memcg: keep only one charge cancelling function
[linux-2.6.git] / mm / memcontrol.c
1 /* memcontrol.c - Memory Controller
2  *
3  * Copyright IBM Corporation, 2007
4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5  *
6  * Copyright 2007 OpenVZ SWsoft Inc
7  * Author: Pavel Emelianov <xemul@openvz.org>
8  *
9  * Memory thresholds
10  * Copyright (C) 2009 Nokia Corporation
11  * Author: Kirill A. Shutemov
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2 of the License, or
16  * (at your option) any later version.
17  *
18  * This program is distributed in the hope that it will be useful,
19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
21  * GNU General Public License for more details.
22  */
23
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
44 #include <linux/fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
51 #include "internal.h"
52
53 #include <asm/uaccess.h>
54
55 #include <trace/events/vmscan.h>
56
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES      5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
60
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
64
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata = 1;
68 #else
69 static int really_do_swap_account __initdata = 0;
70 #endif
71
72 #else
73 #define do_swap_account         (0)
74 #endif
75
76 /*
77  * Per memcg event counter is incremented at every pagein/pageout. This counter
78  * is used for trigger some periodic events. This is straightforward and better
79  * than using jiffies etc. to handle periodic memcg event.
80  *
81  * These values will be used as !((event) & ((1 <<(thresh)) - 1))
82  */
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
85
86 /*
87  * Statistics for memory cgroup.
88  */
89 enum mem_cgroup_stat_index {
90         /*
91          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92          */
93         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
94         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
95         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
96         MEM_CGROUP_STAT_PGPGIN_COUNT,   /* # of pages paged in */
97         MEM_CGROUP_STAT_PGPGOUT_COUNT,  /* # of pages paged out */
98         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
99         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
100         /* incremented at every  pagein/pageout */
101         MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
102         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
103
104         MEM_CGROUP_STAT_NSTATS,
105 };
106
107 struct mem_cgroup_stat_cpu {
108         s64 count[MEM_CGROUP_STAT_NSTATS];
109 };
110
111 /*
112  * per-zone information in memory controller.
113  */
114 struct mem_cgroup_per_zone {
115         /*
116          * spin_lock to protect the per cgroup LRU
117          */
118         struct list_head        lists[NR_LRU_LISTS];
119         unsigned long           count[NR_LRU_LISTS];
120
121         struct zone_reclaim_stat reclaim_stat;
122         struct rb_node          tree_node;      /* RB tree node */
123         unsigned long long      usage_in_excess;/* Set to the value by which */
124                                                 /* the soft limit is exceeded*/
125         bool                    on_tree;
126         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
127                                                 /* use container_of        */
128 };
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
131
132 struct mem_cgroup_per_node {
133         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
134 };
135
136 struct mem_cgroup_lru_info {
137         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
138 };
139
140 /*
141  * Cgroups above their limits are maintained in a RB-Tree, independent of
142  * their hierarchy representation
143  */
144
145 struct mem_cgroup_tree_per_zone {
146         struct rb_root rb_root;
147         spinlock_t lock;
148 };
149
150 struct mem_cgroup_tree_per_node {
151         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
152 };
153
154 struct mem_cgroup_tree {
155         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
156 };
157
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
159
160 struct mem_cgroup_threshold {
161         struct eventfd_ctx *eventfd;
162         u64 threshold;
163 };
164
165 /* For threshold */
166 struct mem_cgroup_threshold_ary {
167         /* An array index points to threshold just below usage. */
168         int current_threshold;
169         /* Size of entries[] */
170         unsigned int size;
171         /* Array of thresholds */
172         struct mem_cgroup_threshold entries[0];
173 };
174
175 struct mem_cgroup_thresholds {
176         /* Primary thresholds array */
177         struct mem_cgroup_threshold_ary *primary;
178         /*
179          * Spare threshold array.
180          * This is needed to make mem_cgroup_unregister_event() "never fail".
181          * It must be able to store at least primary->size - 1 entries.
182          */
183         struct mem_cgroup_threshold_ary *spare;
184 };
185
186 /* for OOM */
187 struct mem_cgroup_eventfd_list {
188         struct list_head list;
189         struct eventfd_ctx *eventfd;
190 };
191
192 static void mem_cgroup_threshold(struct mem_cgroup *mem);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
194
195 /*
196  * The memory controller data structure. The memory controller controls both
197  * page cache and RSS per cgroup. We would eventually like to provide
198  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199  * to help the administrator determine what knobs to tune.
200  *
201  * TODO: Add a water mark for the memory controller. Reclaim will begin when
202  * we hit the water mark. May be even add a low water mark, such that
203  * no reclaim occurs from a cgroup at it's low water mark, this is
204  * a feature that will be implemented much later in the future.
205  */
206 struct mem_cgroup {
207         struct cgroup_subsys_state css;
208         /*
209          * the counter to account for memory usage
210          */
211         struct res_counter res;
212         /*
213          * the counter to account for mem+swap usage.
214          */
215         struct res_counter memsw;
216         /*
217          * Per cgroup active and inactive list, similar to the
218          * per zone LRU lists.
219          */
220         struct mem_cgroup_lru_info info;
221         /*
222          * While reclaiming in a hierarchy, we cache the last child we
223          * reclaimed from.
224          */
225         int last_scanned_child;
226         /*
227          * Should the accounting and control be hierarchical, per subtree?
228          */
229         bool use_hierarchy;
230         atomic_t        oom_lock;
231         atomic_t        refcnt;
232
233         unsigned int    swappiness;
234         /* OOM-Killer disable */
235         int             oom_kill_disable;
236
237         /* set when res.limit == memsw.limit */
238         bool            memsw_is_minimum;
239
240         /* protect arrays of thresholds */
241         struct mutex thresholds_lock;
242
243         /* thresholds for memory usage. RCU-protected */
244         struct mem_cgroup_thresholds thresholds;
245
246         /* thresholds for mem+swap usage. RCU-protected */
247         struct mem_cgroup_thresholds memsw_thresholds;
248
249         /* For oom notifier event fd */
250         struct list_head oom_notify;
251
252         /*
253          * Should we move charges of a task when a task is moved into this
254          * mem_cgroup ? And what type of charges should we move ?
255          */
256         unsigned long   move_charge_at_immigrate;
257         /*
258          * percpu counter.
259          */
260         struct mem_cgroup_stat_cpu *stat;
261         /*
262          * used when a cpu is offlined or other synchronizations
263          * See mem_cgroup_read_stat().
264          */
265         struct mem_cgroup_stat_cpu nocpu_base;
266         spinlock_t pcp_counter_lock;
267 };
268
269 /* Stuffs for move charges at task migration. */
270 /*
271  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
272  * left-shifted bitmap of these types.
273  */
274 enum move_type {
275         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
276         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
277         NR_MOVE_TYPE,
278 };
279
280 /* "mc" and its members are protected by cgroup_mutex */
281 static struct move_charge_struct {
282         spinlock_t        lock; /* for from, to */
283         struct mem_cgroup *from;
284         struct mem_cgroup *to;
285         unsigned long precharge;
286         unsigned long moved_charge;
287         unsigned long moved_swap;
288         struct task_struct *moving_task;        /* a task moving charges */
289         wait_queue_head_t waitq;                /* a waitq for other context */
290 } mc = {
291         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
292         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
293 };
294
295 static bool move_anon(void)
296 {
297         return test_bit(MOVE_CHARGE_TYPE_ANON,
298                                         &mc.to->move_charge_at_immigrate);
299 }
300
301 static bool move_file(void)
302 {
303         return test_bit(MOVE_CHARGE_TYPE_FILE,
304                                         &mc.to->move_charge_at_immigrate);
305 }
306
307 /*
308  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
309  * limit reclaim to prevent infinite loops, if they ever occur.
310  */
311 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
312 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
313
314 enum charge_type {
315         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
316         MEM_CGROUP_CHARGE_TYPE_MAPPED,
317         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
318         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
319         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
320         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
321         NR_CHARGE_TYPE,
322 };
323
324 /* for encoding cft->private value on file */
325 #define _MEM                    (0)
326 #define _MEMSWAP                (1)
327 #define _OOM_TYPE               (2)
328 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
329 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
330 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
331 /* Used for OOM nofiier */
332 #define OOM_CONTROL             (0)
333
334 /*
335  * Reclaim flags for mem_cgroup_hierarchical_reclaim
336  */
337 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
338 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
339 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
340 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
341 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
342 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
343
344 static void mem_cgroup_get(struct mem_cgroup *mem);
345 static void mem_cgroup_put(struct mem_cgroup *mem);
346 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
347 static void drain_all_stock_async(void);
348
349 static struct mem_cgroup_per_zone *
350 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
351 {
352         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
353 }
354
355 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
356 {
357         return &mem->css;
358 }
359
360 static struct mem_cgroup_per_zone *
361 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
362 {
363         int nid = page_to_nid(page);
364         int zid = page_zonenum(page);
365
366         return mem_cgroup_zoneinfo(mem, nid, zid);
367 }
368
369 static struct mem_cgroup_tree_per_zone *
370 soft_limit_tree_node_zone(int nid, int zid)
371 {
372         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
373 }
374
375 static struct mem_cgroup_tree_per_zone *
376 soft_limit_tree_from_page(struct page *page)
377 {
378         int nid = page_to_nid(page);
379         int zid = page_zonenum(page);
380
381         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
382 }
383
384 static void
385 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
386                                 struct mem_cgroup_per_zone *mz,
387                                 struct mem_cgroup_tree_per_zone *mctz,
388                                 unsigned long long new_usage_in_excess)
389 {
390         struct rb_node **p = &mctz->rb_root.rb_node;
391         struct rb_node *parent = NULL;
392         struct mem_cgroup_per_zone *mz_node;
393
394         if (mz->on_tree)
395                 return;
396
397         mz->usage_in_excess = new_usage_in_excess;
398         if (!mz->usage_in_excess)
399                 return;
400         while (*p) {
401                 parent = *p;
402                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
403                                         tree_node);
404                 if (mz->usage_in_excess < mz_node->usage_in_excess)
405                         p = &(*p)->rb_left;
406                 /*
407                  * We can't avoid mem cgroups that are over their soft
408                  * limit by the same amount
409                  */
410                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
411                         p = &(*p)->rb_right;
412         }
413         rb_link_node(&mz->tree_node, parent, p);
414         rb_insert_color(&mz->tree_node, &mctz->rb_root);
415         mz->on_tree = true;
416 }
417
418 static void
419 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
420                                 struct mem_cgroup_per_zone *mz,
421                                 struct mem_cgroup_tree_per_zone *mctz)
422 {
423         if (!mz->on_tree)
424                 return;
425         rb_erase(&mz->tree_node, &mctz->rb_root);
426         mz->on_tree = false;
427 }
428
429 static void
430 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
431                                 struct mem_cgroup_per_zone *mz,
432                                 struct mem_cgroup_tree_per_zone *mctz)
433 {
434         spin_lock(&mctz->lock);
435         __mem_cgroup_remove_exceeded(mem, mz, mctz);
436         spin_unlock(&mctz->lock);
437 }
438
439
440 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
441 {
442         unsigned long long excess;
443         struct mem_cgroup_per_zone *mz;
444         struct mem_cgroup_tree_per_zone *mctz;
445         int nid = page_to_nid(page);
446         int zid = page_zonenum(page);
447         mctz = soft_limit_tree_from_page(page);
448
449         /*
450          * Necessary to update all ancestors when hierarchy is used.
451          * because their event counter is not touched.
452          */
453         for (; mem; mem = parent_mem_cgroup(mem)) {
454                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
455                 excess = res_counter_soft_limit_excess(&mem->res);
456                 /*
457                  * We have to update the tree if mz is on RB-tree or
458                  * mem is over its softlimit.
459                  */
460                 if (excess || mz->on_tree) {
461                         spin_lock(&mctz->lock);
462                         /* if on-tree, remove it */
463                         if (mz->on_tree)
464                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
465                         /*
466                          * Insert again. mz->usage_in_excess will be updated.
467                          * If excess is 0, no tree ops.
468                          */
469                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
470                         spin_unlock(&mctz->lock);
471                 }
472         }
473 }
474
475 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
476 {
477         int node, zone;
478         struct mem_cgroup_per_zone *mz;
479         struct mem_cgroup_tree_per_zone *mctz;
480
481         for_each_node_state(node, N_POSSIBLE) {
482                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
483                         mz = mem_cgroup_zoneinfo(mem, node, zone);
484                         mctz = soft_limit_tree_node_zone(node, zone);
485                         mem_cgroup_remove_exceeded(mem, mz, mctz);
486                 }
487         }
488 }
489
490 static struct mem_cgroup_per_zone *
491 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
492 {
493         struct rb_node *rightmost = NULL;
494         struct mem_cgroup_per_zone *mz;
495
496 retry:
497         mz = NULL;
498         rightmost = rb_last(&mctz->rb_root);
499         if (!rightmost)
500                 goto done;              /* Nothing to reclaim from */
501
502         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
503         /*
504          * Remove the node now but someone else can add it back,
505          * we will to add it back at the end of reclaim to its correct
506          * position in the tree.
507          */
508         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
509         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
510                 !css_tryget(&mz->mem->css))
511                 goto retry;
512 done:
513         return mz;
514 }
515
516 static struct mem_cgroup_per_zone *
517 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
518 {
519         struct mem_cgroup_per_zone *mz;
520
521         spin_lock(&mctz->lock);
522         mz = __mem_cgroup_largest_soft_limit_node(mctz);
523         spin_unlock(&mctz->lock);
524         return mz;
525 }
526
527 /*
528  * Implementation Note: reading percpu statistics for memcg.
529  *
530  * Both of vmstat[] and percpu_counter has threshold and do periodic
531  * synchronization to implement "quick" read. There are trade-off between
532  * reading cost and precision of value. Then, we may have a chance to implement
533  * a periodic synchronizion of counter in memcg's counter.
534  *
535  * But this _read() function is used for user interface now. The user accounts
536  * memory usage by memory cgroup and he _always_ requires exact value because
537  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
538  * have to visit all online cpus and make sum. So, for now, unnecessary
539  * synchronization is not implemented. (just implemented for cpu hotplug)
540  *
541  * If there are kernel internal actions which can make use of some not-exact
542  * value, and reading all cpu value can be performance bottleneck in some
543  * common workload, threashold and synchonization as vmstat[] should be
544  * implemented.
545  */
546 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
547                 enum mem_cgroup_stat_index idx)
548 {
549         int cpu;
550         s64 val = 0;
551
552         get_online_cpus();
553         for_each_online_cpu(cpu)
554                 val += per_cpu(mem->stat->count[idx], cpu);
555 #ifdef CONFIG_HOTPLUG_CPU
556         spin_lock(&mem->pcp_counter_lock);
557         val += mem->nocpu_base.count[idx];
558         spin_unlock(&mem->pcp_counter_lock);
559 #endif
560         put_online_cpus();
561         return val;
562 }
563
564 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
565 {
566         s64 ret;
567
568         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
569         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
570         return ret;
571 }
572
573 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
574                                          bool charge)
575 {
576         int val = (charge) ? 1 : -1;
577         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
578 }
579
580 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
581                                          bool file, int nr_pages)
582 {
583         preempt_disable();
584
585         if (file)
586                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
587         else
588                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
589
590         /* pagein of a big page is an event. So, ignore page size */
591         if (nr_pages > 0)
592                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
593         else {
594                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
595                 nr_pages = -nr_pages; /* for event */
596         }
597
598         __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages);
599
600         preempt_enable();
601 }
602
603 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
604                                         enum lru_list idx)
605 {
606         int nid, zid;
607         struct mem_cgroup_per_zone *mz;
608         u64 total = 0;
609
610         for_each_online_node(nid)
611                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
612                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
613                         total += MEM_CGROUP_ZSTAT(mz, idx);
614                 }
615         return total;
616 }
617
618 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
619 {
620         s64 val;
621
622         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
623
624         return !(val & ((1 << event_mask_shift) - 1));
625 }
626
627 /*
628  * Check events in order.
629  *
630  */
631 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
632 {
633         /* threshold event is triggered in finer grain than soft limit */
634         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
635                 mem_cgroup_threshold(mem);
636                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
637                         mem_cgroup_update_tree(mem, page);
638         }
639 }
640
641 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
642 {
643         return container_of(cgroup_subsys_state(cont,
644                                 mem_cgroup_subsys_id), struct mem_cgroup,
645                                 css);
646 }
647
648 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
649 {
650         /*
651          * mm_update_next_owner() may clear mm->owner to NULL
652          * if it races with swapoff, page migration, etc.
653          * So this can be called with p == NULL.
654          */
655         if (unlikely(!p))
656                 return NULL;
657
658         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
659                                 struct mem_cgroup, css);
660 }
661
662 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
663 {
664         struct mem_cgroup *mem = NULL;
665
666         if (!mm)
667                 return NULL;
668         /*
669          * Because we have no locks, mm->owner's may be being moved to other
670          * cgroup. We use css_tryget() here even if this looks
671          * pessimistic (rather than adding locks here).
672          */
673         rcu_read_lock();
674         do {
675                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
676                 if (unlikely(!mem))
677                         break;
678         } while (!css_tryget(&mem->css));
679         rcu_read_unlock();
680         return mem;
681 }
682
683 /* The caller has to guarantee "mem" exists before calling this */
684 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
685 {
686         struct cgroup_subsys_state *css;
687         int found;
688
689         if (!mem) /* ROOT cgroup has the smallest ID */
690                 return root_mem_cgroup; /*css_put/get against root is ignored*/
691         if (!mem->use_hierarchy) {
692                 if (css_tryget(&mem->css))
693                         return mem;
694                 return NULL;
695         }
696         rcu_read_lock();
697         /*
698          * searching a memory cgroup which has the smallest ID under given
699          * ROOT cgroup. (ID >= 1)
700          */
701         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
702         if (css && css_tryget(css))
703                 mem = container_of(css, struct mem_cgroup, css);
704         else
705                 mem = NULL;
706         rcu_read_unlock();
707         return mem;
708 }
709
710 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
711                                         struct mem_cgroup *root,
712                                         bool cond)
713 {
714         int nextid = css_id(&iter->css) + 1;
715         int found;
716         int hierarchy_used;
717         struct cgroup_subsys_state *css;
718
719         hierarchy_used = iter->use_hierarchy;
720
721         css_put(&iter->css);
722         /* If no ROOT, walk all, ignore hierarchy */
723         if (!cond || (root && !hierarchy_used))
724                 return NULL;
725
726         if (!root)
727                 root = root_mem_cgroup;
728
729         do {
730                 iter = NULL;
731                 rcu_read_lock();
732
733                 css = css_get_next(&mem_cgroup_subsys, nextid,
734                                 &root->css, &found);
735                 if (css && css_tryget(css))
736                         iter = container_of(css, struct mem_cgroup, css);
737                 rcu_read_unlock();
738                 /* If css is NULL, no more cgroups will be found */
739                 nextid = found + 1;
740         } while (css && !iter);
741
742         return iter;
743 }
744 /*
745  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
746  * be careful that "break" loop is not allowed. We have reference count.
