stock_pcp *stock;
1648         bool ret = true;
1649
1650         stock = &get_cpu_var(memcg_stock);
1651         if (mem == stock->cached && stock->charge)
1652                 stock->charge -= PAGE_SIZE;
1653         else /* need to call res_counter_charge */
1654                 ret = false;
1655         put_cpu_var(memcg_stock);
1656         return ret;
1657 }
1658
1659 /*
1660  * Returns stocks cached in percpu to res_counter and reset cached information.
1661  */
1662 static void drain_stock(struct memcg_stock_pcp *stock)
1663 {
1664         struct mem_cgroup *old = stock->cached;
1665
1666         if (stock->charge) {
1667                 res_counter_uncharge(&old->res, stock->charge);
1668                 if (do_swap_account)
1669                         res_counter_uncharge(&old->memsw, stock->charge);
1670         }
1671         stock->cached = NULL;
1672         stock->charge = 0;
1673 }
1674
1675 /*
1676  * This must be called under preempt disabled or must be called by
1677  * a thread which is pinned to local cpu.
1678  */
1679 static void drain_local_stock(struct work_struct *dummy)
1680 {
1681         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1682         drain_stock(stock);
1683 }
1684
1685 /*
1686  * Cache charges(val) which is from res_counter, to local per_cpu area.
1687  * This will be consumed by consume_stock() function, later.
1688  */
1689 static void refill_stock(struct mem_cgroup *mem, int val)
1690 {
1691         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1692
1693         if (stock->cached != mem) { /* reset if necessary */
1694                 drain_stock(stock);
1695                 stock->cached = mem;
1696         }
1697         stock->charge += val;
1698         put_cpu_var(memcg_stock);
1699 }
1700
1701 /*
1702  * Tries to drain stocked charges in other cpus. This function is asynchronous
1703  * and just put a work per cpu for draining localy on each cpu. Caller can
1704  * expects some charges will be back to res_counter later but cannot wait for
1705  * it.
1706  */
1707 static void drain_all_stock_async(void)
1708 {
1709         int cpu;
1710         /* This function is for scheduling "drain" in asynchronous way.
1711          * The result of "drain" is not directly handled by callers. Then,
1712          * if someone is calling drain, we don't have to call drain more.
1713          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1714          * there is a race. We just do loose check here.
1715          */
1716         if (atomic_read(&memcg_drain_count))
1717                 return;
1718         /* Notify other cpus that system-wide "drain" is running */
1719         atomic_inc(&memcg_drain_count);
1720         get_online_cpus();
1721         for_each_online_cpu(cpu) {
1722                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1723                 schedule_work_on(cpu, &stock->work);
1724         }
1725         put_online_cpus();
1726         atomic_dec(&memcg_drain_count);
1727         /* We don't wait for flush_work */
1728 }
1729
1730 /* This is a synchronous drain interface. */
1731 static void drain_all_stock_sync(void)
1732 {
1733         /* called when force_empty is called */
1734         atomic_inc(&memcg_drain_count);
1735         schedule_on_each_cpu(drain_local_stock);
1736         atomic_dec(&memcg_drain_count);
1737 }
1738
1739 /*
1740  * This function drains percpu counter value from DEAD cpu and
1741  * move it to local cpu. Note that this function can be preempted.
1742  */
1743 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1744 {
1745         int i;
1746
1747         spin_lock(&mem->pcp_counter_lock);
1748         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1749                 s64 x = per_cpu(mem->stat->count[i], cpu);
1750
1751                 per_cpu(mem->stat->count[i], cpu) = 0;
1752                 mem->nocpu_base.count[i] += x;
1753         }
1754         spin_unlock(&mem->pcp_counter_lock);
1755 }
1756
1757 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1758                                         unsigned long action,
1759                                         void *hcpu)
1760 {
1761         int cpu = (unsigned long)hcpu;
1762         struct memcg_stock_pcp *stock;
1763         struct mem_cgroup *iter;
1764
1765         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1766                 return NOTIFY_OK;
1767
1768         for_each_mem_cgroup_all(iter)
1769                 mem_cgroup_drain_pcp_counter(iter, cpu);
1770
1771         stock = &per_cpu(memcg_stock, cpu);
1772         drain_stock(stock);
1773         return NOTIFY_OK;
1774 }
1775
1776
1777 /* See __mem_cgroup_try_charge() for details */
1778 enum {
1779         CHARGE_OK,              /* success */
1780         CHARGE_RETRY,           /* need to retry but retry is not bad */
1781         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1782         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1783         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1784 };
1785
1786 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1787                                 int csize, bool oom_check)
1788 {
1789         struct mem_cgroup *mem_over_limit;
1790         struct res_counter *fail_res;
1791         unsigned long flags = 0;
1792         int ret;
1793
1794         ret = res_counter_charge(&mem->res, csize, &fail_res);
1795
1796         if (likely(!ret)) {
1797                 if (!do_swap_account)
1798                         return CHARGE_OK;
1799                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1800                 if (likely(!ret))
1801                         return CHARGE_OK;
1802
1803                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1804                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1805         } else
1806                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1807
1808         if (csize > PAGE_SIZE) /* change csize and retry */
1809                 return CHARGE_RETRY;
1810
1811         if (!(gfp_mask & __GFP_WAIT))
1812                 return CHARGE_WOULDBLOCK;
1813
1814         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1815                                         gfp_mask, flags);
1816         /*
1817          * try_to_free_mem_cgroup_pages() might not give us a full
1818          * picture of reclaim. Some pages are reclaimed and might be
1819          * moved to swap cache or just unmapped from the cgroup.
1820          * Check the limit again to see if the reclaim reduced the
1821          * current usage of the cgroup before giving up
1822          */
1823         if (ret || mem_cgroup_check_under_limit(mem_over_limit))
1824                 return CHARGE_RETRY;
1825
1826         /*
1827          * At task move, charge accounts can be doubly counted. So, it's
1828          * better to wait until the end of task_move if something is going on.
1829          */
1830         if (mem_cgroup_wait_acct_move(mem_over_limit))
1831                 return CHARGE_RETRY;
1832
1833         /* If we don't need to call oom-killer at el, return immediately */
1834         if (!oom_check)
1835                 return CHARGE_NOMEM;
1836         /* check OOM */
1837         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1838                 return CHARGE_OOM_DIE;
1839
1840         return CHARGE_RETRY;
1841 }
1842
1843 /*
1844  * Unlike exported interface, "oom" parameter is added. if oom==true,
1845  * oom-killer can be invoked.
1846  */
1847 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1848                 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1849 {
1850         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1851         struct mem_cgroup *mem = NULL;
1852         int ret;
1853         int csize = CHARGE_SIZE;
1854
1855         /*
1856          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1857          * in system level. So, allow to go ahead dying process in addition to
1858          * MEMDIE process.
1859          */
1860         if (unlikely(test_thread_flag(TIF_MEMDIE)
1861                      || fatal_signal_pending(current)))
1862                 goto bypass;
1863
1864         /*
1865          * We always charge the cgroup the mm_struct belongs to.
1866          * The mm_struct's mem_cgroup changes on task migration if the
1867          * thread group leader migrates. It's possible that mm is not
1868          * set, if so charge the init_mm (happens for pagecache usage).
1869          */
1870         if (!*memcg && !mm)
1871                 goto bypass;
1872 again:
1873         if (*memcg) { /* css should be a valid one */
1874                 mem = *memcg;
1875                 VM_BUG_ON(css_is_removed(&mem->css));
1876                 if (mem_cgroup_is_root(mem))
1877                         goto done;
1878                 if (consume_stock(mem))
1879                         goto done;
1880                 css_get(&mem->css);
1881         } else {
1882                 struct task_struct *p;
1883
1884                 rcu_read_lock();
1885                 p = rcu_dereference(mm->owner);
1886                 VM_BUG_ON(!p);
1887                 /*
1888                  * because we don't have task_lock(), "p" can exit while
1889                  * we're here. In that case, "mem" can point to root
1890                  * cgroup but never be NULL. (and task_struct itself is freed
1891                  * by RCU, cgroup itself is RCU safe.) Then, we have small
1892                  * risk here to get wrong cgroup. But such kind of mis-account
1893                  * by race always happens because we don't have cgroup_mutex().
1894                  * It's overkill and we allow that small race, here.
1895                  */
1896                 mem = mem_cgroup_from_task(p);
1897                 VM_BUG_ON(!mem);
1898                 if (mem_cgroup_is_root(mem)) {
1899                         rcu_read_unlock();
1900                         goto done;
1901                 }
1902                 if (consume_stock(mem)) {
1903                         /*
1904                          * It seems dagerous to access memcg without css_get().
1905                          * But considering how consume_stok works, it's not
1906                          * necessary. If consume_stock success, some charges
1907                          * from this memcg are cached on this cpu. So, we
1908                          * don't need to call css_get()/css_tryget() before
1909                          * calling consume_stock().
1910                          */
1911                         rcu_read_unlock();
1912                         goto done;
1913                 }
1914                 /* after here, we may be blocked. we need to get refcnt */
1915                 if (!css_tryget(&mem->css)) {
1916                         rcu_read_unlock();
1917                         goto again;
1918                 }
1919                 rcu_read_unlock();
1920         }
1921
1922         do {
1923                 bool oom_check;
1924
1925                 /* If killed, bypass charge */
1926                 if (fatal_signal_pending(current)) {
1927                         css_put(&mem->css);
1928                         goto bypass;
1929                 }
1930
1931                 oom_check = false;
1932                 if (oom && !nr_oom_retries) {
1933                         oom_check = true;
1934                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1935                 }
1936
1937                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1938
1939                 switch (ret) {
1940                 case CHARGE_OK:
1941                         break;
1942                 case CHARGE_RETRY: /* not in OOM situation but retry */
1943                         csize = PAGE_SIZE;
1944                         css_put(&mem->css);
1945                         mem = NULL;
1946                         goto again;
1947                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1948                         css_put(&mem->css);
1949                         goto nomem;
1950                 case CHARGE_NOMEM: /* OOM routine works */
1951                         if (!oom) {
1952                                 css_put(&mem->css);
1953                                 goto nomem;
1954                         }
1955                         /* If oom, we never return -ENOMEM */
1956                         nr_oom_retries--;
1957                         break;
1958                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
1959                         css_put(&mem->css);
1960                         goto bypass;
1961                 }
1962         } while (ret != CHARGE_OK);
1963
1964         if (csize > PAGE_SIZE)
1965                 refill_stock(mem, csize - PAGE_SIZE);
1966         css_put(&mem->css);
1967 done:
1968         *memcg = mem;
1969         return 0;
1970 nomem:
1971         *memcg = NULL;
1972         return -ENOMEM;
1973 bypass:
1974         *memcg = NULL;
1975         return 0;
1976 }
1977
1978 /*
1979  * Somemtimes we have to undo a charge we got by try_charge().
1980  * This function is for that and do uncharge, put css's refcnt.
1981  * gotten by try_charge().
1982  */
1983 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1984                                                         unsigned long count)
1985 {
1986         if (!mem_cgroup_is_root(mem)) {
1987                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1988                 if (do_swap_account)
1989                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1990         }
1991 }
1992
1993 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1994 {
1995         __mem_cgroup_cancel_charge(mem, 1);
1996 }
1997
1998 /*
1999  * A helper function to get mem_cgroup from ID. must be called under
2000  * rcu_read_lock(). The caller must check css_is_removed() or some if
2001  * it's concern. (dropping refcnt from swap can be called against removed
2002  * memcg.)
2003  */
2004 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2005 {
2006         struct cgroup_subsys_state *css;
2007
2008         /* ID 0 is unused ID */
2009         if (!id)
2010                 return NULL;
2011         css = css_lookup(&mem_cgroup_subsys, id);
2012         if (!css)
2013                 return NULL;
2014         return container_of(css, struct mem_cgroup, css);
2015 }
2016
2017 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2018 {
2019         struct mem_cgroup *mem = NULL;
2020         struct page_cgroup *pc;
2021         unsigned short id;
2022         swp_entry_t ent;
2023
2024         VM_BUG_ON(!PageLocked(page));
2025
2026         pc = lookup_page_cgroup(page);
2027         lock_page_cgroup(pc);
2028         if (PageCgroupUsed(pc)) {
2029                 mem = pc->mem_cgroup;
2030                 if (mem && !css_tryget(&mem->css))
2031                         mem = NULL;
2032         } else if (PageSwapCache(page)) {
2033                 ent.val = page_private(page);
2034                 id = lookup_swap_cgroup(ent);
2035                 rcu_read_lock();
2036                 mem = mem_cgroup_lookup(id);
2037                 if (mem && !css_tryget(&mem->css))
2038                         mem = NULL;
2039                 rcu_read_unlock();
2040         }
2041         unlock_page_cgroup(pc);
2042         return mem;
2043 }
2044
2045 /*
2046  * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
2047  * USED state. If already USED, uncharge and return.
