memcg: convert per-cpu stock from bytes to page granularity
[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         unsigned int nr_pages;
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, one page 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->nr_pages)
1672                 stock->nr_pages--;
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->nr_pages) {
1687                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1688
1689                 res_counter_uncharge(&old->res, bytes);
1690                 if (do_swap_account)
1691                         res_counter_uncharge(&old->memsw, bytes);
1692                 stock->nr_pages = 0;
1693         }
1694         stock->cached = NULL;
1695 }
1696
1697 /*
1698  * This must be called under preempt disabled or must be called by
1699  * a thread which is pinned to local cpu.
1700  */
1701 static void drain_local_stock(struct work_struct *dummy)
1702 {
1703         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1704         drain_stock(stock);
1705 }
1706
1707 /*
1708  * Cache charges(val) which is from res_counter, to local per_cpu area.
1709  * This will be consumed by consume_stock() function, later.
1710  */
1711 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
1712 {
1713         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1714
1715         if (stock->cached != mem) { /* reset if necessary */
1716                 drain_stock(stock);
1717                 stock->cached = mem;
1718         }
1719         stock->nr_pages += nr_pages;
1720         put_cpu_var(memcg_stock);
1721 }
1722
1723 /*
1724  * Tries to drain stocked charges in other cpus. This function is asynchronous
1725  * and just put a work per cpu for draining localy on each cpu. Caller can
1726  * expects some charges will be back to res_counter later but cannot wait for
1727  * it.
1728  */
1729 static void drain_all_stock_async(void)
1730 {
1731         int cpu;
1732         /* This function is for scheduling "drain" in asynchronous way.
1733          * The result of "drain" is not directly handled by callers. Then,
1734          * if someone is calling drain, we don't have to call drain more.
1735          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1736          * there is a race. We just do loose check here.
1737          */
1738         if (atomic_read(&memcg_drain_count))
1739                 return;
1740         /* Notify other cpus that system-wide "drain" is running */
1741         atomic_inc(&memcg_drain_count);
1742         get_online_cpus();
1743         for_each_online_cpu(cpu) {
1744                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1745                 schedule_work_on(cpu, &stock->work);
1746         }
1747         put_online_cpus();
1748         atomic_dec(&memcg_drain_count);
1749         /* We don't wait for flush_work */
1750 }
1751
1752 /* This is a synchronous drain interface. */
1753 static void drain_all_stock_sync(void)
1754 {
1755         /* called when force_empty is called */
1756         atomic_inc(&memcg_drain_count);
1757         schedule_on_each_cpu(drain_local_stock);
1758         atomic_dec(&memcg_drain_count);
1759 }
1760
1761 /*
1762  * This function drains percpu counter value from DEAD cpu and
1763  * move it to local cpu. Note that this function can be preempted.
1764  */
1765 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1766 {
1767         int i;
1768
1769         spin_lock(&mem->pcp_counter_lock);
1770         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1771                 s64 x = per_cpu(mem->stat->count[i], cpu);
1772
1773                 per_cpu(mem->stat->count[i], cpu) = 0;
1774                 mem->nocpu_base.count[i] += x;
1775         }
1776         /* need to clear ON_MOVE value, works as a kind of lock. */
1777         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1778         spin_unlock(&mem->pcp_counter_lock);
1779 }
1780
1781 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1782 {
1783         int idx = MEM_CGROUP_ON_MOVE;
1784
1785         spin_lock(&mem->pcp_counter_lock);
1786         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1787         spin_unlock(&mem->pcp_counter_lock);
1788 }
1789
1790 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1791                                         unsigned long action,
1792                                         void *hcpu)
1793 {
1794         int cpu = (unsigned long)hcpu;
1795         struct memcg_stock_pcp *stock;
1796         struct mem_cgroup *iter;
1797
1798         if ((action == CPU_ONLINE)) {
1799                 for_each_mem_cgroup_all(iter)
1800                         synchronize_mem_cgroup_on_move(iter, cpu);
1801                 return NOTIFY_OK;
1802         }
1803
1804         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1805                 return NOTIFY_OK;
1806
1807         for_each_mem_cgroup_all(iter)
1808                 mem_cgroup_drain_pcp_counter(iter, cpu);
1809
1810         stock = &per_cpu(memcg_stock, cpu);
1811         drain_stock(stock);
1812         return NOTIFY_OK;
1813 }
1814
1815
1816 /* See __mem_cgroup_try_charge() for details */
1817 enum {
1818         CHARGE_OK,              /* success */
1819         CHARGE_RETRY,           /* need to retry but retry is not bad */
1820         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1821         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1822         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1823 };
1824
1825 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1826                                 int csize, bool oom_check)
1827 {
1828         struct mem_cgroup *mem_over_limit;
1829         struct res_counter *fail_res;
1830         unsigned long flags = 0;
1831         int ret;
1832
1833         ret = res_counter_charge(&mem->res, csize, &fail_res);
1834
1835         if (likely(!ret)) {
1836                 if (!do_swap_account)
1837                         return CHARGE_OK;
1838                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1839                 if (likely(!ret))
1840                         return CHARGE_OK;
1841
1842                 res_counter_uncharge(&mem->res, csize);
1843                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1844                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1845         } else
1846                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1847         /*
1848          * csize can be either a huge page (HPAGE_SIZE), a batch of
1849          * regular pages (CHARGE_SIZE), or a single regular page
1850          * (PAGE_SIZE).
1851          *
1852          * Never reclaim on behalf of optional batching, retry with a
1853          * single page instead.