747  * Instead of that modify "cond" to be false and "continue" to exit the loop.
748  */
749 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
750         for (iter = mem_cgroup_start_loop(root);\
751              iter != NULL;\
752              iter = mem_cgroup_get_next(iter, root, cond))
753
754 #define for_each_mem_cgroup_tree(iter, root) \
755         for_each_mem_cgroup_tree_cond(iter, root, true)
756
757 #define for_each_mem_cgroup_all(iter) \
758         for_each_mem_cgroup_tree_cond(iter, NULL, true)
759
760
761 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
762 {
763         return (mem == root_mem_cgroup);
764 }
765
766 /*
767  * Following LRU functions are allowed to be used without PCG_LOCK.
768  * Operations are called by routine of global LRU independently from memcg.
769  * What we have to take care of here is validness of pc->mem_cgroup.
770  *
771  * Changes to pc->mem_cgroup happens when
772  * 1. charge
773  * 2. moving account
774  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
775  * It is added to LRU before charge.
776  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
777  * When moving account, the page is not on LRU. It's isolated.
778  */
779
780 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
781 {
782         struct page_cgroup *pc;
783         struct mem_cgroup_per_zone *mz;
784
785         if (mem_cgroup_disabled())
786                 return;
787         pc = lookup_page_cgroup(page);
788         /* can happen while we handle swapcache. */
789         if (!TestClearPageCgroupAcctLRU(pc))
790                 return;
791         VM_BUG_ON(!pc->mem_cgroup);
792         /*
793          * We don't check PCG_USED bit. It's cleared when the "page" is finally
794          * removed from global LRU.
795          */
796         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
797         /* huge page split is done under lru_lock. so, we have no races. */
798         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
799         if (mem_cgroup_is_root(pc->mem_cgroup))
800                 return;
801         VM_BUG_ON(list_empty(&pc->lru));
802         list_del_init(&pc->lru);
803 }
804
805 void mem_cgroup_del_lru(struct page *page)
806 {
807         mem_cgroup_del_lru_list(page, page_lru(page));
808 }
809
810 /*
811  * Writeback is about to end against a page which has been marked for immediate
812  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
813  * inactive list.
814  */
815 void mem_cgroup_rotate_reclaimable_page(struct page *page)
816 {
817         struct mem_cgroup_per_zone *mz;
818         struct page_cgroup *pc;
819         enum lru_list lru = page_lru(page);
820
821         if (mem_cgroup_disabled())
822                 return;
823
824         pc = lookup_page_cgroup(page);
825         /* unused or root page is not rotated. */
826         if (!PageCgroupUsed(pc))
827                 return;
828         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
829         smp_rmb();
830         if (mem_cgroup_is_root(pc->mem_cgroup))
831                 return;
832         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
833         list_move_tail(&pc->lru, &mz->lists[lru]);
834 }
835
836 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
837 {
838         struct mem_cgroup_per_zone *mz;
839         struct page_cgroup *pc;
840
841         if (mem_cgroup_disabled())
842                 return;
843
844         pc = lookup_page_cgroup(page);
845         /* unused or root page is not rotated. */
846         if (!PageCgroupUsed(pc))
847                 return;
848         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
849         smp_rmb();
850         if (mem_cgroup_is_root(pc->mem_cgroup))
851                 return;
852         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
853         list_move(&pc->lru, &mz->lists[lru]);
854 }
855
856 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
857 {
858         struct page_cgroup *pc;
859         struct mem_cgroup_per_zone *mz;
860
861         if (mem_cgroup_disabled())
862                 return;
863         pc = lookup_page_cgroup(page);
864         VM_BUG_ON(PageCgroupAcctLRU(pc));
865         if (!PageCgroupUsed(pc))
866                 return;
867         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
868         smp_rmb();
869         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
870         /* huge page split is done under lru_lock. so, we have no races. */
871         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
872         SetPageCgroupAcctLRU(pc);
873         if (mem_cgroup_is_root(pc->mem_cgroup))
874                 return;
875         list_add(&pc->lru, &mz->lists[lru]);
876 }
877
878 /*
879  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
880  * lru because the page may.be reused after it's fully uncharged (because of
881  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
882  * it again. This function is only used to charge SwapCache. It's done under
883  * lock_page and expected that zone->lru_lock is never held.
884  */
885 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
886 {
887         unsigned long flags;
888         struct zone *zone = page_zone(page);
889         struct page_cgroup *pc = lookup_page_cgroup(page);
890
891         spin_lock_irqsave(&zone->lru_lock, flags);
892         /*
893          * Forget old LRU when this page_cgroup is *not* used. This Used bit
894          * is guarded by lock_page() because the page is SwapCache.
895          */
896         if (!PageCgroupUsed(pc))
897                 mem_cgroup_del_lru_list(page, page_lru(page));
898         spin_unlock_irqrestore(&zone->lru_lock, flags);
899 }
900
901 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
902 {
903         unsigned long flags;
904         struct zone *zone = page_zone(page);
905         struct page_cgroup *pc = lookup_page_cgroup(page);
906
907         spin_lock_irqsave(&zone->lru_lock, flags);
908         /* link when the page is linked to LRU but page_cgroup isn't */
909         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
910                 mem_cgroup_add_lru_list(page, page_lru(page));
911         spin_unlock_irqrestore(&zone->lru_lock, flags);
912 }
913
914
915 void mem_cgroup_move_lists(struct page *page,
916                            enum lru_list from, enum lru_list to)
917 {
918         if (mem_cgroup_disabled())
919                 return;
920         mem_cgroup_del_lru_list(page, from);
921         mem_cgroup_add_lru_list(page, to);
922 }
923
924 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
925 {
926         int ret;
927         struct mem_cgroup *curr = NULL;
928         struct task_struct *p;
929
930         p = find_lock_task_mm(task);
931         if (!p)
932                 return 0;
933         curr = try_get_mem_cgroup_from_mm(p->mm);
934         task_unlock(p);
935         if (!curr)
936                 return 0;
937         /*
938          * We should check use_hierarchy of "mem" not "curr". Because checking
939          * use_hierarchy of "curr" here make this function true if hierarchy is
940          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
941          * hierarchy(even if use_hierarchy is disabled in "mem").
942          */
943         if (mem->use_hierarchy)
944                 ret = css_is_ancestor(&curr->css, &mem->css);
945         else
946                 ret = (curr == mem);
947         css_put(&curr->css);
948         return ret;
949 }
950
951 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
952 {
953         unsigned long active;
954         unsigned long inactive;
955         unsigned long gb;
956         unsigned long inactive_ratio;
957
958         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
959         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
960
961         gb = (inactive + active) >> (30 - PAGE_SHIFT);
962         if (gb)
963                 inactive_ratio = int_sqrt(10 * gb);
964         else
965                 inactive_ratio = 1;
966
967         if (present_pages) {
968                 present_pages[0] = inactive;
969                 present_pages[1] = active;
970         }
971
972         return inactive_ratio;
973 }
974
975 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
976 {
977         unsigned long active;
978         unsigned long inactive;
979         unsigned long present_pages[2];
980         unsigned long inactive_ratio;
981
982         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
983
984         inactive = present_pages[0];
985         active = present_pages[1];
986
987         if (inactive * inactive_ratio < active)
988                 return 1;
989
990         return 0;
991 }
992
993 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
994 {
995         unsigned long active;
996         unsigned long inactive;
997
998         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
999         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1000
1001         return (active > inactive);
1002 }
1003
1004 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
1005                                        struct zone *zone,
1006                                        enum lru_list lru)
1007 {
1008         int nid = zone_to_nid(zone);
1009         int zid = zone_idx(zone);
1010         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1011
1012         return MEM_CGROUP_ZSTAT(mz, lru);
1013 }
1014
1015 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1016                                                       struct zone *zone)
1017 {
1018         int nid = zone_to_nid(zone);
1019         int zid = zone_idx(zone);
1020         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1021
1022         return &mz->reclaim_stat;
1023 }
1024
1025 struct zone_reclaim_stat *
1026 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1027 {
1028         struct page_cgroup *pc;
1029         struct mem_cgroup_per_zone *mz;
1030
1031         if (mem_cgroup_disabled())
1032                 return NULL;
1033
1034         pc = lookup_page_cgroup(page);
1035         if (!PageCgroupUsed(pc))
1036                 return NULL;
1037         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1038         smp_rmb();
1039         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1040         return &mz->reclaim_stat;
1041 }
1042
1043 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1044                                         struct list_head *dst,
1045                                         unsigned long *scanned, int order,
1046                                         int mode, struct zone *z,
1047                                         struct mem_cgroup *mem_cont,
1048                                         int active, int file)
1049 {
1050         unsigned long nr_taken = 0;
1051         struct page *page;
1052         unsigned long scan;
1053         LIST_HEAD(pc_list);
1054         struct list_head *src;
1055         struct page_cgroup *pc, *tmp;
1056         int nid = zone_to_nid(z);
1057         int zid = zone_idx(z);
1058         struct mem_cgroup_per_zone *mz;
1059         int lru = LRU_FILE * file + active;
1060         int ret;
1061
1062         BUG_ON(!mem_cont);
1063         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1064         src = &mz->lists[lru];
1065
1066         scan = 0;
1067         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1068                 if (scan >= nr_to_scan)
1069                         break;
1070
1071                 if (unlikely(!PageCgroupUsed(pc)))
1072                         continue;
1073
1074                 page = lookup_cgroup_page(pc);
1075
1076                 if (unlikely(!PageLRU(page)))
1077                         continue;
1078
1079                 scan++;
1080                 ret = __isolate_lru_page(page, mode, file);
1081                 switch (ret) {
1082                 case 0:
1083                         list_move(&page->lru, dst);
1084                         mem_cgroup_del_lru(page);
1085                         nr_taken += hpage_nr_pages(page);
1086                         break;
1087                 case -EBUSY:
1088                         /* we don't affect global LRU but rotate in our LRU */
1089                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1090                         break;
1091                 default:
1092                         break;
1093                 }
1094         }
1095
1096         *scanned = scan;
1097
1098         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1099                                       0, 0, 0, mode);
1100
1101         return nr_taken;
1102 }
1103
1104 #define mem_cgroup_from_res_counter(counter, member)    \
1105         container_of(counter, struct mem_cgroup, member)
1106
1107 /**
1108  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1109  * @mem: the memory cgroup
1110  *
1111  * Returns the maximum amount of memory @mem can be charged with, in
1112  * bytes.
1113  */
1114 static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem)
1115 {
1116         unsigned long long margin;
1117
1118         margin = res_counter_margin(&mem->res);
1119         if (do_swap_account)
1120                 margin = min(margin, res_counter_margin(&mem->memsw));
1121         return margin;
1122 }
1123
1124 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1125 {
1126         struct cgroup *cgrp = memcg->css.cgroup;
1127
1128         /* root ? */
1129         if (cgrp->parent == NULL)
1130                 return vm_swappiness;
1131
1132         return memcg->swappiness;
1133 }
1134
1135 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1136 {
1137         int cpu;
1138
1139         get_online_cpus();
1140         spin_lock(&mem->pcp_counter_lock);
1141         for_each_online_cpu(cpu)
1142                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1143         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1144         spin_unlock(&mem->pcp_counter_lock);
1145         put_online_cpus();
1146
1147         synchronize_rcu();
1148 }
1149
1150 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1151 {
1152         int cpu;
1153
1154         if (!mem)
1155                 return;
1156         get_online_cpus();
1157         spin_lock(&mem->pcp_counter_lock);
1158         for_each_online_cpu(cpu)
1159                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1160         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1161         spin_unlock(&mem->pcp_counter_lock);
1162         put_online_cpus();
1163 }
1164 /*
1165  * 2 routines for checking "mem" is under move_account() or not.
1166  *
1167  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1168  *                        for avoiding race in accounting. If true,
1169  *                        pc->mem_cgroup may be overwritten.
1170  *
1171  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1172  *                        under hierarchy of moving cgroups. This is for
1173  *                        waiting at hith-memory prressure caused by "move".
1174  */
1175
1176 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1177 {
1178         VM_BUG_ON(!rcu_read_lock_held());
1179         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1180 }
1181
1182 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1183 {
1184         struct mem_cgroup *from;
1185         struct mem_cgroup *to;
1186         bool ret = false;
1187         /*
1188          * Unlike task_move routines, we access mc.to, mc.from not under
1189          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1190          */
1191         spin_lock(&mc.lock);
1192         from = mc.from;
1193         to = mc.to;
1194         if (!from)
1195                 goto unlock;
1196         if (from == mem || to == mem
1197             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1198             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1199                 ret = true;
1200 unlock:
1201         spin_unlock(&mc.lock);
1202         return ret;
1203 }
1204
1205 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1206 {
1207         if (mc.moving_task && current != mc.moving_task) {
1208                 if (mem_cgroup_under_move(mem)) {
1209                         DEFINE_WAIT(wait);
1210                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1211                         /* moving charge context might have finished. */
1212                         if (mc.moving_task)
1213                                 schedule();
1214                         finish_wait(&mc.waitq, &wait);
1215                         return true;
1216                 }
1217         }
1218         return false;
1219 }
1220
1221 /**
1222  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1223  * @memcg: The memory cgroup that went over limit
1224  * @p: Task that is going to be killed
1225  *
1226  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1227  * enabled
1228  */
1229 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1230 {
1231         struct cgroup *task_cgrp;
1232         struct cgroup *mem_cgrp;
1233         /*
1234          * Need a buffer in BSS, can't rely on allocations. The code relies
1235          * on the assumption that OOM is serialized for memory controller.
1236          * If this assumption is broken, revisit this code.
1237          */
1238         static char memcg_name[PATH_MAX];
1239         int ret;
1240
1241         if (!memcg || !p)
1242                 return;
1243
1244
1245         rcu_read_lock();
1246
1247         mem_cgrp = memcg->css.cgroup;
1248         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1249
1250         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1251         if (ret < 0) {
1252                 /*
1253                  * Unfortunately, we are unable to convert to a useful name
1254                  * But we'll still print out the usage information
1255                  */
1256                 rcu_read_unlock();
1257                 goto done;
1258         }
1259         rcu_read_unlock();
1260
1261         printk(KERN_INFO "Task in %s killed", memcg_name);
1262
1263         rcu_read_lock();
1264         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1265         if (ret < 0) {
1266                 rcu_read_unlock();
1267                 goto done;
1268         }
1269         rcu_read_unlock();
1270
1271         /*
1272          * Continues from above, so we don't need an KERN_ level
1273          */
1274         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1275 done:
1276
1277         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1278                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1279                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1280                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1281         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1282                 "failcnt %llu\n",
1283                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1284                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1285                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1286 }
1287
1288 /*
1289  * This function returns the number of memcg under hierarchy tree. Returns
1290  * 1(self count) if no children.
1291  */
1292 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1293 {
1294         int num = 0;
1295         struct mem_cgroup *iter;
1296
1297         for_each_mem_cgroup_tree(iter, mem)
1298                 num++;
1299         return num;
1300 }
1301
1302 /*
1303  * Return the memory (and swap, if configured) limit for a memcg.
1304  */
1305 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1306 {
1307         u64 limit;
1308         u64 memsw;
1309
1310         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1311         limit += total_swap_pages << PAGE_SHIFT;
1312
1313         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1314         /*
1315          * If memsw is finite and limits the amount of swap space available
1316          * to this memcg, return that limit.
1317          */
1318         return min(limit, memsw);
1319 }
1320
1321 /*
1322  * Visit the first child (need not be the first child as per the ordering
1323  * of the cgroup list, since we track last_scanned_child) of @mem and use
1324  * that to reclaim free pages from.
1325  */
1326 static struct mem_cgroup *
1327 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1328 {
1329         struct mem_cgroup *ret = NULL;
1330         struct cgroup_subsys_state *css;
1331         int nextid, found;
1332
1333         if (!root_mem->use_hierarchy) {
1334                 css_get(&root_mem->css);
1335                 ret = root_mem;
1336         }
1337
1338         while (!ret) {
1339                 rcu_read_lock();
1340                 nextid = root_mem->last_scanned_child + 1;
1341                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1342                                    &found);
1343                 if (css && css_tryget(css))
1344                         ret = container_of(css, struct mem_cgroup, css);
1345
1346                 rcu_read_unlock();
1347                 /* Updates scanning parameter */
1348                 if (!css) {
1349                         /* this means start scan from ID:1 */
1350                         root_mem->last_scanned_child = 0;
1351                 } else
1352                         root_mem->last_scanned_child = found;
1353         }
1354
1355         return ret;
1356 }
1357
1358 /*
1359  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1360  * we reclaimed from, so that we don't end up penalizing one child extensively
1361  * based on its position in the children list.
1362  *
1363  * root_mem is the original ancestor that we've been reclaim from.
1364  *
1365  * We give up and return to the caller when we visit root_mem twice.
1366  * (other groups can be removed while we're walking....)
1367  *
1368  * If shrink==true, for avoiding to free too much, this returns immedieately.
1369  */
1370 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1371                                                 struct zone *zone,
1372                                                 gfp_t gfp_mask,
1373                                                 unsigned long reclaim_options)
1374 {
1375         struct mem_cgroup *victim;
1376         int ret, total = 0;
1377         int loop = 0;
1378         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1379         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1380         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1381         unsigned long excess;
1382
1383         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1384
1385         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1386         if (root_mem->memsw_is_minimum)
1387                 noswap = true;
1388
1389         while (1) {
1390                 victim = mem_cgroup_select_victim(root_mem);
1391                 if (victim == root_mem) {
1392                         loop++;
1393                         if (loop >= 1)
1394                                 drain_all_stock_async();
1395                         if (loop >= 2) {
1396                                 /*
1397                                  * If we have not been able to reclaim
1398                                  * anything, it might because there are
1399                                  * no reclaimable pages under this hierarchy
1400                                  */
1401                                 if (!check_soft || !total) {
1402                                         css_put(&victim->css);
1403                                         break;
1404                                 }
1405                                 /*
1406                                  * We want to do more targetted reclaim.