2048  */
2049
2050 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2051                                      struct page_cgroup *pc,
2052                                      enum charge_type ctype)
2053 {
2054         /* try_charge() can return NULL to *memcg, taking care of it. */
2055         if (!mem)
2056                 return;
2057
2058         lock_page_cgroup(pc);
2059         if (unlikely(PageCgroupUsed(pc))) {
2060                 unlock_page_cgroup(pc);
2061                 mem_cgroup_cancel_charge(mem);
2062                 return;
2063         }
2064
2065         pc->mem_cgroup = mem;
2066         /*
2067          * We access a page_cgroup asynchronously without lock_page_cgroup().
2068          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2069          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2070          * before USED bit, we need memory barrier here.
2071          * See mem_cgroup_add_lru_list(), etc.
2072          */
2073         smp_wmb();
2074         switch (ctype) {
2075         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2076         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2077                 SetPageCgroupCache(pc);
2078                 SetPageCgroupUsed(pc);
2079                 break;
2080         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2081                 ClearPageCgroupCache(pc);
2082                 SetPageCgroupUsed(pc);
2083                 break;
2084         default:
2085                 break;
2086         }
2087
2088         mem_cgroup_charge_statistics(mem, pc, true);
2089
2090         unlock_page_cgroup(pc);
2091         /*
2092          * "charge_statistics" updated event counter. Then, check it.
2093          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2094          * if they exceeds softlimit.
2095          */
2096         memcg_check_events(mem, pc->page);
2097 }
2098
2099 /**
2100  * __mem_cgroup_move_account - move account of the page
2101  * @pc: page_cgroup of the page.
2102  * @from: mem_cgroup which the page is moved from.
2103  * @to: mem_cgroup which the page is moved to. @from != @to.
2104  * @uncharge: whether we should call uncharge and css_put against @from.
2105  *
2106  * The caller must confirm following.
2107  * - page is not on LRU (isolate_page() is useful.)
2108  * - the pc is locked, used, and ->mem_cgroup points to @from.
2109  *
2110  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2111  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2112  * true, this function does "uncharge" from old cgroup, but it doesn't if
2113  * @uncharge is false, so a caller should do "uncharge".
2114  */
2115
2116 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2117         struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2118 {
2119         VM_BUG_ON(from == to);
2120         VM_BUG_ON(PageLRU(pc->page));
2121         VM_BUG_ON(!PageCgroupLocked(pc));
2122         VM_BUG_ON(!PageCgroupUsed(pc));
2123         VM_BUG_ON(pc->mem_cgroup != from);
2124
2125         if (PageCgroupFileMapped(pc)) {
2126                 /* Update mapped_file data for mem_cgroup */
2127                 preempt_disable();
2128                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2129                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2130                 preempt_enable();
2131         }
2132         mem_cgroup_charge_statistics(from, pc, false);
2133         if (uncharge)
2134                 /* This is not "cancel", but cancel_charge does all we need. */
2135                 mem_cgroup_cancel_charge(from);
2136
2137         /* caller should have done css_get */
2138         pc->mem_cgroup = to;
2139         mem_cgroup_charge_statistics(to, pc, true);
2140         /*
2141          * We charges against "to" which may not have any tasks. Then, "to"
2142          * can be under rmdir(). But in current implementation, caller of
2143          * this function is just force_empty() and move charge, so it's
2144          * garanteed that "to" is never removed. So, we don't check rmdir
2145          * status here.
2146          */
2147 }
2148
2149 /*
2150  * check whether the @pc is valid for moving account and call
2151  * __mem_cgroup_move_account()
2152  */
2153 static int mem_cgroup_move_account(struct page_cgroup *pc,
2154                 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
2155 {
2156         int ret = -EINVAL;
2157         lock_page_cgroup(pc);
2158         if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2159                 __mem_cgroup_move_account(pc, from, to, uncharge);
2160                 ret = 0;
2161         }
2162         unlock_page_cgroup(pc);
2163         /*
2164          * check events
2165          */
2166         memcg_check_events(to, pc->page);
2167         memcg_check_events(from, pc->page);
2168         return ret;
2169 }
2170
2171 /*
2172  * move charges to its parent.
2173  */
2174
2175 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2176                                   struct mem_cgroup *child,
2177                                   gfp_t gfp_mask)
2178 {
2179         struct page *page = pc->page;
2180         struct cgroup *cg = child->css.cgroup;
2181         struct cgroup *pcg = cg->parent;
2182         struct mem_cgroup *parent;
2183         int ret;
2184
2185         /* Is ROOT ? */
2186         if (!pcg)
2187                 return -EINVAL;
2188
2189         ret = -EBUSY;
2190         if (!get_page_unless_zero(page))
2191                 goto out;
2192         if (isolate_lru_page(page))
2193                 goto put;
2194
2195         parent = mem_cgroup_from_cont(pcg);
2196         ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
2197         if (ret || !parent)
2198                 goto put_back;
2199
2200         ret = mem_cgroup_move_account(pc, child, parent, true);
2201         if (ret)
2202                 mem_cgroup_cancel_charge(parent);
2203 put_back:
2204         putback_lru_page(page);
2205 put:
2206         put_page(page);
2207 out:
2208         return ret;
2209 }
2210
2211 /*
2212  * Charge the memory controller for page usage.
2213  * Return
2214  * 0 if the charge was successful
2215  * < 0 if the cgroup is over its limit
2216  */
2217 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2218                                 gfp_t gfp_mask, enum charge_type ctype)
2219 {
2220         struct mem_cgroup *mem = NULL;
2221         struct page_cgroup *pc;
2222         int ret;
2223
2224         pc = lookup_page_cgroup(page);
2225         /* can happen at boot */
2226         if (unlikely(!pc))
2227                 return 0;
2228         prefetchw(pc);
2229
2230         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
2231         if (ret || !mem)
2232                 return ret;
2233
2234         __mem_cgroup_commit_charge(mem, pc, ctype);
2235         return 0;
2236 }
2237
2238 int mem_cgroup_newpage_charge(struct page *page,
2239                               struct mm_struct *mm, gfp_t gfp_mask)
2240 {
2241         if (mem_cgroup_disabled())
2242                 return 0;
2243         if (PageCompound(page))
2244                 return 0;
2245         /*
2246          * If already mapped, we don't have to account.
2247          * If page cache, page->mapping has address_space.
2248          * But page->mapping may have out-of-use anon_vma pointer,
2249          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2250          * is NULL.
2251          */
2252         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2253                 return 0;
2254         if (unlikely(!mm))
2255                 mm = &init_mm;
2256         return mem_cgroup_charge_common(page, mm, gfp_mask,
2257                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2258 }
2259
2260 static void
2261 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2262                                         enum charge_type ctype);
2263
2264 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2265                                 gfp_t gfp_mask)
2266 {
2267         int ret;
2268
2269         if (mem_cgroup_disabled())
2270                 return 0;
2271         if (PageCompound(page))
2272                 return 0;
2273         /*
2274          * Corner case handling. This is called from add_to_page_cache()
2275          * in usual. But some FS (shmem) precharges this page before calling it
2276          * and call add_to_page_cache() with GFP_NOWAIT.
2277          *
2278          * For GFP_NOWAIT case, the page may be pre-charged before calling
2279          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2280          * charge twice. (It works but has to pay a bit larger cost.)
2281          * And when the page is SwapCache, it should take swap information
2282          * into account. This is under lock_page() now.
2283          */
2284         if (!(gfp_mask & __GFP_WAIT)) {
2285                 struct page_cgroup *pc;
2286
2287                 pc = lookup_page_cgroup(page);
2288                 if (!pc)
2289                         return 0;
2290                 lock_page_cgroup(pc);
2291                 if (PageCgroupUsed(pc)) {
2292                         unlock_page_cgroup(pc);
2293                         return 0;
2294                 }
2295                 unlock_page_cgroup(pc);
2296         }
2297
2298         if (unlikely(!mm))
2299                 mm = &init_mm;
2300
2301         if (page_is_file_cache(page))
2302                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2303                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2304
2305         /* shmem */
2306         if (PageSwapCache(page)) {
2307                 struct mem_cgroup *mem = NULL;
2308
2309                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2310                 if (!ret)
2311                         __mem_cgroup_commit_charge_swapin(page, mem,
2312                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2313         } else
2314                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2315                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2316
2317         return ret;
2318 }
2319
2320 /*
2321  * While swap-in, try_charge -> commit or cancel, the page is locked.
2322  * And when try_charge() successfully returns, one refcnt to memcg without
2323  * struct page_cgroup is acquired. This refcnt will be consumed by
2324  * "commit()" or removed by "cancel()"
2325  */
2326 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2327                                  struct page *page,
2328                                  gfp_t mask, struct mem_cgroup **ptr)
2329 {
2330         struct mem_cgroup *mem;
2331         int ret;
2332
2333         if (mem_cgroup_disabled())
2334                 return 0;
2335
2336         if (!do_swap_account)
2337                 goto charge_cur_mm;
2338         /*
2339          * A racing thread's fault, or swapoff, may have already updated
2340          * the pte, and even removed page from swap cache: in those cases
2341          * do_swap_page()'s pte_same() test will fail; but there's also a
2342          * KSM case which does need to charge the page.
2343          */
2344         if (!PageSwapCache(page))
2345                 goto charge_cur_mm;
2346         mem = try_get_mem_cgroup_from_page(page);
2347         if (!mem)
2348                 goto charge_cur_mm;
2349         *ptr = mem;
2350         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2351         css_put(&mem->css);
2352         return ret;
2353 charge_cur_mm:
2354         if (unlikely(!mm))
2355                 mm = &init_mm;
2356         return __mem_cgroup_try_charge(mm, mask, ptr, true);
2357 }
2358
2359 static void
2360 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2361                                         enum charge_type ctype)
2362 {
2363         struct page_cgroup *pc;
2364
2365         if (mem_cgroup_disabled())
2366                 return;
2367         if (!ptr)
2368                 return;
2369         cgroup_exclude_rmdir(&ptr->css);
2370         pc = lookup_page_cgroup(page);
2371         mem_cgroup_lru_del_before_commit_swapcache(page);
2372         __mem_cgroup_commit_charge(ptr, pc, ctype);
2373         mem_cgroup_lru_add_after_commit_swapcache(page);
2374         /*
2375          * Now swap is on-memory. This means this page may be
2376          * counted both as mem and swap....double count.
2377          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2378          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2379          * may call delete_from_swap_cache() before reach here.
2380          */
2381         if (do_swap_account && PageSwapCache(page)) {
2382                 swp_entry_t ent = {.val = page_private(page)};
2383                 unsigned short id;
2384                 struct mem_cgroup *memcg;
2385
2386                 id = swap_cgroup_record(ent, 0);
2387                 rcu_read_lock();
2388                 memcg = mem_cgroup_lookup(id);
2389                 if (memcg) {
2390                         /*
2391                          * This recorded memcg can be obsolete one. So, avoid
2392                          * calling css_tryget
2393                          */
2394                         if (!mem_cgroup_is_root(memcg))
2395                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2396                         mem_cgroup_swap_statistics(memcg, false);
2397                         mem_cgroup_put(memcg);
2398                 }
2399                 rcu_read_unlock();
2400         }
2401         /*
2402          * At swapin, we may charge account against cgroup which has no tasks.
2403          * So, rmdir()->pre_destroy() can be called while we do this charge.
2404          * In that case, we need to call pre_destroy() again. check it here.
2405          */
2406         cgroup_release_and_wakeup_rmdir(&ptr->css);
2407 }
2408
2409 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2410 {
2411         __mem_cgroup_commit_charge_swapin(page, ptr,
2412                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2413 }
2414
2415 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2416 {
2417         if (mem_cgroup_disabled())
2418                 return;
2419         if (!mem)
2420                 return;
2421         mem_cgroup_cancel_charge(mem);
2422 }
2423
2424 static void
2425 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2426 {
2427         struct memcg_batch_info *batch = NULL;
2428         bool uncharge_memsw = true;
2429         /* If swapout, usage of swap doesn't decrease */
2430         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2431                 uncharge_memsw = false;
2432
2433         batch = &current->memcg_batch;
2434         /*
2435          * In usual, we do css_get() when we remember memcg pointer.