1854          */
1855         if (csize == CHARGE_SIZE)
1856                 return CHARGE_RETRY;
1857
1858         if (!(gfp_mask & __GFP_WAIT))
1859                 return CHARGE_WOULDBLOCK;
1860
1861         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1862                                               gfp_mask, flags);
1863         if (mem_cgroup_margin(mem_over_limit) >= csize)
1864                 return CHARGE_RETRY;
1865         /*
1866          * Even though the limit is exceeded at this point, reclaim
1867          * may have been able to free some pages.  Retry the charge
1868          * before killing the task.
1869          *
1870          * Only for regular pages, though: huge pages are rather
1871          * unlikely to succeed so close to the limit, and we fall back
1872          * to regular pages anyway in case of failure.
1873          */
1874         if (csize == PAGE_SIZE && ret)
1875                 return CHARGE_RETRY;
1876
1877         /*
1878          * At task move, charge accounts can be doubly counted. So, it's
1879          * better to wait until the end of task_move if something is going on.
1880          */
1881         if (mem_cgroup_wait_acct_move(mem_over_limit))
1882                 return CHARGE_RETRY;
1883
1884         /* If we don't need to call oom-killer at el, return immediately */
1885         if (!oom_check)
1886                 return CHARGE_NOMEM;
1887         /* check OOM */
1888         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1889                 return CHARGE_OOM_DIE;
1890
1891         return CHARGE_RETRY;
1892 }
1893
1894 /*
1895  * Unlike exported interface, "oom" parameter is added. if oom==true,
1896  * oom-killer can be invoked.
1897  */
1898 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1899                                    gfp_t gfp_mask,
1900                                    struct mem_cgroup **memcg, bool oom,
1901                                    int page_size)
1902 {
1903         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1904         struct mem_cgroup *mem = NULL;
1905         int ret;
1906         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1907
1908         /*
1909          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1910          * in system level. So, allow to go ahead dying process in addition to
1911          * MEMDIE process.
1912          */
1913         if (unlikely(test_thread_flag(TIF_MEMDIE)
1914                      || fatal_signal_pending(current)))
1915                 goto bypass;
1916
1917         /*
1918          * We always charge the cgroup the mm_struct belongs to.
1919          * The mm_struct's mem_cgroup changes on task migration if the
1920          * thread group leader migrates. It's possible that mm is not
1921          * set, if so charge the init_mm (happens for pagecache usage).
1922          */
1923         if (!*memcg && !mm)
1924                 goto bypass;
1925 again:
1926         if (*memcg) { /* css should be a valid one */
1927                 mem = *memcg;
1928                 VM_BUG_ON(css_is_removed(&mem->css));
1929                 if (mem_cgroup_is_root(mem))
1930                         goto done;
1931                 if (page_size == PAGE_SIZE && consume_stock(mem))
1932                         goto done;
1933                 css_get(&mem->css);
1934         } else {
1935                 struct task_struct *p;
1936
1937                 rcu_read_lock();
1938                 p = rcu_dereference(mm->owner);
1939                 /*
1940                  * Because we don't have task_lock(), "p" can exit.
1941                  * In that case, "mem" can point to root or p can be NULL with
1942                  * race with swapoff. Then, we have small risk of mis-accouning.
1943                  * But such kind of mis-account by race always happens because
1944                  * we don't have cgroup_mutex(). It's overkill and we allo that
1945                  * small race, here.
1946                  * (*) swapoff at el will charge against mm-struct not against
1947                  * task-struct. So, mm->owner can be NULL.
1948                  */
1949                 mem = mem_cgroup_from_task(p);
1950                 if (!mem || mem_cgroup_is_root(mem)) {
1951                         rcu_read_unlock();
1952                         goto done;
1953                 }
1954                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1955                         /*
1956                          * It seems dagerous to access memcg without css_get().
1957                          * But considering how consume_stok works, it's not
1958                          * necessary. If consume_stock success, some charges
1959                          * from this memcg are cached on this cpu. So, we
1960                          * don't need to call css_get()/css_tryget() before
1961                          * calling consume_stock().
1962                          */
1963                         rcu_read_unlock();
1964                         goto done;
1965                 }
1966                 /* after here, we may be blocked. we need to get refcnt */
1967                 if (!css_tryget(&mem->css)) {
1968                         rcu_read_unlock();
1969                         goto again;
1970                 }
1971                 rcu_read_unlock();
1972         }
1973
1974         do {
1975                 bool oom_check;
1976
1977                 /* If killed, bypass charge */
1978                 if (fatal_signal_pending(current)) {
1979                         css_put(&mem->css);
1980                         goto bypass;
1981                 }
1982
1983                 oom_check = false;
1984                 if (oom && !nr_oom_retries) {
1985                         oom_check = true;
1986                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1987                 }
1988
1989                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1990
1991                 switch (ret) {
1992                 case CHARGE_OK:
1993                         break;
1994                 case CHARGE_RETRY: /* not in OOM situation but retry */
1995                         csize = page_size;
1996                         css_put(&mem->css);
1997                         mem = NULL;
1998                         goto again;
1999                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2000                         css_put(&mem->css);
2001                         goto nomem;
2002                 case CHARGE_NOMEM: /* OOM routine works */
2003                         if (!oom) {
2004                                 css_put(&mem->css);
2005                                 goto nomem;
2006                         }
2007                         /* If oom, we never return -ENOMEM */
2008                         nr_oom_retries--;
2009                         break;
2010                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2011                         css_put(&mem->css);
2012                         goto bypass;
2013                 }
2014         } while (ret != CHARGE_OK);
2015
2016         if (csize > page_size)
2017                 refill_stock(mem, (csize - page_size) >> PAGE_SHIFT);
2018         css_put(&mem->css);
2019 done:
2020         *memcg = mem;
2021         return 0;
2022 nomem:
2023         *memcg = NULL;
2024         return -ENOMEM;
2025 bypass:
2026         *memcg = NULL;
2027         return 0;
2028 }
2029
2030 /*
2031  * Somemtimes we have to undo a charge we got by try_charge().