1407                                  * excess >> 2 is not to excessive so as to
1408                                  * reclaim too much, nor too less that we keep
1409                                  * coming back to reclaim from this cgroup
1410                                  */
1411                                 if (total >= (excess >> 2) ||
1412                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1413                                         css_put(&victim->css);
1414                                         break;
1415                                 }
1416                         }
1417                 }
1418                 if (!mem_cgroup_local_usage(victim)) {
1419                         /* this cgroup's local usage == 0 */
1420                         css_put(&victim->css);
1421                         continue;
1422                 }
1423                 /* we use swappiness of local cgroup */
1424                 if (check_soft)
1425                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1426                                 noswap, get_swappiness(victim), zone);
1427                 else
1428                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1429                                                 noswap, get_swappiness(victim));
1430                 css_put(&victim->css);
1431                 /*
1432                  * At shrinking usage, we can't check we should stop here or
1433                  * reclaim more. It's depends on callers. last_scanned_child
1434                  * will work enough for keeping fairness under tree.
1435                  */
1436                 if (shrink)
1437                         return ret;
1438                 total += ret;
1439                 if (check_soft) {
1440                         if (!res_counter_soft_limit_excess(&root_mem->res))
1441                                 return total;
1442                 } else if (mem_cgroup_margin(root_mem))
1443                         return 1 + total;
1444         }
1445         return total;
1446 }
1447
1448 /*
1449  * Check OOM-Killer is already running under our hierarchy.
1450  * If someone is running, return false.
1451  */
1452 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1453 {
1454         int x, lock_count = 0;
1455         struct mem_cgroup *iter;
1456
1457         for_each_mem_cgroup_tree(iter, mem) {
1458                 x = atomic_inc_return(&iter->oom_lock);
1459                 lock_count = max(x, lock_count);
1460         }
1461
1462         if (lock_count == 1)
1463                 return true;
1464         return false;
1465 }
1466
1467 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1468 {
1469         struct mem_cgroup *iter;
1470
1471         /*
1472          * When a new child is created while the hierarchy is under oom,
1473          * mem_cgroup_oom_lock() may not be called. We have to use
1474          * atomic_add_unless() here.
1475          */
1476         for_each_mem_cgroup_tree(iter, mem)
1477                 atomic_add_unless(&iter->oom_lock, -1, 0);
1478         return 0;
1479 }
1480
1481
1482 static DEFINE_MUTEX(memcg_oom_mutex);
1483 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1484
1485 struct oom_wait_info {
1486         struct mem_cgroup *mem;
1487         wait_queue_t    wait;
1488 };
1489
1490 static int memcg_oom_wake_function(wait_queue_t *wait,
1491         unsigned mode, int sync, void *arg)
1492 {
1493         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1494         struct oom_wait_info *oom_wait_info;
1495
1496         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1497
1498         if (oom_wait_info->mem == wake_mem)
1499                 goto wakeup;
1500         /* if no hierarchy, no match */
1501         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1502                 return 0;
1503         /*
1504          * Both of oom_wait_info->mem and wake_mem are stable under us.
1505          * Then we can use css_is_ancestor without taking care of RCU.
1506          */
1507         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1508             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1509                 return 0;
1510
1511 wakeup:
1512         return autoremove_wake_function(wait, mode, sync, arg);
1513 }
1514
1515 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1516 {
1517         /* for filtering, pass "mem" as argument. */
1518         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1519 }
1520
1521 static void memcg_oom_recover(struct mem_cgroup *mem)
1522 {
1523         if (mem && atomic_read(&mem->oom_lock))
1524                 memcg_wakeup_oom(mem);
1525 }
1526
1527 /*
1528  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1529  */
1530 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1531 {
1532         struct oom_wait_info owait;
1533         bool locked, need_to_kill;
1534
1535         owait.mem = mem;
1536         owait.wait.flags = 0;
1537         owait.wait.func = memcg_oom_wake_function;
1538         owait.wait.private = current;
1539         INIT_LIST_HEAD(&owait.wait.task_list);
1540         need_to_kill = true;
1541         /* At first, try to OOM lock hierarchy under mem.*/
1542         mutex_lock(&memcg_oom_mutex);
1543         locked = mem_cgroup_oom_lock(mem);
1544         /*
1545          * Even if signal_pending(), we can't quit charge() loop without
1546          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1547          * under OOM is always welcomed, use TASK_KILLABLE here.
1548          */
1549         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1550         if (!locked || mem->oom_kill_disable)
1551                 need_to_kill = false;
1552         if (locked)
1553                 mem_cgroup_oom_notify(mem);
1554         mutex_unlock(&memcg_oom_mutex);
1555
1556         if (need_to_kill) {
1557                 finish_wait(&memcg_oom_waitq, &owait.wait);
1558                 mem_cgroup_out_of_memory(mem, mask);
1559         } else {
1560                 schedule();
1561                 finish_wait(&memcg_oom_waitq, &owait.wait);
1562         }
1563         mutex_lock(&memcg_oom_mutex);
1564         mem_cgroup_oom_unlock(mem);
1565         memcg_wakeup_oom(mem);
1566         mutex_unlock(&memcg_oom_mutex);
1567
1568         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1569                 return false;
1570         /* Give chance to dying process */
1571         schedule_timeout(1);
1572         return true;
1573 }
1574
1575 /*
1576  * Currently used to update mapped file statistics, but the routine can be
1577  * generalized to update other statistics as well.
1578  *
1579  * Notes: Race condition
1580  *
1581  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1582  * it tends to be costly. But considering some conditions, we doesn't need
1583  * to do so _always_.
1584  *
1585  * Considering "charge", lock_page_cgroup() is not required because all
1586  * file-stat operations happen after a page is attached to radix-tree. There
1587  * are no race with "charge".
1588  *
1589  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1590  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1591  * if there are race with "uncharge". Statistics itself is properly handled
1592  * by flags.
1593  *
1594  * Considering "move", this is an only case we see a race. To make the race
1595  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1596  * possibility of race condition. If there is, we take a lock.
1597  */
1598
1599 void mem_cgroup_update_page_stat(struct page *page,
1600                                  enum mem_cgroup_page_stat_item idx, int val)
1601 {
1602         struct mem_cgroup *mem;
1603         struct page_cgroup *pc = lookup_page_cgroup(page);
1604         bool need_unlock = false;
1605         unsigned long uninitialized_var(flags);
1606
1607         if (unlikely(!pc))
1608                 return;
1609
1610         rcu_read_lock();
1611         mem = pc->mem_cgroup;
1612         if (unlikely(!mem || !PageCgroupUsed(pc)))
1613                 goto out;
1614         /* pc->mem_cgroup is unstable ? */
1615         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1616                 /* take a lock against to access pc->mem_cgroup */
1617                 move_lock_page_cgroup(pc, &flags);
1618                 need_unlock = true;
1619                 mem = pc->mem_cgroup;
1620                 if (!mem || !PageCgroupUsed(pc))
1621                         goto out;
1622         }
1623
1624         switch (idx) {
1625         case MEMCG_NR_FILE_MAPPED:
1626                 if (val > 0)
1627                         SetPageCgroupFileMapped(pc);
1628                 else if (!page_mapped(page))
1629                         ClearPageCgroupFileMapped(pc);
1630                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1631                 break;
1632         default:
1633                 BUG();
1634         }
1635
1636         this_cpu_add(mem->stat->count[idx], val);
1637
1638 out:
1639         if (unlikely(need_unlock))
1640                 move_unlock_page_cgroup(pc, &flags);
1641         rcu_read_unlock();
1642         return;
1643 }
1644 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1645
1646 /*
1647  * size of first charge trial. "32" comes from vmscan.c's magic value.
1648  * TODO: maybe necessary to use big numbers in big irons.
1649  */
1650 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1651 struct memcg_stock_pcp {
1652         struct mem_cgroup *cached; /* this never be root cgroup */
1653         int charge;
1654         struct work_struct work;
1655 };
1656 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1657 static atomic_t memcg_drain_count;
1658
1659 /*
1660  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1661  * from local stock and true is returned. If the stock is 0 or charges from a
1662  * cgroup which is not current target, returns false. This stock will be
1663  * refilled.
1664  */
1665 static bool consume_stock(struct mem_cgroup *mem)
1666 {
1667         struct memcg_stock_pcp *stock;
1668         bool ret = true;
1669
1670         stock = &get_cpu_var(memcg_stock);
1671         if (mem == stock->cached && stock->charge)
1672                 stock->charge -= PAGE_SIZE;
1673         else /* need to call res_counter_charge */
1674                 ret = false;
1675         put_cpu_var(memcg_stock);
1676         return ret;
1677 }
1678
1679 /*
1680  * Returns stocks cached in percpu to res_counter and reset cached information.
1681  */
1682 static void drain_stock(struct memcg_stock_pcp *stock)
1683 {
1684         struct mem_cgroup *old = stock->cached;
1685
1686         if (stock->charge) {
1687                 res_counter_uncharge(&old->res, stock->charge);
1688                 if (do_swap_account)
1689                         res_counter_uncharge(&old->memsw, stock->charge);
1690         }
1691         stock->cached = NULL;
1692         stock->charge = 0;
1693 }
1694
1695 /*
1696  * This must be called under preempt disabled or must be called by
1697  * a thread which is pinned to local cpu.
1698  */
1699 static void drain_local_stock(struct work_struct *dummy)
1700 {
1701         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1702         drain_stock(stock);
1703 }
1704
1705 /*
1706  * Cache charges(val) which is from res_counter, to local per_cpu area.
1707  * This will be consumed by consume_stock() function, later.
1708  */
1709 static void refill_stock(struct mem_cgroup *mem, int val)
1710 {
1711         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1712
1713         if (stock->cached != mem) { /* reset if necessary */
1714                 drain_stock(stock);
1715                 stock->cached = mem;
1716         }
1717         stock->charge += val;
1718         put_cpu_var(memcg_stock);
1719 }
1720
1721 /*
1722  * Tries to drain stocked charges in other cpus. This function is asynchronous
1723  * and just put a work per cpu for draining localy on each cpu. Caller can
1724  * expects some charges will be back to res_counter later but cannot wait for
1725  * it.
1726  */
1727 static void drain_all_stock_async(void)
1728 {
1729         int cpu;
1730         /* This function is for scheduling "drain" in asynchronous way.
1731          * The result of "drain" is not directly handled by callers. Then,
1732          * if someone is calling drain, we don't have to call drain more.
1733          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1734          * there is a race. We just do loose check here.
1735          */
1736         if (atomic_read(&memcg_drain_count))
1737                 return;
1738         /* Notify other cpus that system-wide "drain" is running */
1739         atomic_inc(&memcg_drain_count);
1740         get_online_cpus();
1741         for_each_online_cpu(cpu) {
1742                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1743                 schedule_work_on(cpu, &stock->work);
1744         }
1745         put_online_cpus();
1746         atomic_dec(&memcg_drain_count);
1747         /* We don't wait for flush_work */
1748 }
1749
1750 /* This is a synchronous drain interface. */
1751 static void drain_all_stock_sync(void)
1752 {
1753         /* called when force_empty is called */
1754         atomic_inc(&memcg_drain_count);
1755         schedule_on_each_cpu(drain_local_stock);
1756         atomic_dec(&memcg_drain_count);
1757 }
1758
1759 /*
1760  * This function drains percpu counter value from DEAD cpu and
1761  * move it to local cpu. Note that this function can be preempted.
1762  */
1763 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1764 {
1765         int i;
1766
1767         spin_lock(&mem->pcp_counter_lock);
1768         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1769                 s64 x = per_cpu(mem->stat->count[i], cpu);
1770
1771                 per_cpu(mem->stat->count[i], cpu) = 0;
1772                 mem->nocpu_base.count[i] += x;
1773         }
1774         /* need to clear ON_MOVE value, works as a kind of lock. */
1775         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1776         spin_unlock(&mem->pcp_counter_lock);
1777 }
1778
1779 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1780 {
1781         int idx = MEM_CGROUP_ON_MOVE;
1782
1783         spin_lock(&mem->pcp_counter_lock);
1784         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1785         spin_unlock(&mem->pcp_counter_lock);
1786 }
1787
1788 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1789                                         unsigned long action,
1790                                         void *hcpu)
1791 {
1792         int cpu = (unsigned long)hcpu;
1793         struct memcg_stock_pcp *stock;
1794         struct mem_cgroup *iter;
1795
1796         if ((action == CPU_ONLINE)) {
1797                 for_each_mem_cgroup_all(iter)
1798                         synchronize_mem_cgroup_on_move(iter, cpu);
1799                 return NOTIFY_OK;
1800         }
1801
1802         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1803                 return NOTIFY_OK;
1804
1805         for_each_mem_cgroup_all(iter)
1806                 mem_cgroup_drain_pcp_counter(iter, cpu);
1807
1808         stock = &per_cpu(memcg_stock, cpu);
1809         drain_stock(stock);
1810         return NOTIFY_OK;
1811 }
1812
1813
1814 /* See __mem_cgroup_try_charge() for details */
1815 enum {
1816         CHARGE_OK,              /* success */
1817         CHARGE_RETRY,           /* need to retry but retry is not bad */
1818         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1819         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1820         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1821 };
1822
1823 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1824                                 int csize, bool oom_check)
1825 {
1826         struct mem_cgroup *mem_over_limit;
1827         struct res_counter *fail_res;
1828         unsigned long flags = 0;
1829         int ret;
1830
1831         ret = res_counter_charge(&mem->res, csize, &fail_res);
1832
1833         if (likely(!ret)) {
1834                 if (!do_swap_account)
1835                         return CHARGE_OK;
1836                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1837                 if (likely(!ret))
1838                         return CHARGE_OK;
1839
1840                 res_counter_uncharge(&mem->res, csize);
1841                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1842                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1843         } else
1844                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1845         /*
1846          * csize can be either a huge page (HPAGE_SIZE), a batch of
1847          * regular pages (CHARGE_SIZE), or a single regular page
1848          * (PAGE_SIZE).
1849          *
1850          * Never reclaim on behalf of optional batching, retry with a
1851          * single page instead.
1852          */
1853         if (csize == CHARGE_SIZE)
1854                 return CHARGE_RETRY;
1855
1856         if (!(gfp_mask & __GFP_WAIT))
1857                 return CHARGE_WOULDBLOCK;
1858
1859         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1860                                               gfp_mask, flags);
1861         if (mem_cgroup_margin(mem_over_limit) >= csize)
1862                 return CHARGE_RETRY;
1863         /*
1864          * Even though the limit is exceeded at this point, reclaim
1865          * may have been able to free some pages.  Retry the charge
1866          * before killing the task.
1867          *
1868          * Only for regular pages, though: huge pages are rather
1869          * unlikely to succeed so close to the limit, and we fall back
1870          * to regular pages anyway in case of failure.
1871          */
1872         if (csize == PAGE_SIZE && ret)
1873                 return CHARGE_RETRY;
1874
1875         /*
1876          * At task move, charge accounts can be doubly counted. So, it's
1877          * better to wait until the end of task_move if something is going on.
1878          */
1879         if (mem_cgroup_wait_acct_move(mem_over_limit))
1880                 return CHARGE_RETRY;
1881
1882         /* If we don't need to call oom-killer at el, return immediately */
1883         if (!oom_check)
1884                 return CHARGE_NOMEM;
1885         /* check OOM */
1886         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1887                 return CHARGE_OOM_DIE;
1888
1889         return CHARGE_RETRY;
1890 }
1891
1892 /*
1893  * Unlike exported interface, "oom" parameter is added. if oom==true,
1894  * oom-killer can be invoked.
1895  */
1896 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1897                                    gfp_t gfp_mask,
1898                                    struct mem_cgroup **memcg, bool oom,
1899                                    int page_size)
1900 {
1901         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1902         struct mem_cgroup *mem = NULL;
1903         int ret;
1904         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1905
1906         /*
1907          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1908          * in system level. So, allow to go ahead dying process in addition to
1909          * MEMDIE process.
1910          */
1911         if (unlikely(test_thread_flag(TIF_MEMDIE)
1912                      || fatal_signal_pending(current)))
1913                 goto bypass;
1914
1915         /*
1916          * We always charge the cgroup the mm_struct belongs to.
1917          * The mm_struct's mem_cgroup changes on task migration if the
1918          * thread group leader migrates. It's possible that mm is not
1919          * set, if so charge the init_mm (happens for pagecache usage).
1920          */
1921         if (!*memcg && !mm)
1922                 goto bypass;
1923 again:
1924         if (*memcg) { /* css should be a valid one */
1925                 mem = *memcg;
1926                 VM_BUG_ON(css_is_removed(&mem->css));
1927                 if (mem_cgroup_is_root(mem))
1928                         goto done;
1929                 if (page_size == PAGE_SIZE && consume_stock(mem))
1930                         goto done;
1931                 css_get(&mem->css);
1932         } else {
1933                 struct task_struct *p;
1934
1935                 rcu_read_lock();
1936                 p = rcu_dereference(mm->owner);
1937                 /*
1938                  * Because we don't have task_lock(), "p" can exit.
1939                  * In that case, "mem" can point to root or p can be NULL with
1940                  * race with swapoff. Then, we have small risk of mis-accouning.
1941                  * But such kind of mis-account by race always happens because
1942                  * we don't have cgroup_mutex(). It's overkill and we allo that
1943                  * small race, here.
1944                  * (*) swapoff at el will charge against mm-struct not against
1945                  * task-struct. So, mm->owner can be NULL.
1946                  */
1947                 mem = mem_cgroup_from_task(p);
1948                 if (!mem || mem_cgroup_is_root(mem)) {
1949                         rcu_read_unlock();
1950                         goto done;
1951                 }
1952                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1953                         /*
1954                          * It seems dagerous to access memcg without css_get().
1955                          * But considering how consume_stok works, it's not
1956                          * necessary. If consume_stock success, some charges
1957                          * from this memcg are cached on this cpu. So, we
1958                          * don't need to call css_get()/css_tryget() before
1959                          * calling consume_stock().
1960                          */
1961                         rcu_read_unlock();
1962                         goto done;
1963                 }
1964                 /* after here, we may be blocked. we need to get refcnt */
1965                 if (!css_tryget(&mem->css)) {
1966                         rcu_read_unlock();
1967                         goto again;
1968                 }
1969                 rcu_read_unlock();
1970         }
1971
1972         do {
1973                 bool oom_check;
1974
1975                 /* If killed, bypass charge */
1976                 if (fatal_signal_pending(current)) {
1977                         css_put(&mem->css);
1978                         goto bypass;
1979                 }
1980
1981                 oom_check = false;
1982                 if (oom && !nr_oom_retries) {
1983                         oom_check = true;
1984                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1985                 }
1986
1987                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1988
1989                 switch (ret) {
1990                 case CHARGE_OK:
1991                         break;
1992                 case CHARGE_RETRY: /* not in OOM situation but retry */
1993                         csize = page_size;
1994                         css_put(&mem->css);
1995                         mem = NULL;
1996                         goto again;
1997                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1998                         css_put(&mem->css);
1999                         goto nomem;
2000                 case CHARGE_NOMEM: /* OOM routine works */
2001                         if (!oom) {
2002                                 css_put(&mem->css);
2003                                 goto nomem;
2004                         }
2005                         /* If oom, we never return -ENOMEM */
2006                         nr_oom_retries--;
2007                         break;
2008                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2009                         css_put(&mem->css);
2010                         goto bypass;
2011                 }
2012         } while (ret != CHARGE_OK);
2013
2014         if (csize > page_size)
2015                 refill_stock(mem, csize - page_size);
2016         css_put(&mem->css);
2017 done:
2018         *memcg = mem;
2019         return 0;
2020 nomem:
2021         *memcg = NULL;
2022         return -ENOMEM;
2023 bypass:
2024         *memcg = NULL;
2025         return 0;
2026 }
2027
2028 /*
2029  * Somemtimes we have to undo a charge we got by try_charge().