2436          * But in this case, we keep res->usage until end of a series of
2437          * uncharges. Then, it's ok to ignore memcg's refcnt.
2438          */
2439         if (!batch->memcg)
2440                 batch->memcg = mem;
2441         /*
2442          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2443          * In those cases, all pages freed continously can be expected to be in
2444          * the same cgroup and we have chance to coalesce uncharges.
2445          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2446          * because we want to do uncharge as soon as possible.
2447          */
2448
2449         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2450                 goto direct_uncharge;
2451
2452         /*
2453          * In typical case, batch->memcg == mem. This means we can
2454          * merge a series of uncharges to an uncharge of res_counter.
2455          * If not, we uncharge res_counter ony by one.
2456          */
2457         if (batch->memcg != mem)
2458                 goto direct_uncharge;
2459         /* remember freed charge and uncharge it later */
2460         batch->bytes += PAGE_SIZE;
2461         if (uncharge_memsw)
2462                 batch->memsw_bytes += PAGE_SIZE;
2463         return;
2464 direct_uncharge:
2465         res_counter_uncharge(&mem->res, PAGE_SIZE);
2466         if (uncharge_memsw)
2467                 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2468         if (unlikely(batch->memcg != mem))
2469                 memcg_oom_recover(mem);
2470         return;
2471 }
2472
2473 /*
2474  * uncharge if !page_mapped(page)
2475  */
2476 static struct mem_cgroup *
2477 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2478 {
2479         struct page_cgroup *pc;
2480         struct mem_cgroup *mem = NULL;
2481
2482         if (mem_cgroup_disabled())
2483                 return NULL;
2484
2485         if (PageSwapCache(page))
2486                 return NULL;
2487
2488         /*
2489          * Check if our page_cgroup is valid
2490          */
2491         pc = lookup_page_cgroup(page);
2492         if (unlikely(!pc || !PageCgroupUsed(pc)))
2493                 return NULL;
2494
2495         lock_page_cgroup(pc);
2496
2497         mem = pc->mem_cgroup;
2498
2499         if (!PageCgroupUsed(pc))
2500                 goto unlock_out;
2501
2502         switch (ctype) {
2503         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2504         case MEM_CGROUP_CHARGE_TYPE_DROP:
2505                 /* See mem_cgroup_prepare_migration() */
2506                 if (page_mapped(page) || PageCgroupMigration(pc))
2507                         goto unlock_out;
2508                 break;
2509         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2510                 if (!PageAnon(page)) {  /* Shared memory */
2511                         if (page->mapping && !page_is_file_cache(page))
2512                                 goto unlock_out;
2513                 } else if (page_mapped(page)) /* Anon */
2514                                 goto unlock_out;
2515                 break;
2516         default:
2517                 break;
2518         }
2519
2520         mem_cgroup_charge_statistics(mem, pc, false);
2521
2522         ClearPageCgroupUsed(pc);
2523         /*
2524          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2525          * freed from LRU. This is safe because uncharged page is expected not
2526          * to be reused (freed soon). Exception is SwapCache, it's handled by
2527          * special functions.
2528          */
2529
2530         unlock_page_cgroup(pc);
2531         /*
2532          * even after unlock, we have mem->res.usage here and this memcg
2533          * will never be freed.
2534          */
2535         memcg_check_events(mem, page);
2536         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2537                 mem_cgroup_swap_statistics(mem, true);
2538                 mem_cgroup_get(mem);
2539         }
2540         if (!mem_cgroup_is_root(mem))
2541                 __do_uncharge(mem, ctype);
2542
2543         return mem;
2544
2545 unlock_out:
2546         unlock_page_cgroup(pc);
2547         return NULL;
2548 }
2549
2550 void mem_cgroup_uncharge_page(struct page *page)
2551 {
2552         /* early check. */
2553         if (page_mapped(page))
2554                 return;
2555         if (page->mapping && !PageAnon(page))
2556                 return;
2557         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2558 }
2559
2560 void mem_cgroup_uncharge_cache_page(struct page *page)
2561 {
2562         VM_BUG_ON(page_mapped(page));
2563         VM_BUG_ON(page->mapping);
2564         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2565 }
2566
2567 /*
2568  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2569  * In that cases, pages are freed continuously and we can expect pages
2570  * are in the same memcg. All these calls itself limits the number of
2571  * pages freed at once, then uncharge_start/end() is called properly.
2572  * This may be called prural(2) times in a context,
2573  */
2574
2575 void mem_cgroup_uncharge_start(void)
2576 {
2577         current->memcg_batch.do_batch++;
2578         /* We can do nest. */
2579         if (current->memcg_batch.do_batch == 1) {
2580                 current->memcg_batch.memcg = NULL;
2581                 current->memcg_batch.bytes = 0;
2582                 current->memcg_batch.memsw_bytes = 0;
2583         }
2584 }
2585
2586 void mem_cgroup_uncharge_end(void)
2587 {
2588         struct memcg_batch_info *batch = &current->memcg_batch;
2589
2590         if (!batch->do_batch)
2591                 return;
2592
2593         batch->do_batch--;
2594         if (batch->do_batch) /* If stacked, do nothing. */
2595                 return;
2596
2597         if (!batch->memcg)
2598                 return;
2599         /*
2600          * This "batch->memcg" is valid without any css_get/put etc...
2601          * bacause we hide charges behind us.
2602          */
2603         if (batch->bytes)
2604                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2605         if (batch->memsw_bytes)
2606                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2607         memcg_oom_recover(batch->memcg);
2608         /* forget this pointer (for sanity check) */
2609         batch->memcg = NULL;
2610 }
2611
2612 #ifdef CONFIG_SWAP
2613 /*
2614  * called after __delete_from_swap_cache() and drop "page" account.
2615  * memcg information is recorded to swap_cgroup of "ent"
2616  */
2617 void
2618 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2619 {
2620         struct mem_cgroup *memcg;
2621         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2622
2623         if (!swapout) /* this was a swap cache but the swap is unused ! */
2624                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2625
2626         memcg = __mem_cgroup_uncharge_common(page, ctype);
2627
2628         /*
2629          * record memcg information,  if swapout && memcg != NULL,
2630          * mem_cgroup_get() was called in uncharge().
2631          */
2632         if (do_swap_account && swapout && memcg)
2633                 swap_cgroup_record(ent, css_id(&memcg->css));
2634 }
2635 #endif
2636
2637 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2638 /*
2639  * called from swap_entry_free(). remove record in swap_cgroup and
2640  * uncharge "memsw" account.
2641  */
2642 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2643 {
2644         struct mem_cgroup *memcg;
2645         unsigned short id;
2646
2647         if (!do_swap_account)
2648                 return;
2649
2650         id = swap_cgroup_record(ent, 0);
2651         rcu_read_lock();
2652         memcg = mem_cgroup_lookup(id);
2653         if (memcg) {
2654                 /*
2655                  * We uncharge this because swap is freed.
2656                  * This memcg can be obsolete one. We avoid calling css_tryget
2657                  */
2658                 if (!mem_cgroup_is_root(memcg))
2659                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2660                 mem_cgroup_swap_statistics(memcg, false);
2661                 mem_cgroup_put(memcg);
2662         }
2663         rcu_read_unlock();
2664 }
2665
2666 /**
2667  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2668  * @entry: swap entry to be moved
2669  * @from:  mem_cgroup which the entry is moved from
2670  * @to:  mem_cgroup which the entry is moved to
2671  * @need_fixup: whether we should fixup res_counters and refcounts.
2672  *
2673  * It succeeds only when the swap_cgroup's record for this entry is the same
2674  * as the mem_cgroup's id of @from.
2675  *
2676  * Returns 0 on success, -EINVAL on failure.
2677  *
2678  * The caller must have charged to @to, IOW, called res_counter_charge() about
2679  * both res and memsw, and called css_get().
2680  */
2681 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2682                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2683 {
2684         unsigned short old_id, new_id;
2685
2686         old_id = css_id(&from->css);
2687         new_id = css_id(&to->css);
2688
2689         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2690                 mem_cgroup_swap_statistics(from, false);
2691                 mem_cgroup_swap_statistics(to, true);
2692                 /*
2693                  * This function is only called from task migration context now.
2694                  * It postpones res_counter and refcount handling till the end
2695                  * of task migration(mem_cgroup_clear_mc()) for performance
2696                  * improvement. But we cannot postpone mem_cgroup_get(to)
2697                  * because if the process that has been moved to @to does
2698                  * swap-in, the refcount of @to might be decreased to 0.
2699                  */
2700                 mem_cgroup_get(to);
2701                 if (need_fixup) {
2702                         if (!mem_cgroup_is_root(from))
2703                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2704                         mem_cgroup_put(from);
2705                         /*
2706                          * we charged both to->res and to->memsw, so we should
2707                          * uncharge to->res.
2708                          */
2709                         if (!mem_cgroup_is_root(to))
2710                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2711                 }
2712                 return 0;
2713         }
2714         return -EINVAL;
2715 }
2716 #else
2717 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2718                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2719 {
2720         return -EINVAL;
2721 }
2722 #endif
2723
2724 /*
2725  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2726  * page belongs to.
2727  */
2728 int mem_cgroup_prepare_migration(struct page *page,
2729         struct page *newpage, struct mem_cgroup **ptr)
2730 {
2731         struct page_cgroup *pc;
2732         struct mem_cgroup *mem = NULL;
2733         enum charge_type ctype;
2734         int ret = 0;
2735
2736         if (mem_cgroup_disabled())
2737                 return 0;
2738
2739         pc = lookup_page_cgroup(page);
2740         lock_page_cgroup(pc);
2741         if (PageCgroupUsed(pc)) {
2742                 mem = pc->mem_cgroup;
2743                 css_get(&mem->css);
2744                 /*
2745                  * At migrating an anonymous page, its mapcount goes down
2746                  * to 0 and uncharge() will be called. But, even if it's fully
2747                  * unmapped, migration may fail and this page has to be
2748                  * charged again. We set MIGRATION flag here and delay uncharge
2749                  * until end_migration() is called
2750                  *
2751                  * Corner Case Thinking
2752                  * A)
2753                  * When the old page was mapped as Anon and it's unmap-and-freed
2754                  * while migration was ongoing.
2755                  * If unmap finds the old page, uncharge() of it will be delayed
2756                  * until end_migration(). If unmap finds a new page, it's
2757                  * uncharged when it make mapcount to be 1->0. If unmap code
2758                  * finds swap_migration_entry, the new page will not be mapped
2759                  * and end_migration() will find it(mapcount==0).
2760                  *
2761                  * B)
2762                  * When the old page was mapped but migraion fails, the kernel
2763                  * remaps it. A charge for it is kept by MIGRATION flag even
2764                  * if mapcount goes down to 0. We can do remap successfully
2765                  * without charging it again.
2766                  *
2767                  * C)
2768                  * The "old" page is under lock_page() until the end of
2769                  * migration, so, the old page itself will not be swapped-out.
2770                  * If the new page is swapped out before end_migraton, our
2771                  * hook to usual swap-out path will catch the event.
2772                  */
2773                 if (PageAnon(page))
2774                         SetPageCgroupMigration(pc);
2775         }
2776         unlock_page_cgroup(pc);
2777         /*
2778          * If the page is not charged at this point,
2779          * we return here.
2780          */
2781         if (!mem)
2782                 return 0;
2783
2784         *ptr = mem;
2785         ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2786         css_put(&mem->css);/* drop extra refcnt */
2787         if (ret || *ptr == NULL) {
2788                 if (PageAnon(page)) {
2789                         lock_page_cgroup(pc);
2790                         ClearPageCgroupMigration(pc);
2791                         unlock_page_cgroup(pc);
2792                         /*
2793                          * The old page may be fully unmapped while we kept it.
2794                          */
2795                         mem_cgroup_uncharge_page(page);
2796                 }
2797                 return -ENOMEM;
2798         }
2799         /*
2800          * We charge new page before it's used/mapped. So, even if unlock_page()
2801          * is called before end_migration, we can catch all events on this new
2802          * page. In the case new page is migrated but not remapped, new page's
2803          * mapcount will be finally 0 and we call uncharge in end_migration().