2032  * This function is for that and do uncharge, put css's refcnt.
2033  * gotten by try_charge().
2034  */
2035 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2036                                        unsigned int nr_pages)
2037 {
2038         if (!mem_cgroup_is_root(mem)) {
2039                 unsigned long bytes = nr_pages * PAGE_SIZE;
2040
2041                 res_counter_uncharge(&mem->res, bytes);
2042                 if (do_swap_account)
2043                         res_counter_uncharge(&mem->memsw, bytes);
2044         }
2045 }
2046
2047 /*
2048  * A helper function to get mem_cgroup from ID. must be called under
2049  * rcu_read_lock(). The caller must check css_is_removed() or some if
2050  * it's concern. (dropping refcnt from swap can be called against removed
2051  * memcg.)
2052  */
2053 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2054 {
2055         struct cgroup_subsys_state *css;
2056
2057         /* ID 0 is unused ID */
2058         if (!id)
2059                 return NULL;
2060         css = css_lookup(&mem_cgroup_subsys, id);
2061         if (!css)
2062                 return NULL;
2063         return container_of(css, struct mem_cgroup, css);
2064 }
2065
2066 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2067 {
2068         struct mem_cgroup *mem = NULL;
2069         struct page_cgroup *pc;
2070         unsigned short id;
2071         swp_entry_t ent;
2072
2073         VM_BUG_ON(!PageLocked(page));
2074
2075         pc = lookup_page_cgroup(page);
2076         lock_page_cgroup(pc);
2077         if (PageCgroupUsed(pc)) {
2078                 mem = pc->mem_cgroup;
2079                 if (mem && !css_tryget(&mem->css))
2080                         mem = NULL;
2081         } else if (PageSwapCache(page)) {
2082                 ent.val = page_private(page);
2083                 id = lookup_swap_cgroup(ent);
2084                 rcu_read_lock();
2085                 mem = mem_cgroup_lookup(id);
2086                 if (mem && !css_tryget(&mem->css))
2087                         mem = NULL;
2088                 rcu_read_unlock();
2089         }
2090         unlock_page_cgroup(pc);
2091         return mem;
2092 }
2093
2094 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2095                                        struct page *page,
2096                                        struct page_cgroup *pc,
2097                                        enum charge_type ctype,
2098                                        int page_size)
2099 {
2100         int nr_pages = page_size >> PAGE_SHIFT;
2101
2102         lock_page_cgroup(pc);
2103         if (unlikely(PageCgroupUsed(pc))) {
2104                 unlock_page_cgroup(pc);
2105                 __mem_cgroup_cancel_charge(mem, nr_pages);
2106                 return;
2107         }
2108         /*
2109          * we don't need page_cgroup_lock about tail pages, becase they are not
2110          * accessed by any other context at this point.
2111          */
2112         pc->mem_cgroup = mem;
2113         /*
2114          * We access a page_cgroup asynchronously without lock_page_cgroup().
2115          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2116          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2117          * before USED bit, we need memory barrier here.
2118          * See mem_cgroup_add_lru_list(), etc.
2119          */
2120         smp_wmb();
2121         switch (ctype) {
2122         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2123         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2124                 SetPageCgroupCache(pc);
2125                 SetPageCgroupUsed(pc);
2126                 break;
2127         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2128                 ClearPageCgroupCache(pc);
2129                 SetPageCgroupUsed(pc);
2130                 break;
2131         default:
2132                 break;
2133         }
2134
2135         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2136         unlock_page_cgroup(pc);
2137         /*
2138          * "charge_statistics" updated event counter. Then, check it.
2139          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2140          * if they exceeds softlimit.
2141          */
2142         memcg_check_events(mem, page);
2143 }
2144
2145 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2146
2147 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2148                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2149 /*
2150  * Because tail pages are not marked as "used", set it. We're under
2151  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2152  */
2153 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2154 {
2155         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2156         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2157         unsigned long flags;
2158
2159         if (mem_cgroup_disabled())
2160                 return;
2161         /*
2162          * We have no races with charge/uncharge but will have races with
2163          * page state accounting.
2164          */
2165         move_lock_page_cgroup(head_pc, &flags);
2166
2167         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2168         smp_wmb(); /* see __commit_charge() */
2169         if (PageCgroupAcctLRU(head_pc)) {
2170                 enum lru_list lru;
2171                 struct mem_cgroup_per_zone *mz;
2172
2173                 /*
2174                  * LRU flags cannot be copied because we need to add tail
2175                  *.page to LRU by generic call and our hook will be called.
2176                  * We hold lru_lock, then, reduce counter directly.
2177                  */
2178                 lru = page_lru(head);
2179                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2180                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2181         }
2182         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2183         move_unlock_page_cgroup(head_pc, &flags);
2184 }
2185 #endif
2186
2187 /**
2188  * mem_cgroup_move_account - move account of the page
2189  * @page: the page
2190  * @pc: page_cgroup of the page.
2191  * @from: mem_cgroup which the page is moved from.
2192  * @to: mem_cgroup which the page is moved to. @from != @to.
2193  * @uncharge: whether we should call uncharge and css_put against @from.
2194  * @charge_size: number of bytes to charge (regular or huge page)
2195  *
2196  * The caller must confirm following.
2197  * - page is not on LRU (isolate_page() is useful.)
2198  * - compound_lock is held when charge_size > PAGE_SIZE
2199  *
2200  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2201  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2202  * true, this function does "uncharge" from old cgroup, but it doesn't if
2203  * @uncharge is false, so a caller should do "uncharge".