2030  * This function is for that and do uncharge, put css's refcnt.
2031  * gotten by try_charge().
2032  */
2033 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2034                                        unsigned int nr_pages)
2035 {
2036         if (!mem_cgroup_is_root(mem)) {
2037                 unsigned long bytes = nr_pages * PAGE_SIZE;
2038
2039                 res_counter_uncharge(&mem->res, bytes);
2040                 if (do_swap_account)
2041                         res_counter_uncharge(&mem->memsw, bytes);
2042         }
2043 }
2044
2045 /*
2046  * A helper function to get mem_cgroup from ID. must be called under
2047  * rcu_read_lock(). The caller must check css_is_removed() or some if
2048  * it's concern. (dropping refcnt from swap can be called against removed
2049  * memcg.)
2050  */
2051 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2052 {
2053         struct cgroup_subsys_state *css;
2054
2055         /* ID 0 is unused ID */
2056         if (!id)
2057                 return NULL;
2058         css = css_lookup(&mem_cgroup_subsys, id);
2059         if (!css)
2060                 return NULL;
2061         return container_of(css, struct mem_cgroup, css);
2062 }
2063
2064 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2065 {
2066         struct mem_cgroup *mem = NULL;
2067         struct page_cgroup *pc;
2068         unsigned short id;
2069         swp_entry_t ent;
2070
2071         VM_BUG_ON(!PageLocked(page));
2072
2073         pc = lookup_page_cgroup(page);
2074         lock_page_cgroup(pc);
2075         if (PageCgroupUsed(pc)) {
2076                 mem = pc->mem_cgroup;
2077                 if (mem && !css_tryget(&mem->css))
2078                         mem = NULL;
2079         } else if (PageSwapCache(page)) {
2080                 ent.val = page_private(page);
2081                 id = lookup_swap_cgroup(ent);
2082                 rcu_read_lock();
2083                 mem = mem_cgroup_lookup(id);
2084                 if (mem && !css_tryget(&mem->css))
2085                         mem = NULL;
2086                 rcu_read_unlock();
2087         }
2088         unlock_page_cgroup(pc);
2089         return mem;
2090 }
2091
2092 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2093                                        struct page *page,
2094                                        struct page_cgroup *pc,
2095                                        enum charge_type ctype,
2096                                        int page_size)
2097 {
2098         int nr_pages = page_size >> PAGE_SHIFT;
2099
2100         lock_page_cgroup(pc);
2101         if (unlikely(PageCgroupUsed(pc))) {
2102                 unlock_page_cgroup(pc);
2103                 __mem_cgroup_cancel_charge(mem, nr_pages);
2104                 return;
2105         }
2106         /*
2107          * we don't need page_cgroup_lock about tail pages, becase they are not
2108          * accessed by any other context at this point.
2109          */
2110         pc->mem_cgroup = mem;
2111         /*
2112          * We access a page_cgroup asynchronously without lock_page_cgroup().
2113          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2114          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2115          * before USED bit, we need memory barrier here.
2116          * See mem_cgroup_add_lru_list(), etc.
2117          */
2118         smp_wmb();
2119         switch (ctype) {
2120         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2121         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2122                 SetPageCgroupCache(pc);
2123                 SetPageCgroupUsed(pc);
2124                 break;
2125         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2126                 ClearPageCgroupCache(pc);
2127                 SetPageCgroupUsed(pc);
2128                 break;
2129         default:
2130                 break;
2131         }
2132
2133         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2134         unlock_page_cgroup(pc);
2135         /*
2136          * "charge_statistics" updated event counter. Then, check it.
2137          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2138          * if they exceeds softlimit.
2139          */
2140         memcg_check_events(mem, page);
2141 }
2142
2143 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2144
2145 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2146                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2147 /*
2148  * Because tail pages are not marked as "used", set it. We're under
2149  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2150  */
2151 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2152 {
2153         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2154         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2155         unsigned long flags;
2156
2157         if (mem_cgroup_disabled())
2158                 return;
2159         /*
2160          * We have no races with charge/uncharge but will have races with
2161          * page state accounting.
2162          */
2163         move_lock_page_cgroup(head_pc, &flags);
2164
2165         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2166         smp_wmb(); /* see __commit_charge() */
2167         if (PageCgroupAcctLRU(head_pc)) {
2168                 enum lru_list lru;
2169                 struct mem_cgroup_per_zone *mz;
2170
2171                 /*
2172                  * LRU flags cannot be copied because we need to add tail
2173                  *.page to LRU by generic call and our hook will be called.
2174                  * We hold lru_lock, then, reduce counter directly.
2175                  */
2176                 lru = page_lru(head);
2177                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2178                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2179         }
2180         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2181         move_unlock_page_cgroup(head_pc, &flags);
2182 }
2183 #endif
2184
2185 /**
2186  * mem_cgroup_move_account - move account of the page
2187  * @page: the page
2188  * @pc: page_cgroup of the page.
2189  * @from: mem_cgroup which the page is moved from.
2190  * @to: mem_cgroup which the page is moved to. @from != @to.
2191  * @uncharge: whether we should call uncharge and css_put against @from.
2192  * @charge_size: number of bytes to charge (regular or huge page)
2193  *
2194  * The caller must confirm following.
2195  * - page is not on LRU (isolate_page() is useful.)
2196  * - compound_lock is held when charge_size > PAGE_SIZE
2197  *
2198  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2199  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2200  * true, this function does "uncharge" from old cgroup, but it doesn't if
2201  * @uncharge is false, so a caller should do "uncharge".
2202  */
2203 static int mem_cgroup_move_account(struct page *page, struct page_cgroup *pc,
2204                                    struct mem_cgroup *from, struct mem_cgroup *to,
2205                                    bool uncharge, int charge_size)
2206 {
2207         int nr_pages = charge_size >> PAGE_SHIFT;
2208         unsigned long flags;
2209         int ret;
2210
2211         VM_BUG_ON(from == to);
2212         VM_BUG_ON(PageLRU(page));
2213         /*
2214          * The page is isolated from LRU. So, collapse function
2215          * will not handle this page. But page splitting can happen.
2216          * Do this check under compound_page_lock(). The caller should
2217          * hold it.
2218          */
2219         ret = -EBUSY;
2220         if (charge_size > PAGE_SIZE && !PageTransHuge(page))
2221                 goto out;
2222
2223         lock_page_cgroup(pc);
2224
2225         ret = -EINVAL;
2226         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2227                 goto unlock;
2228
2229         move_lock_page_cgroup(pc, &flags);
2230
2231         if (PageCgroupFileMapped(pc)) {
2232                 /* Update mapped_file data for mem_cgroup */
2233                 preempt_disable();
2234                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2235                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2236                 preempt_enable();
2237         }
2238         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2239         if (uncharge)
2240                 /* This is not "cancel", but cancel_charge does all we need. */
2241                 __mem_cgroup_cancel_charge(from, nr_pages);
2242
2243         /* caller should have done css_get */
2244         pc->mem_cgroup = to;
2245         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2246         /*
2247          * We charges against "to" which may not have any tasks. Then, "to"
2248          * can be under rmdir(). But in current implementation, caller of
2249          * this function is just force_empty() and move charge, so it's
2250          * garanteed that "to" is never removed. So, we don't check rmdir
2251          * status here.
2252          */
2253         move_unlock_page_cgroup(pc, &flags);
2254         ret = 0;
2255 unlock:
2256         unlock_page_cgroup(pc);
2257         /*
2258          * check events
2259          */
2260         memcg_check_events(to, page);
2261         memcg_check_events(from, page);
2262 out:
2263         return ret;
2264 }
2265
2266 /*
2267  * move charges to its parent.
2268  */
2269
2270 static int mem_cgroup_move_parent(struct page *page,
2271                                   struct page_cgroup *pc,
2272                                   struct mem_cgroup *child,
2273                                   gfp_t gfp_mask)
2274 {
2275         struct cgroup *cg = child->css.cgroup;
2276         struct cgroup *pcg = cg->parent;
2277         struct mem_cgroup *parent;
2278         int page_size = PAGE_SIZE;
2279         unsigned long flags;
2280         int ret;
2281
2282         /* Is ROOT ? */
2283         if (!pcg)
2284                 return -EINVAL;
2285
2286         ret = -EBUSY;
2287         if (!get_page_unless_zero(page))
2288                 goto out;
2289         if (isolate_lru_page(page))
2290                 goto put;
2291
2292         if (PageTransHuge(page))
2293                 page_size = HPAGE_SIZE;
2294
2295         parent = mem_cgroup_from_cont(pcg);
2296         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2297                                 &parent, false, page_size);
2298         if (ret || !parent)
2299                 goto put_back;
2300
2301         if (page_size > PAGE_SIZE)
2302                 flags = compound_lock_irqsave(page);
2303
2304         ret = mem_cgroup_move_account(page, pc, child, parent, true, page_size);
2305         if (ret)
2306                 __mem_cgroup_cancel_charge(parent, page_size >> PAGE_SHIFT);
2307
2308         if (page_size > PAGE_SIZE)
2309                 compound_unlock_irqrestore(page, flags);
2310 put_back:
2311         putback_lru_page(page);
2312 put:
2313         put_page(page);
2314 out:
2315         return ret;
2316 }
2317
2318 /*
2319  * Charge the memory controller for page usage.
2320  * Return
2321  * 0 if the charge was successful
2322  * < 0 if the cgroup is over its limit
2323  */
2324 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2325                                 gfp_t gfp_mask, enum charge_type ctype)
2326 {
2327         struct mem_cgroup *mem = NULL;
2328         int page_size = PAGE_SIZE;
2329         struct page_cgroup *pc;
2330         bool oom = true;
2331         int ret;
2332
2333         if (PageTransHuge(page)) {
2334                 page_size <<= compound_order(page);
2335                 VM_BUG_ON(!PageTransHuge(page));
2336                 /*
2337                  * Never OOM-kill a process for a huge page.  The
2338                  * fault handler will fall back to regular pages.
2339                  */
2340                 oom = false;
2341         }
2342
2343         pc = lookup_page_cgroup(page);
2344         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2345
2346         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2347         if (ret || !mem)
2348                 return ret;
2349
2350         __mem_cgroup_commit_charge(mem, page, pc, ctype, page_size);
2351         return 0;
2352 }
2353
2354 int mem_cgroup_newpage_charge(struct page *page,
2355                               struct mm_struct *mm, gfp_t gfp_mask)
2356 {
2357         if (mem_cgroup_disabled())
2358                 return 0;
2359         /*
2360          * If already mapped, we don't have to account.
2361          * If page cache, page->mapping has address_space.
2362          * But page->mapping may have out-of-use anon_vma pointer,
2363          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2364          * is NULL.
2365          */
2366         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2367                 return 0;
2368         if (unlikely(!mm))
2369                 mm = &init_mm;
2370         return mem_cgroup_charge_common(page, mm, gfp_mask,
2371                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2372 }
2373
2374 static void
2375 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2376                                         enum charge_type ctype);
2377
2378 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2379                                 gfp_t gfp_mask)
2380 {
2381         int ret;
2382
2383         if (mem_cgroup_disabled())
2384                 return 0;
2385         if (PageCompound(page))
2386                 return 0;
2387         /*
2388          * Corner case handling. This is called from add_to_page_cache()
2389          * in usual. But some FS (shmem) precharges this page before calling it
2390          * and call add_to_page_cache() with GFP_NOWAIT.
2391          *
2392          * For GFP_NOWAIT case, the page may be pre-charged before calling
2393          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2394          * charge twice. (It works but has to pay a bit larger cost.)
2395          * And when the page is SwapCache, it should take swap information
2396          * into account. This is under lock_page() now.
2397          */
2398         if (!(gfp_mask & __GFP_WAIT)) {
2399                 struct page_cgroup *pc;
2400
2401                 pc = lookup_page_cgroup(page);
2402                 if (!pc)
2403                         return 0;
2404                 lock_page_cgroup(pc);
2405                 if (PageCgroupUsed(pc)) {
2406                         unlock_page_cgroup(pc);
2407                         return 0;
2408                 }
2409                 unlock_page_cgroup(pc);
2410         }
2411
2412         if (unlikely(!mm))
2413                 mm = &init_mm;
2414
2415         if (page_is_file_cache(page))
2416                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2417                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2418
2419         /* shmem */
2420         if (PageSwapCache(page)) {
2421                 struct mem_cgroup *mem;
2422
2423                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2424                 if (!ret)
2425                         __mem_cgroup_commit_charge_swapin(page, mem,
2426                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2427         } else
2428                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2429                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2430
2431         return ret;
2432 }
2433
2434 /*
2435  * While swap-in, try_charge -> commit or cancel, the page is locked.
2436  * And when try_charge() successfully returns, one refcnt to memcg without
2437  * struct page_cgroup is acquired. This refcnt will be consumed by
2438  * "commit()" or removed by "cancel()"
2439  */
2440 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2441                                  struct page *page,
2442                                  gfp_t mask, struct mem_cgroup **ptr)
2443 {
2444         struct mem_cgroup *mem;
2445         int ret;
2446
2447         *ptr = NULL;
2448
2449         if (mem_cgroup_disabled())
2450                 return 0;
2451
2452         if (!do_swap_account)
2453                 goto charge_cur_mm;
2454         /*
2455          * A racing thread's fault, or swapoff, may have already updated
2456          * the pte, and even removed page from swap cache: in those cases
2457          * do_swap_page()'s pte_same() test will fail; but there's also a
2458          * KSM case which does need to charge the page.
2459          */
2460         if (!PageSwapCache(page))
2461                 goto charge_cur_mm;
2462         mem = try_get_mem_cgroup_from_page(page);
2463         if (!mem)
2464                 goto charge_cur_mm;
2465         *ptr = mem;
2466         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2467         css_put(&mem->css);
2468         return ret;
2469 charge_cur_mm:
2470         if (unlikely(!mm))
2471                 mm = &init_mm;
2472         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2473 }
2474
2475 static void
2476 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2477                                         enum charge_type ctype)
2478 {
2479         struct page_cgroup *pc;
2480
2481         if (mem_cgroup_disabled())
2482                 return;
2483         if (!ptr)
2484                 return;
2485         cgroup_exclude_rmdir(&ptr->css);
2486         pc = lookup_page_cgroup(page);
2487         mem_cgroup_lru_del_before_commit_swapcache(page);
2488         __mem_cgroup_commit_charge(ptr, page, pc, ctype, PAGE_SIZE);
2489         mem_cgroup_lru_add_after_commit_swapcache(page);
2490         /*
2491          * Now swap is on-memory. This means this page may be
2492          * counted both as mem and swap....double count.
2493          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2494          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2495          * may call delete_from_swap_cache() before reach here.
2496          */
2497         if (do_swap_account && PageSwapCache(page)) {
2498                 swp_entry_t ent = {.val = page_private(page)};
2499                 unsigned short id;
2500                 struct mem_cgroup *memcg;
2501
2502                 id = swap_cgroup_record(ent, 0);
2503                 rcu_read_lock();
2504                 memcg = mem_cgroup_lookup(id);
2505                 if (memcg) {
2506                         /*
2507                          * This recorded memcg can be obsolete one. So, avoid
2508                          * calling css_tryget
2509                          */
2510                         if (!mem_cgroup_is_root(memcg))
2511                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2512                         mem_cgroup_swap_statistics(memcg, false);
2513                         mem_cgroup_put(memcg);
2514                 }
2515                 rcu_read_unlock();
2516         }
2517         /*
2518          * At swapin, we may charge account against cgroup which has no tasks.
2519          * So, rmdir()->pre_destroy() can be called while we do this charge.
2520          * In that case, we need to call pre_destroy() again. check it here.
2521          */
2522         cgroup_release_and_wakeup_rmdir(&ptr->css);
2523 }
2524
2525 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2526 {
2527         __mem_cgroup_commit_charge_swapin(page, ptr,
2528                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2529 }
2530
2531 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2532 {
2533         if (mem_cgroup_disabled())
2534                 return;
2535         if (!mem)
2536                 return;
2537         __mem_cgroup_cancel_charge(mem, 1);
2538 }
2539
2540 static void
2541 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2542               int page_size)
2543 {
2544         struct memcg_batch_info *batch = NULL;
2545         bool uncharge_memsw = true;
2546         /* If swapout, usage of swap doesn't decrease */
2547         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2548                 uncharge_memsw = false;
2549
2550         batch = &current->memcg_batch;
2551         /*
2552          * In usual, we do css_get() when we remember memcg pointer.
2553          * But in this case, we keep res->usage until end of a series of
2554          * uncharges. Then, it's ok to ignore memcg's refcnt.
2555          */
2556         if (!batch->memcg)
2557                 batch->memcg = mem;
2558         /*
2559          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2560          * In those cases, all pages freed continously can be expected to be in
2561          * the same cgroup and we have chance to coalesce uncharges.
2562          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2563          * because we want to do uncharge as soon as possible.
2564          */
2565
2566         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2567                 goto direct_uncharge;
2568
2569         if (page_size != PAGE_SIZE)
2570                 goto direct_uncharge;
2571
2572         /*
2573          * In typical case, batch->memcg == mem. This means we can
2574          * merge a series of uncharges to an uncharge of res_counter.
2575          * If not, we uncharge res_counter ony by one.