2804          */
2805         pc = lookup_page_cgroup(newpage);
2806         if (PageAnon(page))
2807                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2808         else if (page_is_file_cache(page))
2809                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2810         else
2811                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2812         __mem_cgroup_commit_charge(mem, pc, ctype);
2813         return ret;
2814 }
2815
2816 /* remove redundant charge if migration failed*/
2817 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2818         struct page *oldpage, struct page *newpage)
2819 {
2820         struct page *used, *unused;
2821         struct page_cgroup *pc;
2822
2823         if (!mem)
2824                 return;
2825         /* blocks rmdir() */
2826         cgroup_exclude_rmdir(&mem->css);
2827         /* at migration success, oldpage->mapping is NULL. */
2828         if (oldpage->mapping) {
2829                 used = oldpage;
2830                 unused = newpage;
2831         } else {
2832                 used = newpage;
2833                 unused = oldpage;
2834         }
2835         /*
2836          * We disallowed uncharge of pages under migration because mapcount
2837          * of the page goes down to zero, temporarly.
2838          * Clear the flag and check the page should be charged.
2839          */
2840         pc = lookup_page_cgroup(oldpage);
2841         lock_page_cgroup(pc);
2842         ClearPageCgroupMigration(pc);
2843         unlock_page_cgroup(pc);
2844
2845         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2846
2847         /*
2848          * If a page is a file cache, radix-tree replacement is very atomic
2849          * and we can skip this check. When it was an Anon page, its mapcount
2850          * goes down to 0. But because we added MIGRATION flage, it's not
2851          * uncharged yet. There are several case but page->mapcount check
2852          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2853          * check. (see prepare_charge() also)
2854          */
2855         if (PageAnon(used))
2856                 mem_cgroup_uncharge_page(used);
2857         /*
2858          * At migration, we may charge account against cgroup which has no
2859          * tasks.
2860          * So, rmdir()->pre_destroy() can be called while we do this charge.
2861          * In that case, we need to call pre_destroy() again. check it here.
2862          */
2863         cgroup_release_and_wakeup_rmdir(&mem->css);
2864 }
2865
2866 /*
2867  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2868  * Calling hierarchical_reclaim is not enough because we should update
2869  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2870  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2871  * not from the memcg which this page would be charged to.
2872  * try_charge_swapin does all of these works properly.
2873  */
2874 int mem_cgroup_shmem_charge_fallback(struct page *page,
2875                             struct mm_struct *mm,
2876                             gfp_t gfp_mask)
2877 {
2878         struct mem_cgroup *mem = NULL;
2879         int ret;
2880
2881         if (mem_cgroup_disabled())
2882                 return 0;
2883
2884         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2885         if (!ret)
2886                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2887
2888         return ret;
2889 }
2890
2891 static DEFINE_MUTEX(set_limit_mutex);
2892
2893 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2894                                 unsigned long long val)
2895 {
2896         int retry_count;
2897         u64 memswlimit, memlimit;
2898         int ret = 0;
2899         int children = mem_cgroup_count_children(memcg);
2900         u64 curusage, oldusage;
2901         int enlarge;
2902
2903         /*
2904          * For keeping hierarchical_reclaim simple, how long we should retry
2905          * is depends on callers. We set our retry-count to be function
2906          * of # of children which we should visit in this loop.
2907          */
2908         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2909
2910         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2911
2912         enlarge = 0;
2913         while (retry_count) {
2914                 if (signal_pending(current)) {
2915                         ret = -EINTR;
2916                         break;
2917                 }
2918                 /*
2919                  * Rather than hide all in some function, I do this in
2920                  * open coded manner. You see what this really does.
2921                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2922                  */
2923                 mutex_lock(&set_limit_mutex);
2924                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2925                 if (memswlimit < val) {
2926                         ret = -EINVAL;
2927                         mutex_unlock(&set_limit_mutex);
2928                         break;
2929                 }
2930
2931                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2932                 if (memlimit < val)
2933                         enlarge = 1;
2934
2935                 ret = res_counter_set_limit(&memcg->res, val);
2936                 if (!ret) {
2937                         if (memswlimit == val)
2938                                 memcg->memsw_is_minimum = true;
2939                         else
2940                                 memcg->memsw_is_minimum = false;
2941                 }
2942                 mutex_unlock(&set_limit_mutex);
2943
2944                 if (!ret)
2945                         break;
2946
2947                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2948                                                 MEM_CGROUP_RECLAIM_SHRINK);
2949                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2950                 /* Usage is reduced ? */
2951                 if (curusage >= oldusage)
2952                         retry_count--;
2953                 else
2954                         oldusage = curusage;
2955         }
2956         if (!ret && enlarge)
2957                 memcg_oom_recover(memcg);
2958
2959         return ret;
2960 }
2961
2962 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2963                                         unsigned long long val)
2964 {
2965         int retry_count;
2966         u64 memlimit, memswlimit, oldusage, curusage;
2967         int children = mem_cgroup_count_children(memcg);
2968         int ret = -EBUSY;
2969         int enlarge = 0;
2970
2971         /* see mem_cgroup_resize_res_limit */
2972         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2973         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2974         while (retry_count) {
2975                 if (signal_pending(current)) {
2976                         ret = -EINTR;
2977                         break;
2978                 }
2979                 /*
2980                  * Rather than hide all in some function, I do this in
2981                  * open coded manner. You see what this really does.
2982                  * We have to guarantee mem->res.limit < mem->memsw.limit.
2983                  */
2984                 mutex_lock(&set_limit_mutex);
2985                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2986                 if (memlimit > val) {
2987                         ret = -EINVAL;
2988                         mutex_unlock(&set_limit_mutex);
2989                         break;
2990                 }
2991                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2992                 if (memswlimit < val)
2993                         enlarge = 1;
2994                 ret = res_counter_set_limit(&memcg->memsw, val);
2995                 if (!ret) {
2996                         if (memlimit == val)
2997                                 memcg->memsw_is_minimum = true;
2998                         else
2999                                 memcg->memsw_is_minimum = false;
3000                 }
3001                 mutex_unlock(&set_limit_mutex);
3002
3003                 if (!ret)
3004                         break;
3005
3006                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3007                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3008                                                 MEM_CGROUP_RECLAIM_SHRINK);
3009                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3010                 /* Usage is reduced ? */
3011                 if (curusage >= oldusage)
3012                         retry_count--;
3013                 else
3014                         oldusage = curusage;
3015         }
3016         if (!ret && enlarge)
3017                 memcg_oom_recover(memcg);
3018         return ret;
3019 }
3020
3021 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3022                                             gfp_t gfp_mask)
3023 {
3024         unsigned long nr_reclaimed = 0;
3025         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3026         unsigned long reclaimed;
3027         int loop = 0;
3028         struct mem_cgroup_tree_per_zone *mctz;
3029         unsigned long long excess;
3030
3031         if (order > 0)
3032                 return 0;
3033
3034         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3035         /*
3036          * This loop can run a while, specially if mem_cgroup's continuously
3037          * keep exceeding their soft limit and putting the system under
3038          * pressure
3039          */
3040         do {
3041                 if (next_mz)
3042                         mz = next_mz;
3043                 else
3044                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3045                 if (!mz)
3046                         break;
3047
3048                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3049                                                 gfp_mask,
3050                                                 MEM_CGROUP_RECLAIM_SOFT);
3051                 nr_reclaimed += reclaimed;
3052                 spin_lock(&mctz->lock);
3053
3054                 /*
3055                  * If we failed to reclaim anything from this memory cgroup
3056                  * it is time to move on to the next cgroup
3057                  */
3058                 next_mz = NULL;
3059                 if (!reclaimed) {
3060                         do {
3061                                 /*
3062                                  * Loop until we find yet another one.
3063                                  *
3064                                  * By the time we get the soft_limit lock
3065                                  * again, someone might have aded the
3066                                  * group back on the RB tree. Iterate to
3067                                  * make sure we get a different mem.
3068                                  * mem_cgroup_largest_soft_limit_node returns
3069                                  * NULL if no other cgroup is present on
3070                                  * the tree
3071                                  */
3072                                 next_mz =
3073                                 __mem_cgroup_largest_soft_limit_node(mctz);
3074                                 if (next_mz == mz) {
3075                                         css_put(&next_mz->mem->css);
3076                                         next_mz = NULL;
3077                                 } else /* next_mz == NULL or other memcg */
3078                                         break;
3079                         } while (1);
3080                 }
3081                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3082                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3083                 /*
3084                  * One school of thought says that we should not add
3085                  * back the node to the tree if reclaim returns 0.
3086                  * But our reclaim could return 0, simply because due
3087                  * to priority we are exposing a smaller subset of
3088                  * memory to reclaim from. Consider this as a longer
3089                  * term TODO.
3090                  */
3091                 /* If excess == 0, no tree ops */
3092                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3093                 spin_unlock(&mctz->lock);
3094                 css_put(&mz->mem->css);
3095                 loop++;
3096                 /*
3097                  * Could not reclaim anything and there are no more
3098                  * mem cgroups to try or we seem to be looping without
3099                  * reclaiming anything.
3100                  */
3101                 if (!nr_reclaimed &&
3102                         (next_mz == NULL ||
3103                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3104                         break;
3105         } while (!nr_reclaimed);
3106         if (next_mz)
3107                 css_put(&next_mz->mem->css);
3108         return nr_reclaimed;
3109 }
3110
3111 /*
3112  * This routine traverse page_cgroup in given list and drop them all.
3113  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3114  */
3115 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3116                                 int node, int zid, enum lru_list lru)
3117 {
3118         struct zone *zone;
3119         struct mem_cgroup_per_zone *mz;
3120         struct page_cgroup *pc, *busy;
3121         unsigned long flags, loop;
3122         struct list_head *list;
3123         int ret = 0;
3124
3125         zone = &NODE_DATA(node)->node_zones[zid];
3126         mz = mem_cgroup_zoneinfo(mem, node, zid);
3127         list = &mz->lists[lru];
3128
3129         loop = MEM_CGROUP_ZSTAT(mz, lru);
3130         /* give some margin against EBUSY etc...*/
3131         loop += 256;
3132         busy = NULL;
3133         while (loop--) {
3134                 ret = 0;
3135                 spin_lock_irqsave(&zone->lru_lock, flags);
3136                 if (list_empty(list)) {
3137                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3138                         break;
3139                 }
3140                 pc = list_entry(list->prev, struct page_cgroup, lru);
3141                 if (busy == pc) {
3142                         list_move(&pc->lru, list);
3143                         busy = NULL;
3144                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3145                         continue;
3146                 }
3147                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3148
3149                 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3150                 if (ret == -ENOMEM)
3151                         break;
3152
3153                 if (ret == -EBUSY || ret == -EINVAL) {
3154                         /* found lock contention or "pc" is obsolete. */
3155                         busy = pc;
3156                         cond_resched();
3157                 } else
3158                         busy = NULL;
3159         }
3160
3161         if (!ret && !list_empty(list))
3162                 return -EBUSY;
3163         return ret;
3164 }
3165
3166 /*
3167  * make mem_cgroup's charge to be 0 if there is no task.
3168  * This enables deleting this mem_cgroup.