2204  */
2205 static int mem_cgroup_move_account(struct page *page, struct page_cgroup *pc,
2206                                    struct mem_cgroup *from, struct mem_cgroup *to,
2207                                    bool uncharge, int charge_size)
2208 {
2209         int nr_pages = charge_size >> PAGE_SHIFT;
2210         unsigned long flags;
2211         int ret;
2212
2213         VM_BUG_ON(from == to);
2214         VM_BUG_ON(PageLRU(page));
2215         /*
2216          * The page is isolated from LRU. So, collapse function
2217          * will not handle this page. But page splitting can happen.
2218          * Do this check under compound_page_lock(). The caller should
2219          * hold it.
2220          */
2221         ret = -EBUSY;
2222         if (charge_size > PAGE_SIZE && !PageTransHuge(page))
2223                 goto out;
2224
2225         lock_page_cgroup(pc);
2226
2227         ret = -EINVAL;
2228         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2229                 goto unlock;
2230
2231         move_lock_page_cgroup(pc, &flags);
2232
2233         if (PageCgroupFileMapped(pc)) {
2234                 /* Update mapped_file data for mem_cgroup */
2235                 preempt_disable();
2236                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2237                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2238                 preempt_enable();
2239         }
2240         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2241         if (uncharge)
2242                 /* This is not "cancel", but cancel_charge does all we need. */
2243                 __mem_cgroup_cancel_charge(from, nr_pages);
2244
2245         /* caller should have done css_get */
2246         pc->mem_cgroup = to;
2247         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2248         /*
2249          * We charges against "to" which may not have any tasks. Then, "to"
2250          * can be under rmdir(). But in current implementation, caller of
2251          * this function is just force_empty() and move charge, so it's
2252          * garanteed that "to" is never removed. So, we don't check rmdir
2253          * status here.
2254          */
2255         move_unlock_page_cgroup(pc, &flags);
2256         ret = 0;
2257 unlock:
2258         unlock_page_cgroup(pc);
2259         /*
2260          * check events
2261          */
2262         memcg_check_events(to, page);
2263         memcg_check_events(from, page);
2264 out:
2265         return ret;
2266 }
2267
2268 /*
2269  * move charges to its parent.
2270  */
2271
2272 static int mem_cgroup_move_parent(struct page *page,
2273                                   struct page_cgroup *pc,
2274                                   struct mem_cgroup *child,
2275                                   gfp_t gfp_mask)
2276 {
2277         struct cgroup *cg = child->css.cgroup;
2278         struct cgroup *pcg = cg->parent;
2279         struct mem_cgroup *parent;
2280         int page_size = PAGE_SIZE;
2281         unsigned long flags;
2282         int ret;
2283
2284         /* Is ROOT ? */
2285         if (!pcg)
2286                 return -EINVAL;
2287
2288         ret = -EBUSY;
2289         if (!get_page_unless_zero(page))
2290                 goto out;
2291         if (isolate_lru_page(page))
2292                 goto put;
2293
2294         if (PageTransHuge(page))
2295                 page_size = HPAGE_SIZE;
2296
2297         parent = mem_cgroup_from_cont(pcg);
2298         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2299                                 &parent, false, page_size);
2300         if (ret || !parent)
2301                 goto put_back;
2302
2303         if (page_size > PAGE_SIZE)
2304                 flags = compound_lock_irqsave(page);
2305
2306         ret = mem_cgroup_move_account(page, pc, child, parent, true, page_size);
2307         if (ret)
2308                 __mem_cgroup_cancel_charge(parent, page_size >> PAGE_SHIFT);
2309
2310         if (page_size > PAGE_SIZE)
2311                 compound_unlock_irqrestore(page, flags);
2312 put_back:
2313         putback_lru_page(page);
2314 put:
2315         put_page(page);
2316 out:
2317         return ret;
2318 }
2319
2320 /*
2321  * Charge the memory controller for page usage.
2322  * Return
2323  * 0 if the charge was successful
2324  * < 0 if the cgroup is over its limit
2325  */
2326 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2327                                 gfp_t gfp_mask, enum charge_type ctype)
2328 {
2329         struct mem_cgroup *mem = NULL;
2330         int page_size = PAGE_SIZE;
2331         struct page_cgroup *pc;
2332         bool oom = true;
2333         int ret;
2334
2335         if (PageTransHuge(page)) {
2336                 page_size <<= compound_order(page);
2337                 VM_BUG_ON(!PageTransHuge(page));
2338                 /*
2339                  * Never OOM-kill a process for a huge page.  The
2340                  * fault handler will fall back to regular pages.
2341                  */
2342                 oom = false;
2343         }
2344
2345         pc = lookup_page_cgroup(page);
2346         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2347
2348         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2349         if (ret || !mem)
2350                 return ret;
2351
2352         __mem_cgroup_commit_charge(mem, page, pc, ctype, page_size);
2353         return 0;
2354 }
2355
2356 int mem_cgroup_newpage_charge(struct page *page,
2357                               struct mm_struct *mm, gfp_t gfp_mask)
2358 {
2359         if (mem_cgroup_disabled())
2360                 return 0;
2361         /*
2362          * If already mapped, we don't have to account.
2363          * If page cache, page->mapping has address_space.
2364          * But page->mapping may have out-of-use anon_vma pointer,
2365          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2366          * is NULL.
2367          */
2368         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2369                 return 0;
2370         if (unlikely(!mm))
2371                 mm = &init_mm;
2372         return mem_cgroup_charge_common(page, mm, gfp_mask,
2373                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2374 }
2375
2376 static void
2377 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2378                                         enum charge_type ctype);
2379
2380 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2381                                 gfp_t gfp_mask)
2382 {
2383         int ret;
2384
2385         if (mem_cgroup_disabled())
2386                 return 0;
2387         if (PageCompound(page))
2388                 return 0;
2389         /*
2390          * Corner case handling. This is called from add_to_page_cache()
2391          * in usual. But some FS (shmem) precharges this page before calling it
2392          * and call add_to_page_cache() with GFP_NOWAIT.