2576          */
2577         if (batch->memcg != mem)
2578                 goto direct_uncharge;
2579         /* remember freed charge and uncharge it later */
2580         batch->bytes += PAGE_SIZE;
2581         if (uncharge_memsw)
2582                 batch->memsw_bytes += PAGE_SIZE;
2583         return;
2584 direct_uncharge:
2585         res_counter_uncharge(&mem->res, page_size);
2586         if (uncharge_memsw)
2587                 res_counter_uncharge(&mem->memsw, page_size);
2588         if (unlikely(batch->memcg != mem))
2589                 memcg_oom_recover(mem);
2590         return;
2591 }
2592
2593 /*
2594  * uncharge if !page_mapped(page)
2595  */
2596 static struct mem_cgroup *
2597 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2598 {
2599         int count;
2600         struct page_cgroup *pc;
2601         struct mem_cgroup *mem = NULL;
2602         int page_size = PAGE_SIZE;
2603
2604         if (mem_cgroup_disabled())
2605                 return NULL;
2606
2607         if (PageSwapCache(page))
2608                 return NULL;
2609
2610         if (PageTransHuge(page)) {
2611                 page_size <<= compound_order(page);
2612                 VM_BUG_ON(!PageTransHuge(page));
2613         }
2614
2615         count = page_size >> PAGE_SHIFT;
2616         /*
2617          * Check if our page_cgroup is valid
2618          */
2619         pc = lookup_page_cgroup(page);
2620         if (unlikely(!pc || !PageCgroupUsed(pc)))
2621                 return NULL;
2622
2623         lock_page_cgroup(pc);
2624
2625         mem = pc->mem_cgroup;
2626
2627         if (!PageCgroupUsed(pc))
2628                 goto unlock_out;
2629
2630         switch (ctype) {
2631         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2632         case MEM_CGROUP_CHARGE_TYPE_DROP:
2633                 /* See mem_cgroup_prepare_migration() */
2634                 if (page_mapped(page) || PageCgroupMigration(pc))
2635                         goto unlock_out;
2636                 break;
2637         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2638                 if (!PageAnon(page)) {  /* Shared memory */
2639                         if (page->mapping && !page_is_file_cache(page))
2640                                 goto unlock_out;
2641                 } else if (page_mapped(page)) /* Anon */
2642                                 goto unlock_out;
2643                 break;
2644         default:
2645                 break;
2646         }
2647
2648         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2649
2650         ClearPageCgroupUsed(pc);
2651         /*
2652          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2653          * freed from LRU. This is safe because uncharged page is expected not
2654          * to be reused (freed soon). Exception is SwapCache, it's handled by
2655          * special functions.
2656          */
2657
2658         unlock_page_cgroup(pc);
2659         /*
2660          * even after unlock, we have mem->res.usage here and this memcg
2661          * will never be freed.
2662          */
2663         memcg_check_events(mem, page);
2664         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2665                 mem_cgroup_swap_statistics(mem, true);
2666                 mem_cgroup_get(mem);
2667         }
2668         if (!mem_cgroup_is_root(mem))
2669                 __do_uncharge(mem, ctype, page_size);
2670
2671         return mem;
2672
2673 unlock_out:
2674         unlock_page_cgroup(pc);
2675         return NULL;
2676 }
2677
2678 void mem_cgroup_uncharge_page(struct page *page)
2679 {
2680         /* early check. */
2681         if (page_mapped(page))
2682                 return;
2683         if (page->mapping && !PageAnon(page))
2684                 return;
2685         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2686 }
2687
2688 void mem_cgroup_uncharge_cache_page(struct page *page)
2689 {
2690         VM_BUG_ON(page_mapped(page));
2691         VM_BUG_ON(page->mapping);
2692         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2693 }
2694
2695 /*
2696  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2697  * In that cases, pages are freed continuously and we can expect pages
2698  * are in the same memcg. All these calls itself limits the number of
2699  * pages freed at once, then uncharge_start/end() is called properly.
2700  * This may be called prural(2) times in a context,
2701  */
2702
2703 void mem_cgroup_uncharge_start(void)
2704 {
2705         current->memcg_batch.do_batch++;
2706         /* We can do nest. */
2707         if (current->memcg_batch.do_batch == 1) {
2708                 current->memcg_batch.memcg = NULL;
2709                 current->memcg_batch.bytes = 0;
2710                 current->memcg_batch.memsw_bytes = 0;
2711         }
2712 }
2713
2714 void mem_cgroup_uncharge_end(void)
2715 {
2716         struct memcg_batch_info *batch = &current->memcg_batch;
2717
2718         if (!batch->do_batch)
2719                 return;
2720
2721         batch->do_batch--;
2722         if (batch->do_batch) /* If stacked, do nothing. */
2723                 return;
2724
2725         if (!batch->memcg)
2726                 return;
2727         /*
2728          * This "batch->memcg" is valid without any css_get/put etc...
2729          * bacause we hide charges behind us.
2730          */
2731         if (batch->bytes)
2732                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2733         if (batch->memsw_bytes)
2734                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2735         memcg_oom_recover(batch->memcg);
2736         /* forget this pointer (for sanity check) */
2737         batch->memcg = NULL;
2738 }
2739
2740 #ifdef CONFIG_SWAP
2741 /*
2742  * called after __delete_from_swap_cache() and drop "page" account.
2743  * memcg information is recorded to swap_cgroup of "ent"
2744  */
2745 void
2746 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2747 {
2748         struct mem_cgroup *memcg;
2749         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2750
2751         if (!swapout) /* this was a swap cache but the swap is unused ! */
2752                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2753
2754         memcg = __mem_cgroup_uncharge_common(page, ctype);
2755
2756         /*
2757          * record memcg information,  if swapout && memcg != NULL,
2758          * mem_cgroup_get() was called in uncharge().
2759          */
2760         if (do_swap_account && swapout && memcg)
2761                 swap_cgroup_record(ent, css_id(&memcg->css));
2762 }
2763 #endif
2764
2765 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2766 /*
2767  * called from swap_entry_free(). remove record in swap_cgroup and
2768  * uncharge "memsw" account.
2769  */
2770 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2771 {
2772         struct mem_cgroup *memcg;
2773         unsigned short id;
2774
2775         if (!do_swap_account)
2776                 return;
2777
2778         id = swap_cgroup_record(ent, 0);
2779         rcu_read_lock();
2780         memcg = mem_cgroup_lookup(id);
2781         if (memcg) {
2782                 /*
2783                  * We uncharge this because swap is freed.
2784                  * This memcg can be obsolete one. We avoid calling css_tryget
2785                  */
2786                 if (!mem_cgroup_is_root(memcg))
2787                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2788                 mem_cgroup_swap_statistics(memcg, false);
2789                 mem_cgroup_put(memcg);
2790         }
2791         rcu_read_unlock();
2792 }
2793
2794 /**
2795  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2796  * @entry: swap entry to be moved
2797  * @from:  mem_cgroup which the entry is moved from
2798  * @to:  mem_cgroup which the entry is moved to
2799  * @need_fixup: whether we should fixup res_counters and refcounts.
2800  *
2801  * It succeeds only when the swap_cgroup's record for this entry is the same
2802  * as the mem_cgroup's id of @from.
2803  *
2804  * Returns 0 on success, -EINVAL on failure.
2805  *
2806  * The caller must have charged to @to, IOW, called res_counter_charge() about
2807  * both res and memsw, and called css_get().
2808  */
2809 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2810                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2811 {
2812         unsigned short old_id, new_id;
2813
2814         old_id = css_id(&from->css);
2815         new_id = css_id(&to->css);
2816
2817         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2818                 mem_cgroup_swap_statistics(from, false);
2819                 mem_cgroup_swap_statistics(to, true);
2820                 /*
2821                  * This function is only called from task migration context now.
2822                  * It postpones res_counter and refcount handling till the end
2823                  * of task migration(mem_cgroup_clear_mc()) for performance
2824                  * improvement. But we cannot postpone mem_cgroup_get(to)
2825                  * because if the process that has been moved to @to does
2826                  * swap-in, the refcount of @to might be decreased to 0.
2827                  */
2828                 mem_cgroup_get(to);
2829                 if (need_fixup) {
2830                         if (!mem_cgroup_is_root(from))
2831                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2832                         mem_cgroup_put(from);
2833                         /*
2834                          * we charged both to->res and to->memsw, so we should
2835                          * uncharge to->res.
2836                          */
2837                         if (!mem_cgroup_is_root(to))
2838                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2839                 }
2840                 return 0;
2841         }
2842         return -EINVAL;
2843 }
2844 #else
2845 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2846                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2847 {
2848         return -EINVAL;
2849 }
2850 #endif
2851
2852 /*
2853  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2854  * page belongs to.
2855  */
2856 int mem_cgroup_prepare_migration(struct page *page,
2857         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2858 {
2859         struct page_cgroup *pc;
2860         struct mem_cgroup *mem = NULL;
2861         enum charge_type ctype;
2862         int ret = 0;
2863
2864         *ptr = NULL;
2865
2866         VM_BUG_ON(PageTransHuge(page));
2867         if (mem_cgroup_disabled())
2868                 return 0;
2869
2870         pc = lookup_page_cgroup(page);
2871         lock_page_cgroup(pc);
2872         if (PageCgroupUsed(pc)) {
2873                 mem = pc->mem_cgroup;
2874                 css_get(&mem->css);
2875                 /*
2876                  * At migrating an anonymous page, its mapcount goes down
2877                  * to 0 and uncharge() will be called. But, even if it's fully
2878                  * unmapped, migration may fail and this page has to be
2879                  * charged again. We set MIGRATION flag here and delay uncharge
2880                  * until end_migration() is called
2881                  *
2882                  * Corner Case Thinking
2883                  * A)
2884                  * When the old page was mapped as Anon and it's unmap-and-freed
2885                  * while migration was ongoing.
2886                  * If unmap finds the old page, uncharge() of it will be delayed
2887                  * until end_migration(). If unmap finds a new page, it's
2888                  * uncharged when it make mapcount to be 1->0. If unmap code
2889                  * finds swap_migration_entry, the new page will not be mapped
2890                  * and end_migration() will find it(mapcount==0).
2891                  *
2892                  * B)
2893                  * When the old page was mapped but migraion fails, the kernel
2894                  * remaps it. A charge for it is kept by MIGRATION flag even
2895                  * if mapcount goes down to 0. We can do remap successfully
2896                  * without charging it again.
2897                  *
2898                  * C)
2899                  * The "old" page is under lock_page() until the end of
2900                  * migration, so, the old page itself will not be swapped-out.
2901                  * If the new page is swapped out before end_migraton, our
2902                  * hook to usual swap-out path will catch the event.
2903                  */
2904                 if (PageAnon(page))
2905                         SetPageCgroupMigration(pc);
2906         }
2907         unlock_page_cgroup(pc);
2908         /*
2909          * If the page is not charged at this point,
2910          * we return here.
2911          */
2912         if (!mem)
2913                 return 0;
2914
2915         *ptr = mem;
2916         ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2917         css_put(&mem->css);/* drop extra refcnt */
2918         if (ret || *ptr == NULL) {
2919                 if (PageAnon(page)) {
2920                         lock_page_cgroup(pc);
2921                         ClearPageCgroupMigration(pc);
2922                         unlock_page_cgroup(pc);
2923                         /*
2924                          * The old page may be fully unmapped while we kept it.
2925                          */
2926                         mem_cgroup_uncharge_page(page);
2927                 }
2928                 return -ENOMEM;
2929         }
2930         /*
2931          * We charge new page before it's used/mapped. So, even if unlock_page()
2932          * is called before end_migration, we can catch all events on this new
2933          * page. In the case new page is migrated but not remapped, new page's
2934          * mapcount will be finally 0 and we call uncharge in end_migration().
2935          */
2936         pc = lookup_page_cgroup(newpage);
2937         if (PageAnon(page))
2938                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2939         else if (page_is_file_cache(page))
2940                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2941         else
2942                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2943         __mem_cgroup_commit_charge(mem, page, pc, ctype, PAGE_SIZE);
2944         return ret;
2945 }
2946
2947 /* remove redundant charge if migration failed*/
2948 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2949         struct page *oldpage, struct page *newpage, bool migration_ok)
2950 {
2951         struct page *used, *unused;
2952         struct page_cgroup *pc;
2953
2954         if (!mem)
2955                 return;
2956         /* blocks rmdir() */
2957         cgroup_exclude_rmdir(&mem->css);
2958         if (!migration_ok) {
2959                 used = oldpage;
2960                 unused = newpage;
2961         } else {
2962                 used = newpage;
2963                 unused = oldpage;
2964         }
2965         /*
2966          * We disallowed uncharge of pages under migration because mapcount
2967          * of the page goes down to zero, temporarly.
2968          * Clear the flag and check the page should be charged.
2969          */
2970         pc = lookup_page_cgroup(oldpage);
2971         lock_page_cgroup(pc);
2972         ClearPageCgroupMigration(pc);
2973         unlock_page_cgroup(pc);
2974
2975         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2976
2977         /*
2978          * If a page is a file cache, radix-tree replacement is very atomic
2979          * and we can skip this check. When it was an Anon page, its mapcount
2980          * goes down to 0. But because we added MIGRATION flage, it's not
2981          * uncharged yet. There are several case but page->mapcount check
2982          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2983          * check. (see prepare_charge() also)
2984          */
2985         if (PageAnon(used))
2986                 mem_cgroup_uncharge_page(used);
2987         /*
2988          * At migration, we may charge account against cgroup which has no
2989          * tasks.
2990          * So, rmdir()->pre_destroy() can be called while we do this charge.
2991          * In that case, we need to call pre_destroy() again. check it here.
2992          */
2993         cgroup_release_and_wakeup_rmdir(&mem->css);
2994 }
2995
2996 /*
2997  * A call to try to shrink memory usage on charge failure at shmem's swapin.
2998  * Calling hierarchical_reclaim is not enough because we should update
2999  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3000  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3001  * not from the memcg which this page would be charged to.
3002  * try_charge_swapin does all of these works properly.
3003  */
3004 int mem_cgroup_shmem_charge_fallback(struct page *page,
3005                             struct mm_struct *mm,
3006                             gfp_t gfp_mask)
3007 {
3008         struct mem_cgroup *mem;
3009         int ret;
3010
3011         if (mem_cgroup_disabled())
3012                 return 0;
3013
3014         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3015         if (!ret)
3016                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3017
3018         return ret;
3019 }
3020
3021 #ifdef CONFIG_DEBUG_VM
3022 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3023 {
3024         struct page_cgroup *pc;
3025
3026         pc = lookup_page_cgroup(page);
3027         if (likely(pc) && PageCgroupUsed(pc))
3028                 return pc;
3029         return NULL;
3030 }
3031
3032 bool mem_cgroup_bad_page_check(struct page *page)
3033 {
3034         if (mem_cgroup_disabled())
3035                 return false;
3036
3037         return lookup_page_cgroup_used(page) != NULL;
3038 }
3039
3040 void mem_cgroup_print_bad_page(struct page *page)
3041 {
3042         struct page_cgroup *pc;
3043
3044         pc = lookup_page_cgroup_used(page);
3045         if (pc) {
3046                 int ret = -1;
3047                 char *path;
3048
3049                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3050                        pc, pc->flags, pc->mem_cgroup);
3051
3052                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3053                 if (path) {
3054                         rcu_read_lock();
3055                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3056                                                         path, PATH_MAX);
3057                         rcu_read_unlock();
3058                 }
3059
3060                 printk(KERN_CONT "(%s)\n",
3061                                 (ret < 0) ? "cannot get the path" : path);
3062                 kfree(path);
3063         }
3064 }
3065 #endif
3066
3067 static DEFINE_MUTEX(set_limit_mutex);
3068
3069 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3070                                 unsigned long long val)
3071 {
3072         int retry_count;
3073         u64 memswlimit, memlimit;
3074         int ret = 0;
3075         int children = mem_cgroup_count_children(memcg);
3076         u64 curusage, oldusage;
3077         int enlarge;
3078
3079         /*
3080          * For keeping hierarchical_reclaim simple, how long we should retry
3081          * is depends on callers. We set our retry-count to be function
3082          * of # of children which we should visit in this loop.
3083          */
3084         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3085
3086         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3087
3088         enlarge = 0;
3089         while (retry_count) {
3090                 if (signal_pending(current)) {
3091                         ret = -EINTR;
3092                         break;
3093                 }
3094                 /*
3095                  * Rather than hide all in some function, I do this in
3096                  * open coded manner. You see what this really does.
3097                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3098                  */
3099                 mutex_lock(&set_limit_mutex);
3100                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3101                 if (memswlimit < val) {
3102                         ret = -EINVAL;
3103                         mutex_unlock(&set_limit_mutex);
3104                         break;
3105                 }
3106
3107                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3108                 if (memlimit < val)
3109                         enlarge = 1;
3110
3111                 ret = res_counter_set_limit(&memcg->res, val);
3112                 if (!ret) {
3113                         if (memswlimit == val)
3114                                 memcg->memsw_is_minimum = true;
3115                         else
3116                                 memcg->memsw_is_minimum = false;
3117                 }
3118                 mutex_unlock(&set_limit_mutex);
3119
3120                 if (!ret)
3121                         break;
3122
3123                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3124                                                 MEM_CGROUP_RECLAIM_SHRINK);
3125                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3126                 /* Usage is reduced ? */
3127                 if (curusage >= oldusage)
3128                         retry_count--;
3129                 else
3130                         oldusage = curusage;
3131         }
3132         if (!ret && enlarge)
3133                 memcg_oom_recover(memcg);
3134
3135         return ret;
3136 }
3137
3138 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3139                                         unsigned long long val)
3140 {
3141         int retry_count;
3142         u64 memlimit, memswlimit, oldusage, curusage;
3143         int children = mem_cgroup_count_children(memcg);
3144         int ret = -EBUSY;
3145         int enlarge = 0;
3146
3147         /* see mem_cgroup_resize_res_limit */
3148         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3149         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3150         while (retry_count) {
3151                 if (signal_pending(current)) {
3152                         ret = -EINTR;
3153                         break;
3154                 }
3155                 /*
3156                  * Rather than hide all in some function, I do this in
3157                  * open coded manner. You see what this really does.
3158                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3159                  */
3160                 mutex_lock(&set_limit_mutex);
3161                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3162                 if (memlimit > val) {
3163                         ret = -EINVAL;
3164                         mutex_unlock(&set_limit_mutex);
3165                         break;
3166                 }
3167                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3168                 if (memswlimit < val)
3169                         enlarge = 1;
3170                 ret = res_counter_set_limit(&memcg->memsw, val);
3171                 if (!ret) {
3172                         if (memlimit == val)
3173                                 memcg->memsw_is_minimum = true;
3174                         else
3175                                 memcg->memsw_is_minimum = false;
3176                 }
3177                 mutex_unlock(&set_limit_mutex);
3178
3179                 if (!ret)
3180                         break;
3181
3182                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3183                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3184                                                 MEM_CGROUP_RECLAIM_SHRINK);
3185                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3186                 /* Usage is reduced ? */
3187                 if (curusage >= oldusage)
3188                         retry_count--;
3189                 else
3190                         oldusage = curusage;
3191         }
3192         if (!ret && enlarge)
3193                 memcg_oom_recover(memcg);
3194         return ret;
3195 }
3196
3197 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3198                                             gfp_t gfp_mask)
3199 {
3200         unsigned long nr_reclaimed = 0;
3201         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3202         unsigned long reclaimed;
3203         int loop = 0;
3204         struct mem_cgroup_tree_per_zone *mctz;
3205         unsigned long long excess;
3206
3207         if (order > 0)
3208                 return 0;
3209
3210         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3211         /*
3212          * This loop can run a while, specially if mem_cgroup's continuously
3213          * keep exceeding their soft limit and putting the system under
3214          * pressure
3215          */
3216         do {
3217                 if (next_mz)
3218                         mz = next_mz;
3219                 else
3220                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3221                 if (!mz)
3222                         break;
3223
3224                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3225                                                 gfp_mask,
3226                                                 MEM_CGROUP_RECLAIM_SOFT);
3227                 nr_reclaimed += reclaimed;
3228                 spin_lock(&mctz->lock);
3229
3230                 /*
3231                  * If we failed to reclaim anything from this memory cgroup
3232                  * it is time to move on to the next cgroup
3233                  */
3234                 next_mz = NULL;
3235                 if (!reclaimed) {
3236                         do {
3237                                 /*
3238                                  * Loop until we find yet another one.