3169  */
3170 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3171 {
3172         int ret;
3173         int node, zid, shrink;
3174         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3175         struct cgroup *cgrp = mem->css.cgroup;
3176
3177         css_get(&mem->css);
3178
3179         shrink = 0;
3180         /* should free all ? */
3181         if (free_all)
3182                 goto try_to_free;
3183 move_account:
3184         do {
3185                 ret = -EBUSY;
3186                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3187                         goto out;
3188                 ret = -EINTR;
3189                 if (signal_pending(current))
3190                         goto out;
3191                 /* This is for making all *used* pages to be on LRU. */
3192                 lru_add_drain_all();
3193                 drain_all_stock_sync();
3194                 ret = 0;
3195                 mem_cgroup_start_move(mem);
3196                 for_each_node_state(node, N_HIGH_MEMORY) {
3197                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3198                                 enum lru_list l;
3199                                 for_each_lru(l) {
3200                                         ret = mem_cgroup_force_empty_list(mem,
3201                                                         node, zid, l);
3202                                         if (ret)
3203                                                 break;
3204                                 }
3205                         }
3206                         if (ret)
3207                                 break;
3208                 }
3209                 mem_cgroup_end_move(mem);
3210                 memcg_oom_recover(mem);
3211                 /* it seems parent cgroup doesn't have enough mem */
3212                 if (ret == -ENOMEM)
3213                         goto try_to_free;
3214                 cond_resched();
3215         /* "ret" should also be checked to ensure all lists are empty. */
3216         } while (mem->res.usage > 0 || ret);
3217 out:
3218         css_put(&mem->css);
3219         return ret;
3220
3221 try_to_free:
3222         /* returns EBUSY if there is a task or if we come here twice. */
3223         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3224                 ret = -EBUSY;
3225                 goto out;
3226         }
3227         /* we call try-to-free pages for make this cgroup empty */
3228         lru_add_drain_all();
3229         /* try to free all pages in this cgroup */
3230         shrink = 1;
3231         while (nr_retries && mem->res.usage > 0) {
3232                 int progress;
3233
3234                 if (signal_pending(current)) {
3235                         ret = -EINTR;
3236                         goto out;
3237                 }
3238                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3239                                                 false, get_swappiness(mem));
3240                 if (!progress) {
3241                         nr_retries--;
3242                         /* maybe some writeback is necessary */
3243                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3244                 }
3245
3246         }
3247         lru_add_drain();
3248         /* try move_account...there may be some *locked* pages. */
3249         goto move_account;
3250 }
3251
3252 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3253 {
3254         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3255 }
3256
3257
3258 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3259 {
3260         return mem_cgroup_from_cont(cont)->use_hierarchy;
3261 }
3262
3263 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3264                                         u64 val)
3265 {
3266         int retval = 0;
3267         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3268         struct cgroup *parent = cont->parent;
3269         struct mem_cgroup *parent_mem = NULL;
3270
3271         if (parent)
3272                 parent_mem = mem_cgroup_from_cont(parent);
3273
3274         cgroup_lock();
3275         /*
3276          * If parent's use_hierarchy is set, we can't make any modifications
3277          * in the child subtrees. If it is unset, then the change can
3278          * occur, provided the current cgroup has no children.
3279          *
3280          * For the root cgroup, parent_mem is NULL, we allow value to be
3281          * set if there are no children.
3282          */
3283         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3284                                 (val == 1 || val == 0)) {
3285                 if (list_empty(&cont->children))
3286                         mem->use_hierarchy = val;
3287                 else
3288                         retval = -EBUSY;
3289         } else
3290                 retval = -EINVAL;
3291         cgroup_unlock();
3292
3293         return retval;
3294 }
3295
3296
3297 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3298                                 enum mem_cgroup_stat_index idx)
3299 {
3300         struct mem_cgroup *iter;
3301         s64 val = 0;
3302
3303         /* each per cpu's value can be minus.Then, use s64 */
3304         for_each_mem_cgroup_tree(iter, mem)
3305                 val += mem_cgroup_read_stat(iter, idx);
3306
3307         if (val < 0) /* race ? */
3308                 val = 0;
3309         return val;
3310 }
3311
3312 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3313 {
3314         u64 val;
3315
3316         if (!mem_cgroup_is_root(mem)) {
3317                 if (!swap)
3318                         return res_counter_read_u64(&mem->res, RES_USAGE);
3319                 else
3320                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3321         }
3322
3323         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3324         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3325
3326         if (swap)
3327                 val += mem_cgroup_get_recursive_idx_stat(mem,
3328                                 MEM_CGROUP_STAT_SWAPOUT);
3329
3330         return val << PAGE_SHIFT;
3331 }
3332
3333 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3334 {
3335         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3336         u64 val;
3337         int type, name;
3338
3339         type = MEMFILE_TYPE(cft->private);
3340         name = MEMFILE_ATTR(cft->private);
3341         switch (type) {
3342         case _MEM:
3343                 if (name == RES_USAGE)
3344                         val = mem_cgroup_usage(mem, false);
3345                 else
3346                         val = res_counter_read_u64(&mem->res, name);
3347                 break;
3348         case _MEMSWAP:
3349                 if (name == RES_USAGE)
3350                         val = mem_cgroup_usage(mem, true);
3351                 else
3352                         val = res_counter_read_u64(&mem->memsw, name);
3353                 break;
3354         default:
3355                 BUG();
3356                 break;
3357         }
3358         return val;
3359 }
3360 /*
3361  * The user of this function is...
3362  * RES_LIMIT.
3363  */
3364 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3365                             const char *buffer)
3366 {
3367         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3368         int type, name;
3369         unsigned long long val;
3370         int ret;
3371
3372         type = MEMFILE_TYPE(cft->private);
3373         name = MEMFILE_ATTR(cft->private);
3374         switch (name) {
3375         case RES_LIMIT:
3376                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3377                         ret = -EINVAL;
3378                         break;
3379                 }
3380                 /* This function does all necessary parse...reuse it */
3381                 ret = res_counter_memparse_write_strategy(buffer, &val);
3382                 if (ret)
3383                         break;
3384                 if (type == _MEM)
3385                         ret = mem_cgroup_resize_limit(memcg, val);
3386                 else
3387                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3388                 break;
3389         case RES_SOFT_LIMIT:
3390                 ret = res_counter_memparse_write_strategy(buffer, &val);
3391                 if (ret)
3392                         break;
3393                 /*
3394                  * For memsw, soft limits are hard to implement in terms
3395                  * of semantics, for now, we support soft limits for
3396                  * control without swap
3397                  */
3398                 if (type == _MEM)
3399                         ret = res_counter_set_soft_limit(&memcg->res, val);
3400                 else
3401                         ret = -EINVAL;
3402                 break;
3403         default:
3404                 ret = -EINVAL; /* should be BUG() ? */
3405                 break;
3406         }
3407         return ret;
3408 }
3409
3410 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3411                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3412 {
3413         struct cgroup *cgroup;
3414         unsigned long long min_limit, min_memsw_limit, tmp;
3415
3416         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3417         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3418         cgroup = memcg->css.cgroup;
3419         if (!memcg->use_hierarchy)
3420                 goto out;
3421
3422         while (cgroup->parent) {
3423                 cgroup = cgroup->parent;
3424                 memcg = mem_cgroup_from_cont(cgroup);
3425                 if (!memcg->use_hierarchy)
3426                         break;
3427                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3428                 min_limit = min(min_limit, tmp);
3429                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3430                 min_memsw_limit = min(min_memsw_limit, tmp);
3431         }
3432 out:
3433         *mem_limit = min_limit;
3434         *memsw_limit = min_memsw_limit;
3435         return;
3436 }
3437
3438 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3439 {
3440         struct mem_cgroup *mem;
3441         int type, name;
3442
3443         mem = mem_cgroup_from_cont(cont);
3444         type = MEMFILE_TYPE(event);
3445         name = MEMFILE_ATTR(event);
3446         switch (name) {
3447         case RES_MAX_USAGE:
3448                 if (type == _MEM)
3449                         res_counter_reset_max(&mem->res);
3450                 else
3451                         res_counter_reset_max(&mem->memsw);
3452                 break;
3453         case RES_FAILCNT:
3454                 if (type == _MEM)
3455                         res_counter_reset_failcnt(&mem->res);
3456                 else
3457                         res_counter_reset_failcnt(&mem->memsw);
3458                 break;
3459         }
3460
3461         return 0;
3462 }
3463
3464 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3465                                         struct cftype *cft)
3466 {
3467         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3468 }
3469
3470 #ifdef CONFIG_MMU
3471 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3472                                         struct cftype *cft, u64 val)
3473 {
3474         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3475
3476         if (val >= (1 << NR_MOVE_TYPE))
3477                 return -EINVAL;
3478         /*
3479          * We check this value several times in both in can_attach() and
3480          * attach(), so we need cgroup lock to prevent this value from being
3481          * inconsistent.
3482          */
3483         cgroup_lock();
3484         mem->move_charge_at_immigrate = val;
3485         cgroup_unlock();
3486
3487         return 0;
3488 }
3489 #else
3490 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3491                                         struct cftype *cft, u64 val)
3492 {
3493         return -ENOSYS;
3494 }
3495 #endif
3496
3497
3498 /* For read statistics */
3499 enum {
3500         MCS_CACHE,
3501         MCS_RSS,
3502         MCS_FILE_MAPPED,
3503         MCS_PGPGIN,
3504         MCS_PGPGOUT,
3505         MCS_SWAP,
3506         MCS_INACTIVE_ANON,
3507         MCS_ACTIVE_ANON,
3508         MCS_INACTIVE_FILE,
3509         MCS_ACTIVE_FILE,
3510         MCS_UNEVICTABLE,
3511         NR_MCS_STAT,
3512 };
3513
3514 struct mcs_total_stat {
3515         s64 stat[NR_MCS_STAT];
3516 };
3517
3518 struct {
3519         char *local_name;
3520         char *total_name;
3521 } memcg_stat_strings[NR_MCS_STAT] = {
3522         {"cache", "total_cache"},
3523         {"rss", "total_rss"},
3524         {"mapped_file", "total_mapped_file"},
3525         {"pgpgin", "total_pgpgin"},
3526         {"pgpgout", "total_pgpgout"},
3527         {"swap", "total_swap"},
3528         {"inactive_anon", "total_inactive_anon"},
3529         {"active_anon", "total_active_anon"},
3530         {"inactive_file", "total_inactive_file"},
3531         {"active_file", "total_active_file"},
3532         {"unevictable", "total_unevictable"}
3533 };
3534
3535
3536 static void
3537 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3538 {
3539         s64 val;
3540
3541         /* per cpu stat */
3542         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3543         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3544         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3545         s->stat[MCS_RSS] += val * PAGE_SIZE;
3546         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3547         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3548         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3549         s->stat[MCS_PGPGIN] += val;
3550         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3551         s->stat[MCS_PGPGOUT] += val;
3552         if (do_swap_account) {
3553                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3554                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3555         }
3556
3557         /* per zone stat */
3558         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3559         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3560         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3561         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3562         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3563         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3564         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3565         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3566         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3567         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3568 }
3569
3570 static void
3571 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3572 {
3573         struct mem_cgroup *iter;
3574
3575         for_each_mem_cgroup_tree(iter, mem)
3576                 mem_cgroup_get_local_stat(iter, s);
3577 }
3578
3579 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3580                                  struct cgroup_map_cb *cb)
3581 {
3582         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3583         struct mcs_total_stat mystat;
3584         int i;
3585
3586         memset(&mystat, 0, sizeof(mystat));
3587         mem_cgroup_get_local_stat(mem_cont, &mystat);
3588
3589         for (i = 0; i < NR_MCS_STAT; i++) {
3590                 if (i == MCS_SWAP && !do_swap_account)
3591                         continue;
3592                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3593         }
3594
3595         /* Hierarchical information */
3596         {
3597                 unsigned long long limit, memsw_limit;
3598                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3599                 cb->fill(cb, "hierarchical_memory_limit", limit);
3600                 if (do_swap_account)
3601                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3602         }
3603
3604         memset(&mystat, 0, sizeof(mystat));
3605         mem_cgroup_get_total_stat(mem_cont, &mystat);
3606         for (i = 0; i < NR_MCS_STAT; i++) {
3607                 if (i == MCS_SWAP && !do_swap_account)
3608                         continue;
3609                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3610         }
3611
3612 #ifdef CONFIG_DEBUG_VM
3613         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3614
3615         {
3616                 int nid, zid;
3617                 struct mem_cgroup_per_zone *mz;
3618                 unsigned long recent_rotated[2] = {0, 0};
3619                 unsigned long recent_scanned[2] = {0, 0};
3620
3621                 for_each_online_node(nid)
3622                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3623                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3624
3625                                 recent_rotated[0] +=
3626                                         mz->reclaim_stat.recent_rotated[0];
3627                                 recent_rotated[1] +=
3628                                         mz->reclaim_stat.recent_rotated[1];
3629                                 recent_scanned[0] +=
3630                                         mz->reclaim_stat.recent_scanned[0];
3631                                 recent_scanned[1] +=
3632                                         mz->reclaim_stat.recent_scanned[1];
3633                         }
3634                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3635                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3636                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3637                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3638         }
3639 #endif
3640
3641         return 0;
3642 }
3643
3644 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3645 {
3646         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3647
3648         return get_swappiness(memcg);
3649 }
3650
3651 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3652                                        u64 val)
3653 {
3654         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3655         struct mem_cgroup *parent;
3656
3657         if (val > 100)
3658                 return -EINVAL;
3659
3660         if (cgrp->parent == NULL)
3661                 return -EINVAL;
3662
3663         parent = mem_cgroup_from_cont(cgrp->parent);
3664
3665         cgroup_lock();
3666
3667         /* If under hierarchy, only empty-root can set this value */
3668         if ((parent->use_hierarchy) ||
3669             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3670                 cgroup_unlock();
3671                 return -EINVAL;
3672         }
3673
3674         spin_lock(&memcg->reclaim_param_lock);
3675         memcg->swappiness = val;
3676         spin_unlock(&memcg->reclaim_param_lock);
3677
3678         cgroup_unlock();
3679
3680         return 0;
3681 }
3682
3683 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3684 {
3685         struct mem_cgroup_threshold_ary *t;
3686         u64 usage;
3687         int i;
3688
3689         rcu_read_lock();
3690         if (!swap)
3691                 t = rcu_dereference(memcg->thresholds.primary);
3692         else
3693                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3694
3695         if (!t)
3696                 goto unlock;
3697
3698         usage = mem_cgroup_usage(memcg, swap);
3699
3700         /*
3701          * current_threshold points to threshold just below usage.