2393          *
2394          * For GFP_NOWAIT case, the page may be pre-charged before calling
2395          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2396          * charge twice. (It works but has to pay a bit larger cost.)
2397          * And when the page is SwapCache, it should take swap information
2398          * into account. This is under lock_page() now.
2399          */
2400         if (!(gfp_mask & __GFP_WAIT)) {
2401                 struct page_cgroup *pc;
2402
2403                 pc = lookup_page_cgroup(page);
2404                 if (!pc)
2405                         return 0;
2406                 lock_page_cgroup(pc);
2407                 if (PageCgroupUsed(pc)) {
2408                         unlock_page_cgroup(pc);
2409                         return 0;
2410                 }
2411                 unlock_page_cgroup(pc);
2412         }
2413
2414         if (unlikely(!mm))
2415                 mm = &init_mm;
2416
2417         if (page_is_file_cache(page))
2418                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2419                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2420
2421         /* shmem */
2422         if (PageSwapCache(page)) {
2423                 struct mem_cgroup *mem;
2424
2425                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2426                 if (!ret)
2427                         __mem_cgroup_commit_charge_swapin(page, mem,
2428                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2429         } else
2430                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2431                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2432
2433         return ret;
2434 }
2435
2436 /*
2437  * While swap-in, try_charge -> commit or cancel, the page is locked.
2438  * And when try_charge() successfully returns, one refcnt to memcg without
2439  * struct page_cgroup is acquired. This refcnt will be consumed by
2440  * "commit()" or removed by "cancel()"
2441  */
2442 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2443                                  struct page *page,
2444                                  gfp_t mask, struct mem_cgroup **ptr)
2445 {
2446         struct mem_cgroup *mem;
2447         int ret;
2448
2449         *ptr = NULL;
2450
2451         if (mem_cgroup_disabled())
2452                 return 0;
2453
2454         if (!do_swap_account)
2455                 goto charge_cur_mm;
2456         /*
2457          * A racing thread's fault, or swapoff, may have already updated
2458          * the pte, and even removed page from swap cache: in those cases
2459          * do_swap_page()'s pte_same() test will fail; but there's also a
2460          * KSM case which does need to charge the page.
2461          */
2462         if (!PageSwapCache(page))
2463                 goto charge_cur_mm;
2464         mem = try_get_mem_cgroup_from_page(page);
2465         if (!mem)
2466                 goto charge_cur_mm;
2467         *ptr = mem;
2468         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2469         css_put(&mem->css);
2470         return ret;
2471 charge_cur_mm:
2472         if (unlikely(!mm))
2473                 mm = &init_mm;
2474         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2475 }
2476
2477 static void
2478 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2479                                         enum charge_type ctype)
2480 {
2481         struct page_cgroup *pc;
2482
2483         if (mem_cgroup_disabled())
2484                 return;
2485         if (!ptr)
2486                 return;
2487         cgroup_exclude_rmdir(&ptr->css);
2488         pc = lookup_page_cgroup(page);
2489         mem_cgroup_lru_del_before_commit_swapcache(page);
2490         __mem_cgroup_commit_charge(ptr, page, pc, ctype, PAGE_SIZE);
2491         mem_cgroup_lru_add_after_commit_swapcache(page);
2492         /*
2493          * Now swap is on-memory. This means this page may be
2494          * counted both as mem and swap....double count.
2495          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2496          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2497          * may call delete_from_swap_cache() before reach here.
2498          */
2499         if (do_swap_account && PageSwapCache(page)) {
2500                 swp_entry_t ent = {.val = page_private(page)};
2501                 unsigned short id;
2502                 struct mem_cgroup *memcg;
2503
2504                 id = swap_cgroup_record(ent, 0);
2505                 rcu_read_lock();
2506                 memcg = mem_cgroup_lookup(id);
2507                 if (memcg) {
2508                         /*
2509                          * This recorded memcg can be obsolete one. So, avoid
2510                          * calling css_tryget
2511                          */
2512                         if (!mem_cgroup_is_root(memcg))
2513                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2514                         mem_cgroup_swap_statistics(memcg, false);
2515                         mem_cgroup_put(memcg);
2516                 }
2517                 rcu_read_unlock();
2518         }
2519         /*
2520          * At swapin, we may charge account against cgroup which has no tasks.
2521          * So, rmdir()->pre_destroy() can be called while we do this charge.
2522          * In that case, we need to call pre_destroy() again. check it here.
2523          */
2524         cgroup_release_and_wakeup_rmdir(&ptr->css);
2525 }
2526
2527 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2528 {
2529         __mem_cgroup_commit_charge_swapin(page, ptr,
2530                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2531 }
2532
2533 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2534 {
2535         if (mem_cgroup_disabled())
2536                 return;
2537         if (!mem)
2538                 return;
2539         __mem_cgroup_cancel_charge(mem, 1);
2540 }
2541
2542 static void
2543 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2544               int page_size)
2545 {
2546         struct memcg_batch_info *batch = NULL;
2547         bool uncharge_memsw = true;
2548         /* If swapout, usage of swap doesn't decrease */
2549         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2550                 uncharge_memsw = false;
2551
2552         batch = &current->memcg_batch;
2553         /*
2554          * In usual, we do css_get() when we remember memcg pointer.
2555          * But in this case, we keep res->usage until end of a series of
2556          * uncharges. Then, it's ok to ignore memcg's refcnt.