3239                                  *
3240                                  * By the time we get the soft_limit lock
3241                                  * again, someone might have aded the
3242                                  * group back on the RB tree. Iterate to
3243                                  * make sure we get a different mem.
3244                                  * mem_cgroup_largest_soft_limit_node returns
3245                                  * NULL if no other cgroup is present on
3246                                  * the tree
3247                                  */
3248                                 next_mz =
3249                                 __mem_cgroup_largest_soft_limit_node(mctz);
3250                                 if (next_mz == mz) {
3251                                         css_put(&next_mz->mem->css);
3252                                         next_mz = NULL;
3253                                 } else /* next_mz == NULL or other memcg */
3254                                         break;
3255                         } while (1);
3256                 }
3257                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3258                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3259                 /*
3260                  * One school of thought says that we should not add
3261                  * back the node to the tree if reclaim returns 0.
3262                  * But our reclaim could return 0, simply because due
3263                  * to priority we are exposing a smaller subset of
3264                  * memory to reclaim from. Consider this as a longer
3265                  * term TODO.
3266                  */
3267                 /* If excess == 0, no tree ops */
3268                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3269                 spin_unlock(&mctz->lock);
3270                 css_put(&mz->mem->css);
3271                 loop++;
3272                 /*
3273                  * Could not reclaim anything and there are no more
3274                  * mem cgroups to try or we seem to be looping without
3275                  * reclaiming anything.
3276                  */
3277                 if (!nr_reclaimed &&
3278                         (next_mz == NULL ||
3279                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3280                         break;
3281         } while (!nr_reclaimed);
3282         if (next_mz)
3283                 css_put(&next_mz->mem->css);
3284         return nr_reclaimed;
3285 }
3286
3287 /*
3288  * This routine traverse page_cgroup in given list and drop them all.
3289  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3290  */
3291 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3292                                 int node, int zid, enum lru_list lru)
3293 {
3294         struct zone *zone;
3295         struct mem_cgroup_per_zone *mz;
3296         struct page_cgroup *pc, *busy;
3297         unsigned long flags, loop;
3298         struct list_head *list;
3299         int ret = 0;
3300
3301         zone = &NODE_DATA(node)->node_zones[zid];
3302         mz = mem_cgroup_zoneinfo(mem, node, zid);
3303         list = &mz->lists[lru];
3304
3305         loop = MEM_CGROUP_ZSTAT(mz, lru);
3306         /* give some margin against EBUSY etc...*/
3307         loop += 256;
3308         busy = NULL;
3309         while (loop--) {
3310                 struct page *page;
3311
3312                 ret = 0;
3313                 spin_lock_irqsave(&zone->lru_lock, flags);
3314                 if (list_empty(list)) {
3315                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3316                         break;
3317                 }
3318                 pc = list_entry(list->prev, struct page_cgroup, lru);
3319                 if (busy == pc) {
3320                         list_move(&pc->lru, list);
3321                         busy = NULL;
3322                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3323                         continue;
3324                 }
3325                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3326
3327                 page = lookup_cgroup_page(pc);
3328
3329                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3330                 if (ret == -ENOMEM)
3331                         break;
3332
3333                 if (ret == -EBUSY || ret == -EINVAL) {
3334                         /* found lock contention or "pc" is obsolete. */
3335                         busy = pc;
3336                         cond_resched();
3337                 } else
3338                         busy = NULL;
3339         }
3340
3341         if (!ret && !list_empty(list))
3342                 return -EBUSY;
3343         return ret;
3344 }
3345
3346 /*
3347  * make mem_cgroup's charge to be 0 if there is no task.
3348  * This enables deleting this mem_cgroup.
3349  */
3350 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3351 {
3352         int ret;
3353         int node, zid, shrink;
3354         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3355         struct cgroup *cgrp = mem->css.cgroup;
3356
3357         css_get(&mem->css);
3358
3359         shrink = 0;
3360         /* should free all ? */
3361         if (free_all)
3362                 goto try_to_free;
3363 move_account:
3364         do {
3365                 ret = -EBUSY;
3366                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3367                         goto out;
3368                 ret = -EINTR;
3369                 if (signal_pending(current))
3370                         goto out;
3371                 /* This is for making all *used* pages to be on LRU. */
3372                 lru_add_drain_all();
3373                 drain_all_stock_sync();
3374                 ret = 0;
3375                 mem_cgroup_start_move(mem);
3376                 for_each_node_state(node, N_HIGH_MEMORY) {
3377                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3378                                 enum lru_list l;
3379                                 for_each_lru(l) {
3380                                         ret = mem_cgroup_force_empty_list(mem,
3381                                                         node, zid, l);
3382                                         if (ret)
3383                                                 break;
3384                                 }
3385                         }
3386                         if (ret)
3387                                 break;
3388                 }
3389                 mem_cgroup_end_move(mem);
3390                 memcg_oom_recover(mem);
3391                 /* it seems parent cgroup doesn't have enough mem */
3392                 if (ret == -ENOMEM)
3393                         goto try_to_free;
3394                 cond_resched();
3395         /* "ret" should also be checked to ensure all lists are empty. */
3396         } while (mem->res.usage > 0 || ret);
3397 out:
3398         css_put(&mem->css);
3399         return ret;
3400
3401 try_to_free:
3402         /* returns EBUSY if there is a task or if we come here twice. */
3403         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3404                 ret = -EBUSY;
3405                 goto out;
3406         }
3407         /* we call try-to-free pages for make this cgroup empty */
3408         lru_add_drain_all();
3409         /* try to free all pages in this cgroup */
3410         shrink = 1;
3411         while (nr_retries && mem->res.usage > 0) {
3412                 int progress;
3413
3414                 if (signal_pending(current)) {
3415                         ret = -EINTR;
3416                         goto out;
3417                 }
3418                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3419                                                 false, get_swappiness(mem));
3420                 if (!progress) {
3421                         nr_retries--;
3422                         /* maybe some writeback is necessary */
3423                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3424                 }
3425
3426         }
3427         lru_add_drain();
3428         /* try move_account...there may be some *locked* pages. */
3429         goto move_account;
3430 }
3431
3432 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3433 {
3434         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3435 }
3436
3437
3438 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3439 {
3440         return mem_cgroup_from_cont(cont)->use_hierarchy;
3441 }
3442
3443 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3444                                         u64 val)
3445 {
3446         int retval = 0;
3447         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3448         struct cgroup *parent = cont->parent;
3449         struct mem_cgroup *parent_mem = NULL;
3450
3451         if (parent)
3452                 parent_mem = mem_cgroup_from_cont(parent);
3453
3454         cgroup_lock();
3455         /*
3456          * If parent's use_hierarchy is set, we can't make any modifications
3457          * in the child subtrees. If it is unset, then the change can
3458          * occur, provided the current cgroup has no children.
3459          *
3460          * For the root cgroup, parent_mem is NULL, we allow value to be
3461          * set if there are no children.
3462          */
3463         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3464                                 (val == 1 || val == 0)) {
3465                 if (list_empty(&cont->children))
3466                         mem->use_hierarchy = val;
3467                 else
3468                         retval = -EBUSY;
3469         } else
3470                 retval = -EINVAL;
3471         cgroup_unlock();
3472
3473         return retval;
3474 }
3475
3476
3477 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3478                                 enum mem_cgroup_stat_index idx)
3479 {
3480         struct mem_cgroup *iter;
3481         s64 val = 0;
3482
3483         /* each per cpu's value can be minus.Then, use s64 */
3484         for_each_mem_cgroup_tree(iter, mem)
3485                 val += mem_cgroup_read_stat(iter, idx);
3486
3487         if (val < 0) /* race ? */
3488                 val = 0;
3489         return val;
3490 }
3491
3492 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3493 {
3494         u64 val;
3495
3496         if (!mem_cgroup_is_root(mem)) {
3497                 if (!swap)
3498                         return res_counter_read_u64(&mem->res, RES_USAGE);
3499                 else
3500                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3501         }
3502
3503         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3504         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3505
3506         if (swap)
3507                 val += mem_cgroup_get_recursive_idx_stat(mem,
3508                                 MEM_CGROUP_STAT_SWAPOUT);
3509
3510         return val << PAGE_SHIFT;
3511 }
3512
3513 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3514 {
3515         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3516         u64 val;
3517         int type, name;
3518
3519         type = MEMFILE_TYPE(cft->private);
3520         name = MEMFILE_ATTR(cft->private);
3521         switch (type) {
3522         case _MEM:
3523                 if (name == RES_USAGE)
3524                         val = mem_cgroup_usage(mem, false);
3525                 else
3526                         val = res_counter_read_u64(&mem->res, name);
3527                 break;
3528         case _MEMSWAP:
3529                 if (name == RES_USAGE)
3530                         val = mem_cgroup_usage(mem, true);
3531                 else
3532                         val = res_counter_read_u64(&mem->memsw, name);
3533                 break;
3534         default:
3535                 BUG();
3536                 break;
3537         }
3538         return val;
3539 }
3540 /*
3541  * The user of this function is...
3542  * RES_LIMIT.
3543  */
3544 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3545                             const char *buffer)
3546 {
3547         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3548         int type, name;
3549         unsigned long long val;
3550         int ret;
3551
3552         type = MEMFILE_TYPE(cft->private);
3553         name = MEMFILE_ATTR(cft->private);
3554         switch (name) {
3555         case RES_LIMIT:
3556                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3557                         ret = -EINVAL;
3558                         break;
3559                 }
3560                 /* This function does all necessary parse...reuse it */
3561                 ret = res_counter_memparse_write_strategy(buffer, &val);
3562                 if (ret)
3563                         break;
3564                 if (type == _MEM)
3565                         ret = mem_cgroup_resize_limit(memcg, val);
3566                 else
3567                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3568                 break;
3569         case RES_SOFT_LIMIT:
3570                 ret = res_counter_memparse_write_strategy(buffer, &val);
3571                 if (ret)
3572                         break;
3573                 /*
3574                  * For memsw, soft limits are hard to implement in terms
3575                  * of semantics, for now, we support soft limits for
3576                  * control without swap
3577                  */
3578                 if (type == _MEM)
3579                         ret = res_counter_set_soft_limit(&memcg->res, val);
3580                 else
3581                         ret = -EINVAL;
3582                 break;
3583         default:
3584                 ret = -EINVAL; /* should be BUG() ? */
3585                 break;
3586         }
3587         return ret;
3588 }
3589
3590 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3591                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3592 {
3593         struct cgroup *cgroup;
3594         unsigned long long min_limit, min_memsw_limit, tmp;
3595
3596         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3597         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3598         cgroup = memcg->css.cgroup;
3599         if (!memcg->use_hierarchy)
3600                 goto out;
3601
3602         while (cgroup->parent) {
3603                 cgroup = cgroup->parent;
3604                 memcg = mem_cgroup_from_cont(cgroup);
3605                 if (!memcg->use_hierarchy)
3606                         break;
3607                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3608                 min_limit = min(min_limit, tmp);
3609                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3610                 min_memsw_limit = min(min_memsw_limit, tmp);
3611         }
3612 out:
3613         *mem_limit = min_limit;
3614         *memsw_limit = min_memsw_limit;
3615         return;
3616 }
3617
3618 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3619 {
3620         struct mem_cgroup *mem;
3621         int type, name;
3622
3623         mem = mem_cgroup_from_cont(cont);
3624         type = MEMFILE_TYPE(event);
3625         name = MEMFILE_ATTR(event);
3626         switch (name) {
3627         case RES_MAX_USAGE:
3628                 if (type == _MEM)
3629                         res_counter_reset_max(&mem->res);
3630                 else
3631                         res_counter_reset_max(&mem->memsw);
3632                 break;
3633         case RES_FAILCNT:
3634                 if (type == _MEM)
3635                         res_counter_reset_failcnt(&mem->res);
3636                 else
3637                         res_counter_reset_failcnt(&mem->memsw);
3638                 break;
3639         }
3640
3641         return 0;
3642 }
3643
3644 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3645                                         struct cftype *cft)
3646 {
3647         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3648 }
3649
3650 #ifdef CONFIG_MMU
3651 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3652                                         struct cftype *cft, u64 val)
3653 {
3654         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3655
3656         if (val >= (1 << NR_MOVE_TYPE))
3657                 return -EINVAL;
3658         /*
3659          * We check this value several times in both in can_attach() and
3660          * attach(), so we need cgroup lock to prevent this value from being
3661          * inconsistent.
3662          */
3663         cgroup_lock();
3664         mem->move_charge_at_immigrate = val;
3665         cgroup_unlock();
3666
3667         return 0;
3668 }
3669 #else
3670 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3671                                         struct cftype *cft, u64 val)
3672 {
3673         return -ENOSYS;
3674 }
3675 #endif
3676
3677
3678 /* For read statistics */
3679 enum {
3680         MCS_CACHE,
3681         MCS_RSS,
3682         MCS_FILE_MAPPED,
3683         MCS_PGPGIN,
3684         MCS_PGPGOUT,
3685         MCS_SWAP,
3686         MCS_INACTIVE_ANON,
3687         MCS_ACTIVE_ANON,
3688         MCS_INACTIVE_FILE,
3689         MCS_ACTIVE_FILE,
3690         MCS_UNEVICTABLE,
3691         NR_MCS_STAT,
3692 };
3693
3694 struct mcs_total_stat {
3695         s64 stat[NR_MCS_STAT];
3696 };
3697
3698 struct {
3699         char *local_name;
3700         char *total_name;
3701 } memcg_stat_strings[NR_MCS_STAT] = {
3702         {"cache", "total_cache"},
3703         {"rss", "total_rss"},
3704         {"mapped_file", "total_mapped_file"},
3705         {"pgpgin", "total_pgpgin"},
3706         {"pgpgout", "total_pgpgout"},
3707         {"swap", "total_swap"},
3708         {"inactive_anon", "total_inactive_anon"},
3709         {"active_anon", "total_active_anon"},
3710         {"inactive_file", "total_inactive_file"},
3711         {"active_file", "total_active_file"},
3712         {"unevictable", "total_unevictable"}
3713 };
3714
3715
3716 static void
3717 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3718 {
3719         s64 val;
3720
3721         /* per cpu stat */
3722         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3723         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3724         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3725         s->stat[MCS_RSS] += val * PAGE_SIZE;
3726         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3727         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3728         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3729         s->stat[MCS_PGPGIN] += val;
3730         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3731         s->stat[MCS_PGPGOUT] += val;
3732         if (do_swap_account) {
3733                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3734                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3735         }
3736
3737         /* per zone stat */
3738         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3739         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3740         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3741         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3742         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3743         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3744         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3745         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3746         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3747         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3748 }
3749
3750 static void
3751 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3752 {
3753         struct mem_cgroup *iter;
3754
3755         for_each_mem_cgroup_tree(iter, mem)
3756                 mem_cgroup_get_local_stat(iter, s);
3757 }
3758
3759 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3760                                  struct cgroup_map_cb *cb)
3761 {
3762         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3763         struct mcs_total_stat mystat;
3764         int i;
3765
3766         memset(&mystat, 0, sizeof(mystat));
3767         mem_cgroup_get_local_stat(mem_cont, &mystat);
3768
3769         for (i = 0; i < NR_MCS_STAT; i++) {
3770                 if (i == MCS_SWAP && !do_swap_account)
3771                         continue;
3772                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3773         }
3774
3775         /* Hierarchical information */
3776         {
3777                 unsigned long long limit, memsw_limit;
3778                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3779                 cb->fill(cb, "hierarchical_memory_limit", limit);
3780                 if (do_swap_account)
3781                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3782         }
3783
3784         memset(&mystat, 0, sizeof(mystat));
3785         mem_cgroup_get_total_stat(mem_cont, &mystat);
3786         for (i = 0; i < NR_MCS_STAT; i++) {
3787                 if (i == MCS_SWAP && !do_swap_account)
3788                         continue;
3789                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3790         }
3791
3792 #ifdef CONFIG_DEBUG_VM
3793         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3794
3795         {
3796                 int nid, zid;
3797                 struct mem_cgroup_per_zone *mz;
3798                 unsigned long recent_rotated[2] = {0, 0};
3799                 unsigned long recent_scanned[2] = {0, 0};
3800
3801                 for_each_online_node(nid)
3802                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3803                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3804
3805                                 recent_rotated[0] +=
3806                                         mz->reclaim_stat.recent_rotated[0];
3807                                 recent_rotated[1] +=
3808                                         mz->reclaim_stat.recent_rotated[1];
3809                                 recent_scanned[0] +=
3810                                         mz->reclaim_stat.recent_scanned[0];
3811                                 recent_scanned[1] +=
3812                                         mz->reclaim_stat.recent_scanned[1];
3813                         }
3814                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3815                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3816                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3817                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3818         }
3819 #endif
3820
3821         return 0;
3822 }
3823
3824 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3825 {
3826         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3827
3828         return get_swappiness(memcg);
3829 }
3830
3831 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3832                                        u64 val)
3833 {
3834         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3835         struct mem_cgroup *parent;
3836
3837         if (val > 100)
3838                 return -EINVAL;
3839
3840         if (cgrp->parent == NULL)
3841                 return -EINVAL;
3842
3843         parent = mem_cgroup_from_cont(cgrp->parent);
3844
3845         cgroup_lock();
3846
3847         /* If under hierarchy, only empty-root can set this value */
3848         if ((parent->use_hierarchy) ||
3849             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3850                 cgroup_unlock();
3851                 return -EINVAL;
3852         }
3853
3854         memcg->swappiness = val;
3855
3856         cgroup_unlock();
3857
3858         return 0;
3859 }
3860
3861 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3862 {
3863         struct mem_cgroup_threshold_ary *t;
3864         u64 usage;
3865         int i;
3866
3867         rcu_read_lock();
3868         if (!swap)
3869                 t = rcu_dereference(memcg->thresholds.primary);
3870         else
3871                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3872
3873         if (!t)
3874                 goto unlock;
3875
3876         usage = mem_cgroup_usage(memcg, swap);
3877
3878         /*
3879          * current_threshold points to threshold just below usage.