3702          * If it's not true, a threshold was crossed after last
3703          * call of __mem_cgroup_threshold().
3704          */
3705         i = t->current_threshold;
3706
3707         /*
3708          * Iterate backward over array of thresholds starting from
3709          * current_threshold and check if a threshold is crossed.
3710          * If none of thresholds below usage is crossed, we read
3711          * only one element of the array here.
3712          */
3713         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3714                 eventfd_signal(t->entries[i].eventfd, 1);
3715
3716         /* i = current_threshold + 1 */
3717         i++;
3718
3719         /*
3720          * Iterate forward over array of thresholds starting from
3721          * current_threshold+1 and check if a threshold is crossed.
3722          * If none of thresholds above usage is crossed, we read
3723          * only one element of the array here.
3724          */
3725         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3726                 eventfd_signal(t->entries[i].eventfd, 1);
3727
3728         /* Update current_threshold */
3729         t->current_threshold = i - 1;
3730 unlock:
3731         rcu_read_unlock();
3732 }
3733
3734 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3735 {
3736         while (memcg) {
3737                 __mem_cgroup_threshold(memcg, false);
3738                 if (do_swap_account)
3739                         __mem_cgroup_threshold(memcg, true);
3740
3741                 memcg = parent_mem_cgroup(memcg);
3742         }
3743 }
3744
3745 static int compare_thresholds(const void *a, const void *b)
3746 {
3747         const struct mem_cgroup_threshold *_a = a;
3748         const struct mem_cgroup_threshold *_b = b;
3749
3750         return _a->threshold - _b->threshold;
3751 }
3752
3753 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3754 {
3755         struct mem_cgroup_eventfd_list *ev;
3756
3757         list_for_each_entry(ev, &mem->oom_notify, list)
3758                 eventfd_signal(ev->eventfd, 1);
3759         return 0;
3760 }
3761
3762 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3763 {
3764         struct mem_cgroup *iter;
3765
3766         for_each_mem_cgroup_tree(iter, mem)
3767                 mem_cgroup_oom_notify_cb(iter);
3768 }
3769
3770 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3771         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3772 {
3773         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3774         struct mem_cgroup_thresholds *thresholds;
3775         struct mem_cgroup_threshold_ary *new;
3776         int type = MEMFILE_TYPE(cft->private);
3777         u64 threshold, usage;
3778         int i, size, ret;
3779
3780         ret = res_counter_memparse_write_strategy(args, &threshold);
3781         if (ret)
3782                 return ret;
3783
3784         mutex_lock(&memcg->thresholds_lock);
3785
3786         if (type == _MEM)
3787                 thresholds = &memcg->thresholds;
3788         else if (type == _MEMSWAP)
3789                 thresholds = &memcg->memsw_thresholds;
3790         else
3791                 BUG();
3792
3793         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3794
3795         /* Check if a threshold crossed before adding a new one */
3796         if (thresholds->primary)
3797                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3798
3799         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3800
3801         /* Allocate memory for new array of thresholds */
3802         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3803                         GFP_KERNEL);
3804         if (!new) {
3805                 ret = -ENOMEM;
3806                 goto unlock;
3807         }
3808         new->size = size;
3809
3810         /* Copy thresholds (if any) to new array */
3811         if (thresholds->primary) {
3812                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3813                                 sizeof(struct mem_cgroup_threshold));
3814         }
3815
3816         /* Add new threshold */
3817         new->entries[size - 1].eventfd = eventfd;
3818         new->entries[size - 1].threshold = threshold;
3819
3820         /* Sort thresholds. Registering of new threshold isn't time-critical */
3821         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3822                         compare_thresholds, NULL);
3823
3824         /* Find current threshold */
3825         new->current_threshold = -1;
3826         for (i = 0; i < size; i++) {
3827                 if (new->entries[i].threshold < usage) {
3828                         /*
3829                          * new->current_threshold will not be used until
3830                          * rcu_assign_pointer(), so it's safe to increment
3831                          * it here.
3832                          */
3833                         ++new->current_threshold;
3834                 }
3835         }
3836
3837         /* Free old spare buffer and save old primary buffer as spare */
3838         kfree(thresholds->spare);
3839         thresholds->spare = thresholds->primary;
3840
3841         rcu_assign_pointer(thresholds->primary, new);
3842
3843         /* To be sure that nobody uses thresholds */
3844         synchronize_rcu();
3845
3846 unlock:
3847         mutex_unlock(&memcg->thresholds_lock);
3848
3849         return ret;
3850 }
3851
3852 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3853         struct cftype *cft, struct eventfd_ctx *eventfd)
3854 {
3855         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3856         struct mem_cgroup_thresholds *thresholds;
3857         struct mem_cgroup_threshold_ary *new;
3858         int type = MEMFILE_TYPE(cft->private);
3859         u64 usage;
3860         int i, j, size;
3861
3862         mutex_lock(&memcg->thresholds_lock);
3863         if (type == _MEM)
3864                 thresholds = &memcg->thresholds;
3865         else if (type == _MEMSWAP)
3866                 thresholds = &memcg->memsw_thresholds;
3867         else
3868                 BUG();
3869
3870         /*
3871          * Something went wrong if we trying to unregister a threshold
3872          * if we don't have thresholds
3873          */
3874         BUG_ON(!thresholds);
3875
3876         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3877
3878         /* Check if a threshold crossed before removing */
3879         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3880
3881         /* Calculate new number of threshold */
3882         size = 0;
3883         for (i = 0; i < thresholds->primary->size; i++) {
3884                 if (thresholds->primary->entries[i].eventfd != eventfd)
3885                         size++;
3886         }
3887
3888         new = thresholds->spare;
3889
3890         /* Set thresholds array to NULL if we don't have thresholds */
3891         if (!size) {
3892                 kfree(new);
3893                 new = NULL;
3894                 goto swap_buffers;
3895         }
3896
3897         new->size = size;
3898
3899         /* Copy thresholds and find current threshold */
3900         new->current_threshold = -1;
3901         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3902                 if (thresholds->primary->entries[i].eventfd == eventfd)
3903                         continue;
3904
3905                 new->entries[j] = thresholds->primary->entries[i];
3906                 if (new->entries[j].threshold < usage) {
3907                         /*
3908                          * new->current_threshold will not be used
3909                          * until rcu_assign_pointer(), so it's safe to increment
3910                          * it here.
3911                          */
3912                         ++new->current_threshold;
3913                 }
3914                 j++;
3915         }
3916
3917 swap_buffers:
3918         /* Swap primary and spare array */
3919         thresholds->spare = thresholds->primary;
3920         rcu_assign_pointer(thresholds->primary, new);
3921
3922         /* To be sure that nobody uses thresholds */
3923         synchronize_rcu();
3924
3925         mutex_unlock(&memcg->thresholds_lock);
3926 }
3927
3928 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3929         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3930 {
3931         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3932         struct mem_cgroup_eventfd_list *event;
3933         int type = MEMFILE_TYPE(cft->private);
3934
3935         BUG_ON(type != _OOM_TYPE);
3936         event = kmalloc(sizeof(*event), GFP_KERNEL);
3937         if (!event)
3938                 return -ENOMEM;
3939
3940         mutex_lock(&memcg_oom_mutex);
3941
3942         event->eventfd = eventfd;
3943         list_add(&event->list, &memcg->oom_notify);
3944
3945         /* already in OOM ? */
3946         if (atomic_read(&memcg->oom_lock))
3947                 eventfd_signal(eventfd, 1);
3948         mutex_unlock(&memcg_oom_mutex);
3949
3950         return 0;
3951 }
3952
3953 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3954         struct cftype *cft, struct eventfd_ctx *eventfd)
3955 {
3956         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3957         struct mem_cgroup_eventfd_list *ev, *tmp;
3958         int type = MEMFILE_TYPE(cft->private);
3959
3960         BUG_ON(type != _OOM_TYPE);
3961
3962         mutex_lock(&memcg_oom_mutex);
3963
3964         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3965                 if (ev->eventfd == eventfd) {
3966                         list_del(&ev->list);
3967                         kfree(ev);
3968                 }
3969         }
3970
3971         mutex_unlock(&memcg_oom_mutex);
3972 }
3973
3974 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
3975         struct cftype *cft,  struct cgroup_map_cb *cb)
3976 {
3977         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3978
3979         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
3980
3981         if (atomic_read(&mem->oom_lock))
3982                 cb->fill(cb, "under_oom", 1);
3983         else
3984                 cb->fill(cb, "under_oom", 0);
3985         return 0;
3986 }
3987
3988 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
3989         struct cftype *cft, u64 val)
3990 {
3991         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3992         struct mem_cgroup *parent;
3993
3994         /* cannot set to root cgroup and only 0 and 1 are allowed */
3995         if (!cgrp->parent || !((val == 0) || (val == 1)))
3996                 return -EINVAL;
3997
3998         parent = mem_cgroup_from_cont(cgrp->parent);
3999
4000         cgroup_lock();
4001         /* oom-kill-disable is a flag for subhierarchy. */
4002         if ((parent->use_hierarchy) ||
4003             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4004                 cgroup_unlock();
4005                 return -EINVAL;
4006         }
4007         mem->oom_kill_disable = val;
4008         if (!val)
4009                 memcg_oom_recover(mem);
4010         cgroup_unlock();
4011         return 0;
4012 }
4013
4014 static struct cftype mem_cgroup_files[] = {
4015         {
4016                 .name = "usage_in_bytes",
4017                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4018                 .read_u64 = mem_cgroup_read,
4019                 .register_event = mem_cgroup_usage_register_event,
4020                 .unregister_event = mem_cgroup_usage_unregister_event,
4021         },
4022         {
4023                 .name = "max_usage_in_bytes",
4024                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4025                 .trigger = mem_cgroup_reset,
4026                 .read_u64 = mem_cgroup_read,
4027         },
4028         {
4029                 .name = "limit_in_bytes",
4030                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4031                 .write_string = mem_cgroup_write,
4032                 .read_u64 = mem_cgroup_read,
4033         },
4034         {
4035                 .name = "soft_limit_in_bytes",
4036                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4037                 .write_string = mem_cgroup_write,
4038                 .read_u64 = mem_cgroup_read,
4039         },
4040         {
4041                 .name = "failcnt",
4042                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4043                 .trigger = mem_cgroup_reset,
4044                 .read_u64 = mem_cgroup_read,
4045         },
4046         {
4047                 .name = "stat",
4048                 .read_map = mem_control_stat_show,
4049         },
4050         {
4051                 .name = "force_empty",
4052                 .trigger = mem_cgroup_force_empty_write,
4053         },
4054         {
4055                 .name = "use_hierarchy",
4056                 .write_u64 = mem_cgroup_hierarchy_write,
4057                 .read_u64 = mem_cgroup_hierarchy_read,
4058         },
4059         {
4060                 .name = "swappiness",
4061                 .read_u64 = mem_cgroup_swappiness_read,
4062                 .write_u64 = mem_cgroup_swappiness_write,
4063         },
4064         {
4065                 .name = "move_charge_at_immigrate",
4066                 .read_u64 = mem_cgroup_move_charge_read,
4067                 .write_u64 = mem_cgroup_move_charge_write,
4068         },
4069         {
4070                 .name = "oom_control",
4071                 .read_map = mem_cgroup_oom_control_read,
4072                 .write_u64 = mem_cgroup_oom_control_write,
4073                 .register_event = mem_cgroup_oom_register_event,
4074                 .unregister_event = mem_cgroup_oom_unregister_event,
4075                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4076         },
4077 };
4078
4079 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4080 static struct cftype memsw_cgroup_files[] = {
4081         {
4082                 .name = "memsw.usage_in_bytes",
4083                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4084                 .read_u64 = mem_cgroup_read,
4085                 .register_event = mem_cgroup_usage_register_event,
4086                 .unregister_event = mem_cgroup_usage_unregister_event,
4087         },
4088         {
4089                 .name = "memsw.max_usage_in_bytes",
4090                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4091                 .trigger = mem_cgroup_reset,
4092                 .read_u64 = mem_cgroup_read,
4093         },
4094         {
4095                 .name = "memsw.limit_in_bytes",
4096                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4097                 .write_string = mem_cgroup_write,
4098                 .read_u64 = mem_cgroup_read,
4099         },
4100         {
4101                 .name = "memsw.failcnt",
4102                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4103                 .trigger = mem_cgroup_reset,
4104                 .read_u64 = mem_cgroup_read,
4105         },
4106 };
4107
4108 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4109 {
4110         if (!do_swap_account)
4111                 return 0;
4112         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4113                                 ARRAY_SIZE(memsw_cgroup_files));
4114 };
4115 #else
4116 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4117 {
4118         return 0;
4119 }
4120 #endif
4121
4122 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4123 {
4124         struct mem_cgroup_per_node *pn;
4125         struct mem_cgroup_per_zone *mz;
4126         enum lru_list l;
4127         int zone, tmp = node;
4128         /*
4129          * This routine is called against possible nodes.