2557          */
2558         if (!batch->memcg)
2559                 batch->memcg = mem;
2560         /*
2561          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2562          * In those cases, all pages freed continously can be expected to be in
2563          * the same cgroup and we have chance to coalesce uncharges.
2564          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2565          * because we want to do uncharge as soon as possible.
2566          */
2567
2568         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2569                 goto direct_uncharge;
2570
2571         if (page_size != PAGE_SIZE)
2572                 goto direct_uncharge;
2573
2574         /*
2575          * In typical case, batch->memcg == mem. This means we can
2576          * merge a series of uncharges to an uncharge of res_counter.
2577          * If not, we uncharge res_counter ony by one.
2578          */
2579         if (batch->memcg != mem)
2580                 goto direct_uncharge;
2581         /* remember freed charge and uncharge it later */
2582         batch->bytes += PAGE_SIZE;
2583         if (uncharge_memsw)
2584                 batch->memsw_bytes += PAGE_SIZE;
2585         return;
2586 direct_uncharge:
2587         res_counter_uncharge(&mem->res, page_size);
2588         if (uncharge_memsw)
2589                 res_counter_uncharge(&mem->memsw, page_size);
2590         if (unlikely(batch->memcg != mem))
2591                 memcg_oom_recover(mem);
2592         return;
2593 }
2594
2595 /*
2596  * uncharge if !page_mapped(page)
2597  */
2598 static struct mem_cgroup *
2599 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2600 {
2601         int count;
2602         struct page_cgroup *pc;
2603         struct mem_cgroup *mem = NULL;
2604         int page_size = PAGE_SIZE;
2605
2606         if (mem_cgroup_disabled())
2607                 return NULL;
2608
2609         if (PageSwapCache(page))
2610                 return NULL;
2611
2612         if (PageTransHuge(page)) {
2613                 page_size <<= compound_order(page);
2614                 VM_BUG_ON(!PageTransHuge(page));
2615         }
2616
2617         count = page_size >> PAGE_SHIFT;
2618         /*
2619          * Check if our page_cgroup is valid
2620          */
2621         pc = lookup_page_cgroup(page);
2622         if (unlikely(!pc || !PageCgroupUsed(pc)))
2623                 return NULL;
2624
2625         lock_page_cgroup(pc);
2626
2627         mem = pc->mem_cgroup;
2628
2629         if (!PageCgroupUsed(pc))
2630                 goto unlock_out;
2631
2632         switch (ctype) {
2633         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2634         case MEM_CGROUP_CHARGE_TYPE_DROP:
2635                 /* See mem_cgroup_prepare_migration() */
2636                 if (page_mapped(page) || PageCgroupMigration(pc))
2637                         goto unlock_out;
2638                 break;
2639         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2640                 if (!PageAnon(page)) {  /* Shared memory */
2641                         if (page->mapping && !page_is_file_cache(page))
2642                                 goto unlock_out;
2643                 } else if (page_mapped(page)) /* Anon */
2644                                 goto unlock_out;
2645                 break;
2646         default:
2647                 break;
2648         }
2649
2650         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2651
2652         ClearPageCgroupUsed(pc);
2653         /*
2654          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2655          * freed from LRU. This is safe because uncharged page is expected not
2656          * to be reused (freed soon). Exception is SwapCache, it's handled by
2657          * special functions.
2658          */
2659
2660         unlock_page_cgroup(pc);
2661         /*
2662          * even after unlock, we have mem->res.usage here and this memcg
2663          * will never be freed.
2664          */
2665         memcg_check_events(mem, page);
2666         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2667                 mem_cgroup_swap_statistics(mem, true);
2668                 mem_cgroup_get(mem);
2669         }
2670         if (!mem_cgroup_is_root(mem))
2671                 __do_uncharge(mem, ctype, page_size);
2672
2673         return mem;
2674
2675 unlock_out:
2676         unlock_page_cgroup(pc);
2677         return NULL;
2678 }
2679
2680 void mem_cgroup_uncharge_page(struct page *page)
2681 {
2682         /* early check. */
2683         if (page_mapped(page))
2684                 return;
2685         if (page->mapping && !PageAnon(page))
2686                 return;
2687         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2688 }
2689
2690 void mem_cgroup_uncharge_cache_page(struct page *page)
2691 {
2692         VM_BUG_ON(page_mapped(page));
2693         VM_BUG_ON(page->mapping);
2694         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2695 }
2696
2697 /*
2698  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2699  * In that cases, pages are freed continuously and we can expect pages
2700  * are in the same memcg. All these calls itself limits the number of
2701  * pages freed at once, then uncharge_start/end() is called properly.
2702  * This may be called prural(2) times in a context,
2703  */
2704
2705 void mem_cgroup_uncharge_start(void)
2706 {
2707         current->memcg_batch.do_batch++;
2708         /* We can do nest. */
2709         if (current->memcg_batch.do_batch == 1) {
2710                 current->memcg_batch.memcg = NULL;
2711                 current->memcg_batch.bytes = 0;
2712                 current->memcg_batch.memsw_bytes = 0;
2713         }
2714 }
2715
2716 void mem_cgroup_uncharge_end(void)
2717 {
2718         struct memcg_batch_info *batch = &current->memcg_batch;
2719
2720         if (!batch->do_batch)
2721                 return;
2722
2723         batch->do_batch--;
2724         if (batch->do_batch) /* If stacked, do nothing. */
2725                 return;
2726
2727         if (!batch->memcg)
2728                 return;
2729         /*
2730          * This "batch->memcg" is valid without any css_get/put etc...
2731          * bacause we hide charges behind us.
