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