4130          * But it's BUG to call kmalloc() against offline node.
4131          *
4132          * TODO: this routine can waste much memory for nodes which will
4133          *       never be onlined. It's better to use memory hotplug callback
4134          *       function.
4135          */
4136         if (!node_state(node, N_NORMAL_MEMORY))
4137                 tmp = -1;
4138         pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4139         if (!pn)
4140                 return 1;
4141
4142         mem->info.nodeinfo[node] = pn;
4143         memset(pn, 0, sizeof(*pn));
4144
4145         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4146                 mz = &pn->zoneinfo[zone];
4147                 for_each_lru(l)
4148                         INIT_LIST_HEAD(&mz->lists[l]);
4149                 mz->usage_in_excess = 0;
4150                 mz->on_tree = false;
4151                 mz->mem = mem;
4152         }
4153         return 0;
4154 }
4155
4156 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4157 {
4158         kfree(mem->info.nodeinfo[node]);
4159 }
4160
4161 static struct mem_cgroup *mem_cgroup_alloc(void)
4162 {
4163         struct mem_cgroup *mem;
4164         int size = sizeof(struct mem_cgroup);
4165
4166         /* Can be very big if MAX_NUMNODES is very big */
4167         if (size < PAGE_SIZE)
4168                 mem = kmalloc(size, GFP_KERNEL);
4169         else
4170                 mem = vmalloc(size);
4171
4172         if (!mem)
4173                 return NULL;
4174
4175         memset(mem, 0, size);
4176         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4177         if (!mem->stat) {
4178                 if (size < PAGE_SIZE)
4179                         kfree(mem);
4180                 else
4181                         vfree(mem);
4182                 mem = NULL;
4183         }
4184         spin_lock_init(&mem->pcp_counter_lock);
4185         return mem;
4186 }
4187
4188 /*
4189  * At destroying mem_cgroup, references from swap_cgroup can remain.
4190  * (scanning all at force_empty is too costly...)
4191  *
4192  * Instead of clearing all references at force_empty, we remember
4193  * the number of reference from swap_cgroup and free mem_cgroup when
4194  * it goes down to 0.
4195  *
4196  * Removal of cgroup itself succeeds regardless of refs from swap.
4197  */
4198
4199 static void __mem_cgroup_free(struct mem_cgroup *mem)
4200 {
4201         int node;
4202
4203         mem_cgroup_remove_from_trees(mem);
4204         free_css_id(&mem_cgroup_subsys, &mem->css);
4205
4206         for_each_node_state(node, N_POSSIBLE)
4207                 free_mem_cgroup_per_zone_info(mem, node);
4208
4209         free_percpu(mem->stat);
4210         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4211                 kfree(mem);
4212         else
4213                 vfree(mem);
4214 }
4215
4216 static void mem_cgroup_get(struct mem_cgroup *mem)
4217 {
4218         atomic_inc(&mem->refcnt);
4219 }
4220
4221 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4222 {
4223         if (atomic_sub_and_test(count, &mem->refcnt)) {
4224                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4225                 __mem_cgroup_free(mem);
4226                 if (parent)
4227                         mem_cgroup_put(parent);
4228         }
4229 }
4230
4231 static void mem_cgroup_put(struct mem_cgroup *mem)
4232 {
4233         __mem_cgroup_put(mem, 1);
4234 }
4235
4236 /*
4237  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4238  */
4239 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4240 {
4241         if (!mem->res.parent)
4242                 return NULL;
4243         return mem_cgroup_from_res_counter(mem->res.parent, res);
4244 }
4245
4246 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4247 static void __init enable_swap_cgroup(void)
4248 {
4249         if (!mem_cgroup_disabled() && really_do_swap_account)
4250                 do_swap_account = 1;
4251 }
4252 #else
4253 static void __init enable_swap_cgroup(void)
4254 {
4255 }
4256 #endif
4257
4258 static int mem_cgroup_soft_limit_tree_init(void)
4259 {
4260         struct mem_cgroup_tree_per_node *rtpn;
4261         struct mem_cgroup_tree_per_zone *rtpz;
4262         int tmp, node, zone;
4263
4264         for_each_node_state(node, N_POSSIBLE) {
4265                 tmp = node;
4266                 if (!node_state(node, N_NORMAL_MEMORY))
4267                         tmp = -1;
4268                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4269                 if (!rtpn)
4270                         return 1;
4271
4272                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4273
4274                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4275                         rtpz = &rtpn->rb_tree_per_zone[zone];
4276                         rtpz->rb_root = RB_ROOT;
4277                         spin_lock_init(&rtpz->lock);
4278                 }
4279         }
4280         return 0;
4281 }
4282
4283 static struct cgroup_subsys_state * __ref
4284 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4285 {
4286         struct mem_cgroup *mem, *parent;
4287         long error = -ENOMEM;
4288         int node;
4289
4290         mem = mem_cgroup_alloc();
4291         if (!mem)
4292                 return ERR_PTR(error);
4293
4294         for_each_node_state(node, N_POSSIBLE)
4295                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4296                         goto free_out;
4297
4298         /* root ? */
4299         if (cont->parent == NULL) {
4300                 int cpu;
4301                 enable_swap_cgroup();
4302                 parent = NULL;
4303                 root_mem_cgroup = mem;
4304                 if (mem_cgroup_soft_limit_tree_init())
4305                         goto free_out;
4306                 for_each_possible_cpu(cpu) {
4307                         struct memcg_stock_pcp *stock =
4308                                                 &per_cpu(memcg_stock, cpu);
4309                         INIT_WORK(&stock->work, drain_local_stock);
4310                 }
4311                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4312         } else {
4313                 parent = mem_cgroup_from_cont(cont->parent);
4314                 mem->use_hierarchy = parent->use_hierarchy;
4315                 mem->oom_kill_disable = parent->oom_kill_disable;
4316         }
4317
4318         if (parent && parent->use_hierarchy) {
4319                 res_counter_init(&mem->res, &parent->res);
4320                 res_counter_init(&mem->memsw, &parent->memsw);
4321                 /*
4322                  * We increment refcnt of the parent to ensure that we can
4323                  * safely access it on res_counter_charge/uncharge.
4324                  * This refcnt will be decremented when freeing this
4325                  * mem_cgroup(see mem_cgroup_put).
4326                  */
4327                 mem_cgroup_get(parent);
4328         } else {
4329                 res_counter_init(&mem->res, NULL);
4330                 res_counter_init(&mem->memsw, NULL);
4331         }
4332         mem->last_scanned_child = 0;
4333         spin_lock_init(&mem->reclaim_param_lock);
4334         INIT_LIST_HEAD(&mem->oom_notify);
4335
4336         if (parent)
4337                 mem->swappiness = get_swappiness(parent);
4338         atomic_set(&mem->refcnt, 1);
4339         mem->move_charge_at_immigrate = 0;
4340         mutex_init(&mem->thresholds_lock);
4341         return &mem->css;
4342 free_out:
4343         __mem_cgroup_free(mem);
4344         root_mem_cgroup = NULL;
4345         return ERR_PTR(error);
4346 }
4347
4348 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4349                                         struct cgroup *cont)
4350 {
4351         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4352
4353         return mem_cgroup_force_empty(mem, false);
4354 }
4355
4356 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4357                                 struct cgroup *cont)
4358 {
4359         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4360
4361         mem_cgroup_put(mem);
4362 }
4363
4364 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4365                                 struct cgroup *cont)
4366 {
4367         int ret;
4368
4369         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4370                                 ARRAY_SIZE(mem_cgroup_files));
4371
4372         if (!ret)
4373                 ret = register_memsw_files(cont, ss);
4374         return ret;
4375 }
4376
4377 #ifdef CONFIG_MMU
4378 /* Handlers for move charge at task migration. */
4379 #define PRECHARGE_COUNT_AT_ONCE 256
4380 static int mem_cgroup_do_precharge(unsigned long count)
4381 {
4382         int ret = 0;
4383         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4384         struct mem_cgroup *mem = mc.to;
4385
4386         if (mem_cgroup_is_root(mem)) {
4387                 mc.precharge += count;
4388                 /* we don't need css_get for root */
4389                 return ret;
4390         }
4391         /* try to charge at once */
4392         if (count > 1) {
4393                 struct res_counter *dummy;
4394                 /*
4395                  * "mem" cannot be under rmdir() because we've already checked
4396                  * by cgroup_lock_live_cgroup() that it is not removed and we
4397                  * are still under the same cgroup_mutex. So we can postpone
4398                  * css_get().
4399                  */
4400                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4401                         goto one_by_one;
4402                 if (do_swap_account && res_counter_charge(&mem->memsw,
4403                                                 PAGE_SIZE * count, &dummy)) {
4404                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4405                         goto one_by_one;
4406                 }
4407                 mc.precharge += count;
4408                 return ret;
4409         }
4410 one_by_one:
4411         /* fall back to one by one charge */
4412         while (count--) {
4413                 if (signal_pending(current)) {
4414                         ret = -EINTR;
4415                         break;
4416                 }
4417                 if (!batch_count--) {
4418                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4419                         cond_resched();
4420                 }
4421                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4422                 if (ret || !mem)
4423                         /* mem_cgroup_clear_mc() will do uncharge later */
4424                         return -ENOMEM;
4425                 mc.precharge++;
4426         }
4427         return ret;
4428 }
4429
4430 /**
4431  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4432  * @vma: the vma the pte to be checked belongs
4433  * @addr: the address corresponding to the pte to be checked
4434  * @ptent: the pte to be checked
4435  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4436  *
4437  * Returns
4438  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4439  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4440  *     move charge. if @target is not NULL, the page is stored in target->page
4441  *     with extra refcnt got(Callers should handle it).
4442  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4443  *     target for charge migration. if @target is not NULL, the entry is stored
4444  *     in target->ent.
4445  *
4446  * Called with pte lock held.