3880          * If it's not true, a threshold was crossed after last
3881          * call of __mem_cgroup_threshold().
3882          */
3883         i = t->current_threshold;
3884
3885         /*
3886          * Iterate backward over array of thresholds starting from
3887          * current_threshold and check if a threshold is crossed.
3888          * If none of thresholds below usage is crossed, we read
3889          * only one element of the array here.
3890          */
3891         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3892                 eventfd_signal(t->entries[i].eventfd, 1);
3893
3894         /* i = current_threshold + 1 */
3895         i++;
3896
3897         /*
3898          * Iterate forward over array of thresholds starting from
3899          * current_threshold+1 and check if a threshold is crossed.
3900          * If none of thresholds above usage is crossed, we read
3901          * only one element of the array here.
3902          */
3903         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3904                 eventfd_signal(t->entries[i].eventfd, 1);
3905
3906         /* Update current_threshold */
3907         t->current_threshold = i - 1;
3908 unlock:
3909         rcu_read_unlock();
3910 }
3911
3912 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3913 {
3914         while (memcg) {
3915                 __mem_cgroup_threshold(memcg, false);
3916                 if (do_swap_account)
3917                         __mem_cgroup_threshold(memcg, true);
3918
3919                 memcg = parent_mem_cgroup(memcg);
3920         }
3921 }
3922
3923 static int compare_thresholds(const void *a, const void *b)
3924 {
3925         const struct mem_cgroup_threshold *_a = a;
3926         const struct mem_cgroup_threshold *_b = b;
3927
3928         return _a->threshold - _b->threshold;
3929 }
3930
3931 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3932 {
3933         struct mem_cgroup_eventfd_list *ev;
3934
3935         list_for_each_entry(ev, &mem->oom_notify, list)
3936                 eventfd_signal(ev->eventfd, 1);
3937         return 0;
3938 }
3939
3940 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3941 {
3942         struct mem_cgroup *iter;
3943
3944         for_each_mem_cgroup_tree(iter, mem)
3945                 mem_cgroup_oom_notify_cb(iter);
3946 }
3947
3948 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3949         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3950 {
3951         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3952         struct mem_cgroup_thresholds *thresholds;
3953         struct mem_cgroup_threshold_ary *new;
3954         int type = MEMFILE_TYPE(cft->private);
3955         u64 threshold, usage;
3956         int i, size, ret;
3957
3958         ret = res_counter_memparse_write_strategy(args, &threshold);
3959         if (ret)
3960                 return ret;
3961
3962         mutex_lock(&memcg->thresholds_lock);
3963
3964         if (type == _MEM)
3965                 thresholds = &memcg->thresholds;
3966         else if (type == _MEMSWAP)
3967                 thresholds = &memcg->memsw_thresholds;
3968         else
3969                 BUG();
3970
3971         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3972
3973         /* Check if a threshold crossed before adding a new one */
3974         if (thresholds->primary)
3975                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3976
3977         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3978
3979         /* Allocate memory for new array of thresholds */
3980         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3981                         GFP_KERNEL);
3982         if (!new) {
3983                 ret = -ENOMEM;
3984                 goto unlock;
3985         }
3986         new->size = size;
3987
3988         /* Copy thresholds (if any) to new array */
3989         if (thresholds->primary) {
3990                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3991                                 sizeof(struct mem_cgroup_threshold));
3992         }
3993
3994         /* Add new threshold */
3995         new->entries[size - 1].eventfd = eventfd;
3996         new->entries[size - 1].threshold = threshold;
3997
3998         /* Sort thresholds. Registering of new threshold isn't time-critical */
3999         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4000                         compare_thresholds, NULL);
4001
4002         /* Find current threshold */
4003         new->current_threshold = -1;
4004         for (i = 0; i < size; i++) {
4005                 if (new->entries[i].threshold < usage) {
4006                         /*
4007                          * new->current_threshold will not be used until
4008                          * rcu_assign_pointer(), so it's safe to increment
4009                          * it here.
4010                          */
4011                         ++new->current_threshold;
4012                 }
4013         }
4014
4015         /* Free old spare buffer and save old primary buffer as spare */
4016         kfree(thresholds->spare);
4017         thresholds->spare = thresholds->primary;
4018
4019         rcu_assign_pointer(thresholds->primary, new);
4020
4021         /* To be sure that nobody uses thresholds */
4022         synchronize_rcu();
4023
4024 unlock:
4025         mutex_unlock(&memcg->thresholds_lock);
4026
4027         return ret;
4028 }
4029
4030 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4031         struct cftype *cft, struct eventfd_ctx *eventfd)
4032 {
4033         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4034         struct mem_cgroup_thresholds *thresholds;
4035         struct mem_cgroup_threshold_ary *new;
4036         int type = MEMFILE_TYPE(cft->private);
4037         u64 usage;
4038         int i, j, size;
4039
4040         mutex_lock(&memcg->thresholds_lock);
4041         if (type == _MEM)
4042                 thresholds = &memcg->thresholds;
4043         else if (type == _MEMSWAP)
4044                 thresholds = &memcg->memsw_thresholds;
4045         else
4046                 BUG();
4047
4048         /*
4049          * Something went wrong if we trying to unregister a threshold
4050          * if we don't have thresholds
4051          */
4052         BUG_ON(!thresholds);
4053
4054         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4055
4056         /* Check if a threshold crossed before removing */
4057         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4058
4059         /* Calculate new number of threshold */
4060         size = 0;
4061         for (i = 0; i < thresholds->primary->size; i++) {
4062                 if (thresholds->primary->entries[i].eventfd != eventfd)
4063                         size++;
4064         }
4065
4066         new = thresholds->spare;
4067
4068         /* Set thresholds array to NULL if we don't have thresholds */
4069         if (!size) {
4070                 kfree(new);
4071                 new = NULL;
4072                 goto swap_buffers;
4073         }
4074
4075         new->size = size;
4076
4077         /* Copy thresholds and find current threshold */
4078         new->current_threshold = -1;
4079         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4080                 if (thresholds->primary->entries[i].eventfd == eventfd)
4081                         continue;
4082
4083                 new->entries[j] = thresholds->primary->entries[i];
4084                 if (new->entries[j].threshold < usage) {
4085                         /*
4086                          * new->current_threshold will not be used
4087                          * until rcu_assign_pointer(), so it's safe to increment
4088                          * it here.
4089                          */
4090                         ++new->current_threshold;
4091                 }
4092                 j++;
4093         }
4094
4095 swap_buffers:
4096         /* Swap primary and spare array */
4097         thresholds->spare = thresholds->primary;
4098         rcu_assign_pointer(thresholds->primary, new);
4099
4100         /* To be sure that nobody uses thresholds */
4101         synchronize_rcu();
4102
4103         mutex_unlock(&memcg->thresholds_lock);
4104 }
4105
4106 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4107         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4108 {
4109         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4110         struct mem_cgroup_eventfd_list *event;
4111         int type = MEMFILE_TYPE(cft->private);
4112
4113         BUG_ON(type != _OOM_TYPE);
4114         event = kmalloc(sizeof(*event), GFP_KERNEL);
4115         if (!event)
4116                 return -ENOMEM;
4117
4118         mutex_lock(&memcg_oom_mutex);
4119
4120         event->eventfd = eventfd;
4121         list_add(&event->list, &memcg->oom_notify);
4122
4123         /* already in OOM ? */
4124         if (atomic_read(&memcg->oom_lock))
4125                 eventfd_signal(eventfd, 1);
4126         mutex_unlock(&memcg_oom_mutex);
4127
4128         return 0;
4129 }
4130
4131 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4132         struct cftype *cft, struct eventfd_ctx *eventfd)
4133 {
4134         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4135         struct mem_cgroup_eventfd_list *ev, *tmp;
4136         int type = MEMFILE_TYPE(cft->private);
4137
4138         BUG_ON(type != _OOM_TYPE);
4139
4140         mutex_lock(&memcg_oom_mutex);
4141
4142         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4143                 if (ev->eventfd == eventfd) {
4144                         list_del(&ev->list);
4145                         kfree(ev);
4146                 }
4147         }
4148
4149         mutex_unlock(&memcg_oom_mutex);
4150 }
4151
4152 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4153         struct cftype *cft,  struct cgroup_map_cb *cb)
4154 {
4155         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4156
4157         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4158
4159         if (atomic_read(&mem->oom_lock))
4160                 cb->fill(cb, "under_oom", 1);
4161         else
4162                 cb->fill(cb, "under_oom", 0);
4163         return 0;
4164 }
4165
4166 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4167         struct cftype *cft, u64 val)
4168 {
4169         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4170         struct mem_cgroup *parent;
4171
4172         /* cannot set to root cgroup and only 0 and 1 are allowed */
4173         if (!cgrp->parent || !((val == 0) || (val == 1)))
4174                 return -EINVAL;
4175
4176         parent = mem_cgroup_from_cont(cgrp->parent);
4177
4178         cgroup_lock();
4179         /* oom-kill-disable is a flag for subhierarchy. */
4180         if ((parent->use_hierarchy) ||
4181             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4182                 cgroup_unlock();
4183                 return -EINVAL;
4184         }
4185         mem->oom_kill_disable = val;
4186         if (!val)
4187                 memcg_oom_recover(mem);
4188         cgroup_unlock();
4189         return 0;
4190 }
4191
4192 static struct cftype mem_cgroup_files[] = {
4193         {
4194                 .name = "usage_in_bytes",
4195                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4196                 .read_u64 = mem_cgroup_read,
4197                 .register_event = mem_cgroup_usage_register_event,
4198                 .unregister_event = mem_cgroup_usage_unregister_event,
4199         },
4200         {
4201                 .name = "max_usage_in_bytes",
4202                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4203                 .trigger = mem_cgroup_reset,
4204                 .read_u64 = mem_cgroup_read,
4205         },
4206         {
4207                 .name = "limit_in_bytes",
4208                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4209                 .write_string = mem_cgroup_write,
4210                 .read_u64 = mem_cgroup_read,
4211         },
4212         {
4213                 .name = "soft_limit_in_bytes",
4214                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4215                 .write_string = mem_cgroup_write,
4216                 .read_u64 = mem_cgroup_read,
4217         },
4218         {
4219                 .name = "failcnt",
4220                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4221                 .trigger = mem_cgroup_reset,
4222                 .read_u64 = mem_cgroup_read,
4223         },
4224         {
4225                 .name = "stat",
4226                 .read_map = mem_control_stat_show,
4227         },
4228         {
4229                 .name = "force_empty",
4230                 .trigger = mem_cgroup_force_empty_write,
4231         },
4232         {
4233                 .name = "use_hierarchy",
4234                 .write_u64 = mem_cgroup_hierarchy_write,
4235                 .read_u64 = mem_cgroup_hierarchy_read,
4236         },
4237         {
4238                 .name = "swappiness",
4239                 .read_u64 = mem_cgroup_swappiness_read,
4240                 .write_u64 = mem_cgroup_swappiness_write,
4241         },
4242         {
4243                 .name = "move_charge_at_immigrate",
4244                 .read_u64 = mem_cgroup_move_charge_read,
4245                 .write_u64 = mem_cgroup_move_charge_write,
4246         },
4247         {
4248                 .name = "oom_control",
4249                 .read_map = mem_cgroup_oom_control_read,
4250                 .write_u64 = mem_cgroup_oom_control_write,
4251                 .register_event = mem_cgroup_oom_register_event,
4252                 .unregister_event = mem_cgroup_oom_unregister_event,
4253                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4254         },
4255 };
4256
4257 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4258 static struct cftype memsw_cgroup_files[] = {
4259         {
4260                 .name = "memsw.usage_in_bytes",
4261                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4262                 .read_u64 = mem_cgroup_read,
4263                 .register_event = mem_cgroup_usage_register_event,
4264                 .unregister_event = mem_cgroup_usage_unregister_event,
4265         },
4266         {
4267                 .name = "memsw.max_usage_in_bytes",
4268                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4269                 .trigger = mem_cgroup_reset,
4270                 .read_u64 = mem_cgroup_read,
4271         },
4272         {
4273                 .name = "memsw.limit_in_bytes",
4274                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4275                 .write_string = mem_cgroup_write,
4276                 .read_u64 = mem_cgroup_read,
4277         },
4278         {
4279                 .name = "memsw.failcnt",
4280                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4281                 .trigger = mem_cgroup_reset,
4282                 .read_u64 = mem_cgroup_read,
4283         },
4284 };
4285
4286 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4287 {
4288         if (!do_swap_account)
4289                 return 0;
4290         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4291                                 ARRAY_SIZE(memsw_cgroup_files));
4292 };
4293 #else
4294 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4295 {
4296         return 0;
4297 }
4298 #endif
4299
4300 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4301 {
4302         struct mem_cgroup_per_node *pn;
4303         struct mem_cgroup_per_zone *mz;
4304         enum lru_list l;
4305         int zone, tmp = node;
4306         /*
4307          * This routine is called against possible nodes.
4308          * But it's BUG to call kmalloc() against offline node.
4309          *
4310          * TODO: this routine can waste much memory for nodes which will
4311          *       never be onlined. It's better to use memory hotplug callback
4312          *       function.
4313          */
4314         if (!node_state(node, N_NORMAL_MEMORY))
4315                 tmp = -1;
4316         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4317         if (!pn)
4318                 return 1;
4319
4320         mem->info.nodeinfo[node] = pn;
4321         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4322                 mz = &pn->zoneinfo[zone];
4323                 for_each_lru(l)
4324                         INIT_LIST_HEAD(&mz->lists[l]);
4325                 mz->usage_in_excess = 0;
4326                 mz->on_tree = false;
4327                 mz->mem = mem;
4328         }
4329         return 0;
4330 }
4331
4332 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4333 {
4334         kfree(mem->info.nodeinfo[node]);
4335 }
4336
4337 static struct mem_cgroup *mem_cgroup_alloc(void)
4338 {
4339         struct mem_cgroup *mem;
4340         int size = sizeof(struct mem_cgroup);
4341
4342         /* Can be very big if MAX_NUMNODES is very big */
4343         if (size < PAGE_SIZE)
4344                 mem = kzalloc(size, GFP_KERNEL);
4345         else
4346                 mem = vzalloc(size);
4347
4348         if (!mem)
4349                 return NULL;
4350
4351         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4352         if (!mem->stat)
4353                 goto out_free;
4354         spin_lock_init(&mem->pcp_counter_lock);
4355         return mem;
4356
4357 out_free:
4358         if (size < PAGE_SIZE)
4359                 kfree(mem);
4360         else
4361                 vfree(mem);
4362         return NULL;
4363 }
4364
4365 /*
4366  * At destroying mem_cgroup, references from swap_cgroup can remain.
4367  * (scanning all at force_empty is too costly...)
4368  *
4369  * Instead of clearing all references at force_empty, we remember
4370  * the number of reference from swap_cgroup and free mem_cgroup when
4371  * it goes down to 0.
4372  *
4373  * Removal of cgroup itself succeeds regardless of refs from swap.
4374  */
4375
4376 static void __mem_cgroup_free(struct mem_cgroup *mem)
4377 {
4378         int node;
4379
4380         mem_cgroup_remove_from_trees(mem);
4381         free_css_id(&mem_cgroup_subsys, &mem->css);
4382
4383         for_each_node_state(node, N_POSSIBLE)
4384                 free_mem_cgroup_per_zone_info(mem, node);
4385
4386         free_percpu(mem->stat);
4387         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4388                 kfree(mem);
4389         else
4390                 vfree(mem);
4391 }
4392
4393 static void mem_cgroup_get(struct mem_cgroup *mem)
4394 {
4395         atomic_inc(&mem->refcnt);
4396 }
4397
4398 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4399 {
4400         if (atomic_sub_and_test(count, &mem->refcnt)) {
4401                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4402                 __mem_cgroup_free(mem);
4403                 if (parent)
4404                         mem_cgroup_put(parent);
4405         }
4406 }
4407
4408 static void mem_cgroup_put(struct mem_cgroup *mem)
4409 {
4410         __mem_cgroup_put(mem, 1);
4411 }
4412
4413 /*
4414  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4415  */
4416 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4417 {
4418         if (!mem->res.parent)
4419                 return NULL;
4420         return mem_cgroup_from_res_counter(mem->res.parent, res);
4421 }
4422
4423 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4424 static void __init enable_swap_cgroup(void)
4425 {
4426         if (!mem_cgroup_disabled() && really_do_swap_account)
4427                 do_swap_account = 1;
4428 }
4429 #else
4430 static void __init enable_swap_cgroup(void)
4431 {
4432 }
4433 #endif
4434
4435 static int mem_cgroup_soft_limit_tree_init(void)
4436 {
4437         struct mem_cgroup_tree_per_node *rtpn;
4438         struct mem_cgroup_tree_per_zone *rtpz;
4439         int tmp, node, zone;
4440
4441         for_each_node_state(node, N_POSSIBLE) {
4442                 tmp = node;
4443                 if (!node_state(node, N_NORMAL_MEMORY))
4444                         tmp = -1;
4445                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4446                 if (!rtpn)
4447                         return 1;
4448
4449                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4450
4451                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4452                         rtpz = &rtpn->rb_tree_per_zone[zone];
4453                         rtpz->rb_root = RB_ROOT;
4454                         spin_lock_init(&rtpz->lock);
4455                 }
4456         }
4457         return 0;
4458 }
4459
4460 static struct cgroup_subsys_state * __ref
4461 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4462 {
4463         struct mem_cgroup *mem, *parent;
4464         long error = -ENOMEM;
4465         int node;
4466
4467         mem = mem_cgroup_alloc();
4468         if (!mem)
4469                 return ERR_PTR(error);
4470
4471         for_each_node_state(node, N_POSSIBLE)
4472                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4473                         goto free_out;
4474
4475         /* root ? */
4476         if (cont->parent == NULL) {
4477                 int cpu;
4478                 enable_swap_cgroup();
4479                 parent = NULL;
4480                 root_mem_cgroup = mem;
4481                 if (mem_cgroup_soft_limit_tree_init())
4482                         goto free_out;
4483                 for_each_possible_cpu(cpu) {
4484                         struct memcg_stock_pcp *stock =
4485                                                 &per_cpu(memcg_stock, cpu);
4486                         INIT_WORK(&stock->work, drain_local_stock);
4487                 }
4488                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4489         } else {
4490                 parent = mem_cgroup_from_cont(cont->parent);
4491                 mem->use_hierarchy = parent->use_hierarchy;
4492                 mem->oom_kill_disable = parent->oom_kill_disable;
4493         }
4494
4495         if (parent && parent->use_hierarchy) {
4496                 res_counter_init(&mem->res, &parent->res);
4497                 res_counter_init(&mem->memsw, &parent->memsw);
4498                 /*
4499                  * We increment refcnt of the parent to ensure that we can
4500                  * safely access it on res_counter_charge/uncharge.