2732          */
2733         if (batch->bytes)
2734                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2735         if (batch->memsw_bytes)
2736                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2737         memcg_oom_recover(batch->memcg);
2738         /* forget this pointer (for sanity check) */
2739         batch->memcg = NULL;
2740 }
2741
2742 #ifdef CONFIG_SWAP
2743 /*
2744  * called after __delete_from_swap_cache() and drop "page" account.
2745  * memcg information is recorded to swap_cgroup of "ent"
2746  */
2747 void
2748 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2749 {
2750         struct mem_cgroup *memcg;
2751         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2752
2753         if (!swapout) /* this was a swap cache but the swap is unused ! */
2754                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2755
2756         memcg = __mem_cgroup_uncharge_common(page, ctype);
2757
2758         /*
2759          * record memcg information,  if swapout && memcg != NULL,
2760          * mem_cgroup_get() was called in uncharge().
2761          */
2762         if (do_swap_account && swapout && memcg)
2763                 swap_cgroup_record(ent, css_id(&memcg->css));
2764 }
2765 #endif
2766
2767 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2768 /*
2769  * called from swap_entry_free(). remove record in swap_cgroup and
2770  * uncharge "memsw" account.
2771  */
2772 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2773 {
2774         struct mem_cgroup *memcg;
2775         unsigned short id;
2776
2777         if (!do_swap_account)
2778                 return;
2779
2780         id = swap_cgroup_record(ent, 0);
2781         rcu_read_lock();
2782         memcg = mem_cgroup_lookup(id);
2783         if (memcg) {
2784                 /*
2785                  * We uncharge this because swap is freed.
2786                  * This memcg can be obsolete one. We avoid calling css_tryget
2787                  */
2788                 if (!mem_cgroup_is_root(memcg))
2789                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2790                 mem_cgroup_swap_statistics(memcg, false);
2791                 mem_cgroup_put(memcg);
2792         }
2793         rcu_read_unlock();
2794 }
2795
2796 /**
2797  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2798  * @entry: swap entry to be moved
2799  * @from:  mem_cgroup which the entry is moved from
2800  * @to:  mem_cgroup which the entry is moved to
2801  * @need_fixup: whether we should fixup res_counters and refcounts.
2802  *
2803  * It succeeds only when the swap_cgroup's record for this entry is the same
2804  * as the mem_cgroup's id of @from.
2805  *
2806  * Returns 0 on success, -EINVAL on failure.
2807  *
2808  * The caller must have charged to @to, IOW, called res_counter_charge() about
2809  * both res and memsw, and called css_get().
2810  */
2811 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2812                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2813 {
2814         unsigned short old_id, new_id;
2815
2816         old_id = css_id(&from->css);
2817         new_id = css_id(&to->css);
2818
2819         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2820                 mem_cgroup_swap_statistics(from, false);
2821                 mem_cgroup_swap_statistics(to, true);
2822                 /*
2823                  * This function is only called from task migration context now.
2824                  * It postpones res_counter and refcount handling till the end
2825                  * of task migration(mem_cgroup_clear_mc()) for performance
2826                  * improvement. But we cannot postpone mem_cgroup_get(to)
2827                  * because if the process that has been moved to @to does
2828                  * swap-in, the refcount of @to might be decreased to 0.
2829                  */
2830                 mem_cgroup_get(to);
2831                 if (need_fixup) {
2832                         if (!mem_cgroup_is_root(from))
2833                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2834                         mem_cgroup_put(from);
2835                         /*
2836                          * we charged both to->res and to->memsw, so we should
2837                          * uncharge to->res.
2838                          */
2839                         if (!mem_cgroup_is_root(to))
2840                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2841                 }
2842                 return 0;
2843         }
2844         return -EINVAL;
2845 }
2846 #else
2847 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2848                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2849 {
2850         return -EINVAL;
2851 }
2852 #endif
2853
2854 /*
2855  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2856  * page belongs to.
2857  */
2858 int mem_cgroup_prepare_migration(struct page *page,
2859         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2860 {
2861         struct page_cgroup *pc;
2862         struct mem_cgroup *mem = NULL;
2863         enum charge_type ctype;
2864         int ret = 0;
2865
2866         *ptr = NULL;
2867
2868         VM_BUG_ON(PageTransHuge(page));
2869         if (mem_cgroup_disabled())
2870                 return 0;
2871
2872         pc = lookup_page_cgroup(page);
2873         lock_page_cgroup(pc);
2874         if (PageCgroupUsed(pc)) {
2875                 mem = pc->mem_cgroup;
2876                 css_get(&mem->css);
2877                 /*
2878                  * At migrating an anonymous page, its mapcount goes down
2879                  * to 0 and uncharge() will be called. But, even if it's fully
2880                  * unmapped, migration may fail and this page has to be
2881                  * charged again. We set MIGRATION flag here and delay uncharge
2882                  * until end_migration() is called
2883                  *
2884                  * Corner Case Thinking
2885                  * A)
2886                  * When the old page was mapped as Anon and it's unmap-and-freed
2887                  * while migration was ongoing.
2888                  * If unmap finds the old page, uncharge() of it will be delayed
2889                  * until end_migration(). If unmap finds a new page, it's
2890                  * uncharged when it make mapcount to be 1->0. If unmap code
2891                  * finds swap_migration_entry, the new page will not be mapped
2892                  * and end_migration() will find it(mapcount==0).
2893                  *
2894                  * B)
2895                  * When the old page was mapped but migraion fails, the kernel
2896                  * remaps it. A charge for it is kept by MIGRATION flag even
2897                  * if mapcount goes down to 0. We can do remap successfully
2898                  * without charging it again.
2899                  *
2900                  * C)
2901                  * The "old" page is under lock_page() until the end of
2902                  * migration, so, the old page itself will not be swapped-out.