4447  */
4448 union mc_target {
4449         struct page     *page;
4450         swp_entry_t     ent;
4451 };
4452
4453 enum mc_target_type {
4454         MC_TARGET_NONE, /* not used */
4455         MC_TARGET_PAGE,
4456         MC_TARGET_SWAP,
4457 };
4458
4459 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4460                                                 unsigned long addr, pte_t ptent)
4461 {
4462         struct page *page = vm_normal_page(vma, addr, ptent);
4463
4464         if (!page || !page_mapped(page))
4465                 return NULL;
4466         if (PageAnon(page)) {
4467                 /* we don't move shared anon */
4468                 if (!move_anon() || page_mapcount(page) > 2)
4469                         return NULL;
4470         } else if (!move_file())
4471                 /* we ignore mapcount for file pages */
4472                 return NULL;
4473         if (!get_page_unless_zero(page))
4474                 return NULL;
4475
4476         return page;
4477 }
4478
4479 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4480                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4481 {
4482         int usage_count;
4483         struct page *page = NULL;
4484         swp_entry_t ent = pte_to_swp_entry(ptent);
4485
4486         if (!move_anon() || non_swap_entry(ent))
4487                 return NULL;
4488         usage_count = mem_cgroup_count_swap_user(ent, &page);
4489         if (usage_count > 1) { /* we don't move shared anon */
4490                 if (page)
4491                         put_page(page);
4492                 return NULL;
4493         }
4494         if (do_swap_account)
4495                 entry->val = ent.val;
4496
4497         return page;
4498 }
4499
4500 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4501                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4502 {
4503         struct page *page = NULL;
4504         struct inode *inode;
4505         struct address_space *mapping;
4506         pgoff_t pgoff;
4507
4508         if (!vma->vm_file) /* anonymous vma */
4509                 return NULL;
4510         if (!move_file())
4511                 return NULL;
4512
4513         inode = vma->vm_file->f_path.dentry->d_inode;
4514         mapping = vma->vm_file->f_mapping;
4515         if (pte_none(ptent))
4516                 pgoff = linear_page_index(vma, addr);
4517         else /* pte_file(ptent) is true */
4518                 pgoff = pte_to_pgoff(ptent);
4519
4520         /* page is moved even if it's not RSS of this task(page-faulted). */
4521         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4522                 page = find_get_page(mapping, pgoff);
4523         } else { /* shmem/tmpfs file. we should take account of swap too. */
4524                 swp_entry_t ent;
4525                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4526                 if (do_swap_account)
4527                         entry->val = ent.val;
4528         }
4529
4530         return page;
4531 }
4532
4533 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4534                 unsigned long addr, pte_t ptent, union mc_target *target)
4535 {
4536         struct page *page = NULL;
4537         struct page_cgroup *pc;
4538         int ret = 0;
4539         swp_entry_t ent = { .val = 0 };
4540
4541         if (pte_present(ptent))
4542                 page = mc_handle_present_pte(vma, addr, ptent);
4543         else if (is_swap_pte(ptent))
4544                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4545         else if (pte_none(ptent) || pte_file(ptent))
4546                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4547
4548         if (!page && !ent.val)
4549                 return 0;
4550         if (page) {
4551                 pc = lookup_page_cgroup(page);
4552                 /*
4553                  * Do only loose check w/o page_cgroup lock.
4554                  * mem_cgroup_move_account() checks the pc is valid or not under
4555                  * the lock.
4556                  */
4557                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4558                         ret = MC_TARGET_PAGE;
4559                         if (target)
4560                                 target->page = page;
4561                 }
4562                 if (!ret || !target)
4563                         put_page(page);
4564         }
4565         /* There is a swap entry and a page doesn't exist or isn't charged */
4566         if (ent.val && !ret &&
4567                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4568                 ret = MC_TARGET_SWAP;
4569                 if (target)
4570                         target->ent = ent;
4571         }
4572         return ret;
4573 }
4574
4575 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4576                                         unsigned long addr, unsigned long end,
4577                                         struct mm_walk *walk)
4578 {
4579         struct vm_area_struct *vma = walk->private;
4580         pte_t *pte;
4581         spinlock_t *ptl;
4582
4583         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4584         for (; addr != end; pte++, addr += PAGE_SIZE)
4585                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4586                         mc.precharge++; /* increment precharge temporarily */
4587         pte_unmap_unlock(pte - 1, ptl);
4588         cond_resched();
4589
4590         return 0;
4591 }
4592
4593 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4594 {
4595         unsigned long precharge;
4596         struct vm_area_struct *vma;
4597
4598         down_read(&mm->mmap_sem);
4599         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4600                 struct mm_walk mem_cgroup_count_precharge_walk = {
4601                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4602                         .mm = mm,
4603                         .private = vma,
4604                 };
4605                 if (is_vm_hugetlb_page(vma))
4606                         continue;
4607                 walk_page_range(vma->vm_start, vma->vm_end,
4608                                         &mem_cgroup_count_precharge_walk);
4609         }
4610         up_read(&mm->mmap_sem);
4611
4612         precharge = mc.precharge;
4613         mc.precharge = 0;
4614
4615         return precharge;
4616 }
4617
4618 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4619 {
4620         return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4621 }
4622
4623 static void mem_cgroup_clear_mc(void)
4624 {
4625         struct mem_cgroup *from = mc.from;
4626         struct mem_cgroup *to = mc.to;
4627
4628         /* we must uncharge all the leftover precharges from mc.to */
4629         if (mc.precharge) {
4630                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4631                 mc.precharge = 0;
4632         }
4633         /*
4634          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4635          * we must uncharge here.
4636          */
4637         if (mc.moved_charge) {
4638                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4639                 mc.moved_charge = 0;
4640         }
4641         /* we must fixup refcnts and charges */
4642         if (mc.moved_swap) {
4643                 /* uncharge swap account from the old cgroup */
4644                 if (!mem_cgroup_is_root(mc.from))
4645                         res_counter_uncharge(&mc.from->memsw,
4646                                                 PAGE_SIZE * mc.moved_swap);
4647                 __mem_cgroup_put(mc.from, mc.moved_swap);
4648
4649                 if (!mem_cgroup_is_root(mc.to)) {
4650                         /*
4651                          * we charged both to->res and to->memsw, so we should
4652                          * uncharge to->res.
4653                          */
4654                         res_counter_uncharge(&mc.to->res,
4655                                                 PAGE_SIZE * mc.moved_swap);
4656                 }
4657                 /* we've already done mem_cgroup_get(mc.to) */
4658
4659                 mc.moved_swap = 0;
4660         }
4661         spin_lock(&mc.lock);
4662         mc.from = NULL;
4663         mc.to = NULL;
4664         mc.moving_task = NULL;
4665         spin_unlock(&mc.lock);
4666         mem_cgroup_end_move(from);
4667         memcg_oom_recover(from);
4668         memcg_oom_recover(to);
4669         wake_up_all(&mc.waitq);
4670 }
4671
4672 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4673                                 struct cgroup *cgroup,
4674                                 struct task_struct *p,
4675                                 bool threadgroup)
4676 {
4677         int ret = 0;
4678         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4679
4680         if (mem->move_charge_at_immigrate) {
4681                 struct mm_struct *mm;
4682                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4683
4684                 VM_BUG_ON(from == mem);
4685
4686                 mm = get_task_mm(p);
4687                 if (!mm)
4688                         return 0;
4689                 /* We move charges only when we move a owner of the mm */
4690                 if (mm->owner == p) {
4691                         VM_BUG_ON(mc.from);
4692                         VM_BUG_ON(mc.to);
4693                         VM_BUG_ON(mc.precharge);
4694                         VM_BUG_ON(mc.moved_charge);
4695                         VM_BUG_ON(mc.moved_swap);
4696                         VM_BUG_ON(mc.moving_task);
4697                         mem_cgroup_start_move(from);
4698                         spin_lock(&mc.lock);
4699                         mc.from = from;
4700                         mc.to = mem;
4701                         mc.precharge = 0;
4702                         mc.moved_charge = 0;
4703                         mc.moved_swap = 0;
4704                         mc.moving_task = current;
4705                         spin_unlock(&mc.lock);
4706
4707                         ret = mem_cgroup_precharge_mc(mm);
4708                         if (ret)
4709                                 mem_cgroup_clear_mc();
4710                 }
4711                 mmput(mm);
4712         }
4713         return ret;
4714 }
4715
4716 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4717                                 struct cgroup *cgroup,
4718                                 struct task_struct *p,
4719                                 bool threadgroup)
4720 {
4721         mem_cgroup_clear_mc();
4722 }
4723
4724 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4725                                 unsigned long addr, unsigned long end,
4726                                 struct mm_walk *walk)
4727 {
4728         int ret = 0;
4729         struct vm_area_struct *vma = walk->private;
4730         pte_t *pte;
4731         spinlock_t *ptl;
4732
4733 retry:
4734         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4735         for (; addr != end; addr += PAGE_SIZE) {
4736                 pte_t ptent = *(pte++);
4737                 union mc_target target;
4738                 int type;
4739                 struct page *page;
4740                 struct page_cgroup *pc;
4741                 swp_entry_t ent;
4742
4743                 if (!mc.precharge)
4744                         break;
4745
4746                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4747                 switch (type) {
4748                 case MC_TARGET_PAGE:
4749                         page = target.page;
4750                         if (isolate_lru_page(page))
4751                                 goto put;
4752                         pc = lookup_page_cgroup(page);
4753                         if (!mem_cgroup_move_account(pc,
4754                                                 mc.from, mc.to, false)) {
4755                                 mc.precharge--;
4756                                 /* we uncharge from mc.from later. */
4757                                 mc.moved_charge++;
4758                         }
4759                         putback_lru_page(page);
4760 put:                    /* is_target_pte_for_mc() gets the page */
4761                         put_page(page);
4762                         break;
4763                 case MC_TARGET_SWAP:
4764                         ent = target.ent;
4765                         if (!mem_cgroup_move_swap_account(ent,
4766                                                 mc.from, mc.to, false)) {
4767                                 mc.precharge--;
4768                                 /* we fixup refcnts and charges later. */
4769                                 mc.moved_swap++;
4770                         }
4771                         break;
4772                 default:
4773                         break;
4774                 }
4775         }
4776         pte_unmap_unlock(pte - 1, ptl);
4777         cond_resched();
4778
4779         if (addr != end) {
4780                 /*
4781                  * We have consumed all precharges we got in can_attach().
4782                  * We try charge one by one, but don't do any additional
4783                  * charges to mc.to if we have failed in charge once in attach()
4784                  * phase.
4785                  */
4786                 ret = mem_cgroup_do_precharge(1);
4787                 if (!ret)
4788                         goto retry;
4789         }
4790
4791         return ret;
4792 }
4793
4794 static void mem_cgroup_move_charge(struct mm_struct *mm)
4795 {
4796         struct vm_area_struct *vma;
4797
4798         lru_add_drain_all();
4799         down_read(&mm->mmap_sem);
4800         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4801                 int ret;
4802                 struct mm_walk mem_cgroup_move_charge_walk = {
4803                         .pmd_entry = mem_cgroup_move_charge_pte_range,
4804                         .mm = mm,
4805                         .private = vma,
4806                 };
4807                 if (is_vm_hugetlb_page(vma))
4808                         continue;
4809                 ret = walk_page_range(vma->vm_start, vma->vm_end,
4810                                                 &mem_cgroup_move_charge_walk);
4811                 if (ret)
4812                         /*
4813                          * means we have consumed all precharges and failed in
4814                          * doing additional charge. Just abandon here.
4815                          */
4816                         break;
4817         }
4818         up_read(&mm->mmap_sem);
4819 }
4820
4821 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4822                                 struct cgroup *cont,
4823                                 struct cgroup *old_cont,
4824                                 struct task_struct *p,
4825                                 bool threadgroup)
4826 {
4827         struct mm_struct *mm;
4828
4829         if (!mc.to)
4830                 /* no need to move charge */
4831                 return;
4832
4833         mm = get_task_mm(p);
4834         if (mm) {
4835                 mem_cgroup_move_charge(mm);
4836                 mmput(mm);
4837         }
4838         mem_cgroup_clear_mc();
4839 }
4840 #else   /* !CONFIG_MMU */
4841 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4842                                 struct cgroup *cgroup,
4843                                 struct task_struct *p,
4844                                 bool threadgroup)
4845 {
4846         return 0;
4847 }
4848 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4849                                 struct cgroup *cgroup,
4850                                 struct task_struct *p,
4851                                 bool threadgroup)
4852 {
4853 }
4854 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4855                                 struct cgroup *cont,
4856                                 struct cgroup *old_cont,
4857                                 struct task_struct *p,
4858                                 bool threadgroup)
4859 {
4860 }
4861 #endif
4862
4863 struct cgroup_subsys mem_cgroup_subsys = {
4864         .name = "memory",
4865         .subsys_id = mem_cgroup_subsys_id,
4866         .create = mem_cgroup_create,
4867         .pre_destroy = mem_cgroup_pre_destroy,
4868         .destroy = mem_cgroup_destroy,
4869         .populate = mem_cgroup_populate,
4870         .can_attach = mem_cgroup_can_attach,
4871         .cancel_attach = mem_cgroup_cancel_attach,
4872         .attach = mem_cgroup_move_task,
4873         .early_init = 0,
4874         .use_id = 1,
4875 };
4876
4877 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4878
4879 static int __init disable_swap_account(char *s)
4880 {
4881         really_do_swap_account = 0;
4882         return 1;
4883 }
4884 __setup("noswapaccount", disable_swap_account);
4885 #endif