4501                  * This refcnt will be decremented when freeing this
4502                  * mem_cgroup(see mem_cgroup_put).
4503                  */
4504                 mem_cgroup_get(parent);
4505         } else {
4506                 res_counter_init(&mem->res, NULL);
4507                 res_counter_init(&mem->memsw, NULL);
4508         }
4509         mem->last_scanned_child = 0;
4510         INIT_LIST_HEAD(&mem->oom_notify);
4511
4512         if (parent)
4513                 mem->swappiness = get_swappiness(parent);
4514         atomic_set(&mem->refcnt, 1);
4515         mem->move_charge_at_immigrate = 0;
4516         mutex_init(&mem->thresholds_lock);
4517         return &mem->css;
4518 free_out:
4519         __mem_cgroup_free(mem);
4520         root_mem_cgroup = NULL;
4521         return ERR_PTR(error);
4522 }
4523
4524 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4525                                         struct cgroup *cont)
4526 {
4527         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4528
4529         return mem_cgroup_force_empty(mem, false);
4530 }
4531
4532 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4533                                 struct cgroup *cont)
4534 {
4535         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4536
4537         mem_cgroup_put(mem);
4538 }
4539
4540 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4541                                 struct cgroup *cont)
4542 {
4543         int ret;
4544
4545         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4546                                 ARRAY_SIZE(mem_cgroup_files));
4547
4548         if (!ret)
4549                 ret = register_memsw_files(cont, ss);
4550         return ret;
4551 }
4552
4553 #ifdef CONFIG_MMU
4554 /* Handlers for move charge at task migration. */
4555 #define PRECHARGE_COUNT_AT_ONCE 256
4556 static int mem_cgroup_do_precharge(unsigned long count)
4557 {
4558         int ret = 0;
4559         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4560         struct mem_cgroup *mem = mc.to;
4561
4562         if (mem_cgroup_is_root(mem)) {
4563                 mc.precharge += count;
4564                 /* we don't need css_get for root */
4565                 return ret;
4566         }
4567         /* try to charge at once */
4568         if (count > 1) {
4569                 struct res_counter *dummy;
4570                 /*
4571                  * "mem" cannot be under rmdir() because we've already checked
4572                  * by cgroup_lock_live_cgroup() that it is not removed and we
4573                  * are still under the same cgroup_mutex. So we can postpone
4574                  * css_get().
4575                  */
4576                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4577                         goto one_by_one;
4578                 if (do_swap_account && res_counter_charge(&mem->memsw,
4579                                                 PAGE_SIZE * count, &dummy)) {
4580                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4581                         goto one_by_one;
4582                 }
4583                 mc.precharge += count;
4584                 return ret;
4585         }
4586 one_by_one:
4587         /* fall back to one by one charge */
4588         while (count--) {
4589                 if (signal_pending(current)) {
4590                         ret = -EINTR;
4591                         break;
4592                 }
4593                 if (!batch_count--) {
4594                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4595                         cond_resched();
4596                 }
4597                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4598                                               PAGE_SIZE);
4599                 if (ret || !mem)
4600                         /* mem_cgroup_clear_mc() will do uncharge later */
4601                         return -ENOMEM;
4602                 mc.precharge++;
4603         }
4604         return ret;
4605 }
4606
4607 /**
4608  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4609  * @vma: the vma the pte to be checked belongs
4610  * @addr: the address corresponding to the pte to be checked
4611  * @ptent: the pte to be checked
4612  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4613  *
4614  * Returns
4615  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4616  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4617  *     move charge. if @target is not NULL, the page is stored in target->page
4618  *     with extra refcnt got(Callers should handle it).
4619  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4620  *     target for charge migration. if @target is not NULL, the entry is stored
4621  *     in target->ent.
4622  *
4623  * Called with pte lock held.
4624  */
4625 union mc_target {
4626         struct page     *page;
4627         swp_entry_t     ent;
4628 };
4629
4630 enum mc_target_type {
4631         MC_TARGET_NONE, /* not used */
4632         MC_TARGET_PAGE,
4633         MC_TARGET_SWAP,
4634 };
4635
4636 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4637                                                 unsigned long addr, pte_t ptent)
4638 {
4639         struct page *page = vm_normal_page(vma, addr, ptent);
4640
4641         if (!page || !page_mapped(page))
4642                 return NULL;
4643         if (PageAnon(page)) {
4644                 /* we don't move shared anon */
4645                 if (!move_anon() || page_mapcount(page) > 2)
4646                         return NULL;
4647         } else if (!move_file())
4648                 /* we ignore mapcount for file pages */
4649                 return NULL;
4650         if (!get_page_unless_zero(page))
4651                 return NULL;
4652
4653         return page;
4654 }
4655
4656 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4657                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4658 {
4659         int usage_count;
4660         struct page *page = NULL;
4661         swp_entry_t ent = pte_to_swp_entry(ptent);
4662
4663         if (!move_anon() || non_swap_entry(ent))
4664                 return NULL;
4665         usage_count = mem_cgroup_count_swap_user(ent, &page);
4666         if (usage_count > 1) { /* we don't move shared anon */
4667                 if (page)
4668                         put_page(page);
4669                 return NULL;
4670         }
4671         if (do_swap_account)
4672                 entry->val = ent.val;
4673
4674         return page;
4675 }
4676
4677 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4678                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4679 {
4680         struct page *page = NULL;
4681         struct inode *inode;
4682         struct address_space *mapping;
4683         pgoff_t pgoff;
4684
4685         if (!vma->vm_file) /* anonymous vma */
4686                 return NULL;
4687         if (!move_file())
4688                 return NULL;
4689
4690         inode = vma->vm_file->f_path.dentry->d_inode;
4691         mapping = vma->vm_file->f_mapping;
4692         if (pte_none(ptent))
4693                 pgoff = linear_page_index(vma, addr);
4694         else /* pte_file(ptent) is true */
4695                 pgoff = pte_to_pgoff(ptent);
4696
4697         /* page is moved even if it's not RSS of this task(page-faulted). */
4698         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4699                 page = find_get_page(mapping, pgoff);
4700         } else { /* shmem/tmpfs file. we should take account of swap too. */
4701                 swp_entry_t ent;
4702                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4703                 if (do_swap_account)
4704                         entry->val = ent.val;
4705         }
4706
4707         return page;
4708 }
4709
4710 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4711                 unsigned long addr, pte_t ptent, union mc_target *target)
4712 {
4713         struct page *page = NULL;
4714         struct page_cgroup *pc;
4715         int ret = 0;
4716         swp_entry_t ent = { .val = 0 };
4717
4718         if (pte_present(ptent))
4719                 page = mc_handle_present_pte(vma, addr, ptent);
4720         else if (is_swap_pte(ptent))
4721                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4722         else if (pte_none(ptent) || pte_file(ptent))
4723                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4724
4725         if (!page && !ent.val)
4726                 return 0;
4727         if (page) {
4728                 pc = lookup_page_cgroup(page);
4729                 /*
4730                  * Do only loose check w/o page_cgroup lock.
4731                  * mem_cgroup_move_account() checks the pc is valid or not under
4732                  * the lock.
4733                  */
4734                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4735                         ret = MC_TARGET_PAGE;
4736                         if (target)
4737                                 target->page = page;
4738                 }
4739                 if (!ret || !target)
4740                         put_page(page);
4741         }
4742         /* There is a swap entry and a page doesn't exist or isn't charged */
4743         if (ent.val && !ret &&
4744                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4745                 ret = MC_TARGET_SWAP;
4746                 if (target)
4747                         target->ent = ent;
4748         }
4749         return ret;
4750 }
4751
4752 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4753                                         unsigned long addr, unsigned long end,
4754                                         struct mm_walk *walk)
4755 {
4756         struct vm_area_struct *vma = walk->private;
4757         pte_t *pte;
4758         spinlock_t *ptl;
4759
4760         split_huge_page_pmd(walk->mm, pmd);
4761
4762         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4763         for (; addr != end; pte++, addr += PAGE_SIZE)
4764                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4765                         mc.precharge++; /* increment precharge temporarily */
4766         pte_unmap_unlock(pte - 1, ptl);
4767         cond_resched();
4768
4769         return 0;
4770 }
4771
4772 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4773 {
4774         unsigned long precharge;
4775         struct vm_area_struct *vma;
4776
4777         down_read(&mm->mmap_sem);
4778         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4779                 struct mm_walk mem_cgroup_count_precharge_walk = {
4780                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4781                         .mm = mm,
4782                         .private = vma,
4783                 };
4784                 if (is_vm_hugetlb_page(vma))
4785                         continue;
4786                 walk_page_range(vma->vm_start, vma->vm_end,
4787                                         &mem_cgroup_count_precharge_walk);
4788         }
4789         up_read(&mm->mmap_sem);
4790
4791         precharge = mc.precharge;
4792         mc.precharge = 0;
4793
4794         return precharge;
4795 }
4796
4797 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4798 {
4799         unsigned long precharge = mem_cgroup_count_precharge(mm);
4800
4801         VM_BUG_ON(mc.moving_task);
4802         mc.moving_task = current;
4803         return mem_cgroup_do_precharge(precharge);
4804 }
4805
4806 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4807 static void __mem_cgroup_clear_mc(void)
4808 {
4809         struct mem_cgroup *from = mc.from;
4810         struct mem_cgroup *to = mc.to;
4811
4812         /* we must uncharge all the leftover precharges from mc.to */
4813         if (mc.precharge) {
4814                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4815                 mc.precharge = 0;
4816         }
4817         /*
4818          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4819          * we must uncharge here.
4820          */
4821         if (mc.moved_charge) {
4822                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4823                 mc.moved_charge = 0;
4824         }
4825         /* we must fixup refcnts and charges */
4826         if (mc.moved_swap) {
4827                 /* uncharge swap account from the old cgroup */
4828                 if (!mem_cgroup_is_root(mc.from))
4829                         res_counter_uncharge(&mc.from->memsw,
4830                                                 PAGE_SIZE * mc.moved_swap);
4831                 __mem_cgroup_put(mc.from, mc.moved_swap);
4832
4833                 if (!mem_cgroup_is_root(mc.to)) {
4834                         /*
4835                          * we charged both to->res and to->memsw, so we should
4836                          * uncharge to->res.
4837                          */
4838                         res_counter_uncharge(&mc.to->res,
4839                                                 PAGE_SIZE * mc.moved_swap);
4840                 }
4841                 /* we've already done mem_cgroup_get(mc.to) */
4842                 mc.moved_swap = 0;
4843         }
4844         memcg_oom_recover(from);
4845         memcg_oom_recover(to);
4846         wake_up_all(&mc.waitq);
4847 }
4848
4849 static void mem_cgroup_clear_mc(void)
4850 {
4851         struct mem_cgroup *from = mc.from;
4852
4853         /*
4854          * we must clear moving_task before waking up waiters at the end of
4855          * task migration.
4856          */
4857         mc.moving_task = NULL;
4858         __mem_cgroup_clear_mc();
4859         spin_lock(&mc.lock);
4860         mc.from = NULL;
4861         mc.to = NULL;
4862         spin_unlock(&mc.lock);
4863         mem_cgroup_end_move(from);
4864 }
4865
4866 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4867                                 struct cgroup *cgroup,
4868                                 struct task_struct *p,
4869                                 bool threadgroup)
4870 {
4871         int ret = 0;
4872         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4873
4874         if (mem->move_charge_at_immigrate) {
4875                 struct mm_struct *mm;
4876                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4877
4878                 VM_BUG_ON(from == mem);
4879
4880                 mm = get_task_mm(p);
4881                 if (!mm)
4882                         return 0;
4883                 /* We move charges only when we move a owner of the mm */
4884                 if (mm->owner == p) {
4885                         VM_BUG_ON(mc.from);
4886                         VM_BUG_ON(mc.to);
4887                         VM_BUG_ON(mc.precharge);
4888                         VM_BUG_ON(mc.moved_charge);
4889                         VM_BUG_ON(mc.moved_swap);
4890                         mem_cgroup_start_move(from);
4891                         spin_lock(&mc.lock);
4892                         mc.from = from;
4893                         mc.to = mem;
4894                         spin_unlock(&mc.lock);
4895                         /* We set mc.moving_task later */
4896
4897                         ret = mem_cgroup_precharge_mc(mm);
4898                         if (ret)
4899                                 mem_cgroup_clear_mc();
4900                 }
4901                 mmput(mm);
4902         }
4903         return ret;
4904 }
4905
4906 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4907                                 struct cgroup *cgroup,
4908                                 struct task_struct *p,
4909                                 bool threadgroup)
4910 {
4911         mem_cgroup_clear_mc();
4912 }
4913
4914 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4915                                 unsigned long addr, unsigned long end,
4916                                 struct mm_walk *walk)
4917 {
4918         int ret = 0;
4919         struct vm_area_struct *vma = walk->private;
4920         pte_t *pte;
4921         spinlock_t *ptl;
4922
4923         split_huge_page_pmd(walk->mm, pmd);
4924 retry:
4925         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4926         for (; addr != end; addr += PAGE_SIZE) {
4927                 pte_t ptent = *(pte++);
4928                 union mc_target target;
4929                 int type;
4930                 struct page *page;
4931                 struct page_cgroup *pc;
4932                 swp_entry_t ent;
4933
4934                 if (!mc.precharge)
4935                         break;
4936
4937                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4938                 switch (type) {
4939                 case MC_TARGET_PAGE:
4940                         page = target.page;
4941                         if (isolate_lru_page(page))
4942                                 goto put;
4943                         pc = lookup_page_cgroup(page);
4944                         if (!mem_cgroup_move_account(page, pc,
4945                                         mc.from, mc.to, false, PAGE_SIZE)) {
4946                                 mc.precharge--;
4947                                 /* we uncharge from mc.from later. */
4948                                 mc.moved_charge++;
4949                         }
4950                         putback_lru_page(page);
4951 put:                    /* is_target_pte_for_mc() gets the page */
4952                         put_page(page);
4953                         break;
4954                 case MC_TARGET_SWAP:
4955                         ent = target.ent;
4956                         if (!mem_cgroup_move_swap_account(ent,
4957                                                 mc.from, mc.to, false)) {
4958                                 mc.precharge--;
4959                                 /* we fixup refcnts and charges later. */
4960                                 mc.moved_swap++;
4961                         }
4962                         break;
4963                 default:
4964                         break;
4965                 }
4966         }
4967         pte_unmap_unlock(pte - 1, ptl);
4968         cond_resched();
4969
4970         if (addr != end) {
4971                 /*
4972                  * We have consumed all precharges we got in can_attach().
4973                  * We try charge one by one, but don't do any additional
4974                  * charges to mc.to if we have failed in charge once in attach()
4975                  * phase.
4976                  */
4977                 ret = mem_cgroup_do_precharge(1);
4978                 if (!ret)
4979                         goto retry;
4980         }
4981
4982         return ret;
4983 }
4984
4985 static void mem_cgroup_move_charge(struct mm_struct *mm)
4986 {
4987         struct vm_area_struct *vma;
4988
4989         lru_add_drain_all();
4990 retry:
4991         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4992                 /*
4993                  * Someone who are holding the mmap_sem might be waiting in
4994                  * waitq. So we cancel all extra charges, wake up all waiters,
4995                  * and retry. Because we cancel precharges, we might not be able
4996                  * to move enough charges, but moving charge is a best-effort
4997                  * feature anyway, so it wouldn't be a big problem.
4998                  */
4999                 __mem_cgroup_clear_mc();
5000                 cond_resched();
5001                 goto retry;
5002         }
5003         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5004                 int ret;
5005                 struct mm_walk mem_cgroup_move_charge_walk = {
5006                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5007                         .mm = mm,
5008                         .private = vma,
5009                 };
5010                 if (is_vm_hugetlb_page(vma))
5011                         continue;
5012                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5013                                                 &mem_cgroup_move_charge_walk);
5014                 if (ret)
5015                         /*
5016                          * means we have consumed all precharges and failed in
5017                          * doing additional charge. Just abandon here.
5018                          */
5019                         break;
5020         }
5021         up_read(&mm->mmap_sem);
5022 }
5023
5024 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5025                                 struct cgroup *cont,
5026                                 struct cgroup *old_cont,
5027                                 struct task_struct *p,
5028                                 bool threadgroup)
5029 {
5030         struct mm_struct *mm;
5031
5032         if (!mc.to)
5033                 /* no need to move charge */
5034                 return;
5035
5036         mm = get_task_mm(p);
5037         if (mm) {
5038                 mem_cgroup_move_charge(mm);
5039                 mmput(mm);
5040         }
5041         mem_cgroup_clear_mc();
5042 }
5043 #else   /* !CONFIG_MMU */
5044 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5045                                 struct cgroup *cgroup,
5046                                 struct task_struct *p,
5047                                 bool threadgroup)
5048 {
5049         return 0;
5050 }
5051 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5052                                 struct cgroup *cgroup,
5053                                 struct task_struct *p,
5054                                 bool threadgroup)
5055 {
5056 }
5057 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5058                                 struct cgroup *cont,
5059                                 struct cgroup *old_cont,
5060                                 struct task_struct *p,
5061                                 bool threadgroup)
5062 {
5063 }
5064 #endif
5065
5066 struct cgroup_subsys mem_cgroup_subsys = {
5067         .name = "memory",
5068         .subsys_id = mem_cgroup_subsys_id,
5069         .create = mem_cgroup_create,
5070         .pre_destroy = mem_cgroup_pre_destroy,
5071         .destroy = mem_cgroup_destroy,
5072         .populate = mem_cgroup_populate,
5073         .can_attach = mem_cgroup_can_attach,
5074         .cancel_attach = mem_cgroup_cancel_attach,
5075         .attach = mem_cgroup_move_task,
5076         .early_init = 0,
5077         .use_id = 1,
5078 };
5079
5080 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5081 static int __init enable_swap_account(char *s)
5082 {
5083         /* consider enabled if no parameter or 1 is given */
5084         if (!(*s) || !strcmp(s, "=1"))
5085                 really_do_swap_account = 1;
5086         else if (!strcmp(s, "=0"))
5087                 really_do_swap_account = 0;
5088         return 1;
5089 }
5090 __setup("swapaccount", enable_swap_account);
5091
5092 static int __init disable_swap_account(char *s)
5093 {
5094         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5095         enable_swap_account("=0");
5096         return 1;
5097 }
5098 __setup("noswapaccount", disable_swap_account);
5099 #endif