2903                  * If the new page is swapped out before end_migraton, our
2904                  * hook to usual swap-out path will catch the event.
2905                  */
2906                 if (PageAnon(page))
2907                         SetPageCgroupMigration(pc);
2908         }
2909         unlock_page_cgroup(pc);
2910         /*
2911          * If the page is not charged at this point,
2912          * we return here.
2913          */
2914         if (!mem)
2915                 return 0;
2916
2917         *ptr = mem;
2918         ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2919         css_put(&mem->css);/* drop extra refcnt */
2920         if (ret || *ptr == NULL) {
2921                 if (PageAnon(page)) {
2922                         lock_page_cgroup(pc);
2923                         ClearPageCgroupMigration(pc);
2924                         unlock_page_cgroup(pc);
2925                         /*
2926                          * The old page may be fully unmapped while we kept it.
2927                          */
2928                         mem_cgroup_uncharge_page(page);
2929                 }
2930                 return -ENOMEM;
2931         }
2932         /*
2933          * We charge new page before it's used/mapped. So, even if unlock_page()
2934          * is called before end_migration, we can catch all events on this new
2935          * page. In the case new page is migrated but not remapped, new page's
2936          * mapcount will be finally 0 and we call uncharge in end_migration().
2937          */
2938         pc = lookup_page_cgroup(newpage);
2939         if (PageAnon(page))
2940                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2941         else if (page_is_file_cache(page))
2942                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2943         else
2944                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2945         __mem_cgroup_commit_charge(mem, page, pc, ctype, PAGE_SIZE);
2946         return ret;
2947 }
2948
2949 /* remove redundant charge if migration failed*/
2950 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2951         struct page *oldpage, struct page *newpage, bool migration_ok)
2952 {
2953         struct page *used, *unused;
2954         struct page_cgroup *pc;
2955
2956         if (!mem)
2957                 return;
2958         /* blocks rmdir() */
2959         cgroup_exclude_rmdir(&mem->css);
2960         if (!migration_ok) {
2961                 used = oldpage;
2962                 unused = newpage;
2963         } else {
2964                 used = newpage;
2965                 unused = oldpage;
2966         }
2967         /*
2968          * We disallowed uncharge of pages under migration because mapcount
2969          * of the page goes down to zero, temporarly.
2970          * Clear the flag and check the page should be charged.
2971          */
2972         pc = lookup_page_cgroup(oldpage);
2973         lock_page_cgroup(pc);
2974         ClearPageCgroupMigration(pc);
2975         unlock_page_cgroup(pc);
2976
2977         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2978
2979         /*
2980          * If a page is a file cache, radix-tree replacement is very atomic
2981          * and we can skip this check. When it was an Anon page, its mapcount
2982          * goes down to 0. But because we added MIGRATION flage, it's not
2983          * uncharged yet. There are several case but page->mapcount check
2984          * and USED bit check in mem_cgroup_uncharge_page() will do enough
2985          * check. (see prepare_charge() also)
2986          */
2987         if (PageAnon(used))
2988                 mem_cgroup_uncharge_page(used);
2989         /*
2990          * At migration, we may charge account against cgroup which has no
2991          * tasks.
2992          * So, rmdir()->pre_destroy() can be called while we do this charge.
2993          * In that case, we need to call pre_destroy() again. check it here.
2994          */
2995         cgroup_release_and_wakeup_rmdir(&mem->css);
2996 }
2997
2998 /*
2999  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3000  * Calling hierarchical_reclaim is not enough because we should update
3001  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3002  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3003  * not from the memcg which this page would be charged to.
3004  * try_charge_swapin does all of these works properly.
3005  */
3006 int mem_cgroup_shmem_charge_fallback(struct page *page,
3007                             struct mm_struct *mm,
3008                             gfp_t gfp_mask)
3009 {
3010         struct mem_cgroup *mem;
3011         int ret;
3012
3013         if (mem_cgroup_disabled())
3014                 return 0;
3015
3016         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3017         if (!ret)
3018                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3019
3020         return ret;
3021 }
3022
3023 #ifdef CONFIG_DEBUG_VM
3024 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3025 {
3026         struct page_cgroup *pc;
3027
3028         pc = lookup_page_cgroup(page);
3029         if (likely(pc) && PageCgroupUsed(pc))
3030                 return pc;
3031         return NULL;
3032 }
3033
3034 bool mem_cgroup_bad_page_check(struct page *page)
3035 {
3036         if (mem_cgroup_disabled())
3037                 return false;
3038
3039         return lookup_page_cgroup_used(page) != NULL;
3040 }
3041
3042 void mem_cgroup_print_bad_page(struct page *page)
3043 {
3044         struct page_cgroup *pc;
3045
3046         pc = lookup_page_cgroup_used(page);
3047         if (pc) {
3048                 int ret = -1;
3049                 char *path;
3050
3051                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3052                        pc, pc->flags, pc->mem_cgroup);
3053
3054                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3055                 if (path) {
3056                         rcu_read_lock();
3057                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3058                                                         path, PATH_MAX);
3059                         rcu_read_unlock();
3060                 }
3061
3062                 printk(KERN_CONT "(%s)\n",
3063                                 (ret < 0) ? "cannot get the path" : path);
3064                 kfree(path);
3065         }
3066 }
3067 #endif
3068
3069 static DEFINE_MUTEX(set_limit_mutex);
3070
3071 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3072                                 unsigned long long val)
3073 {
3074         int retry_count;
3075         u64 memswlimit, memlimit;
3076         int ret = 0;
3077         int children = mem_cgroup_count_children(memcg);
3078         u64 curusage, oldusage;
3079         int enlarge;
3080
3081         /*
3082          * For keeping hierarchical_reclaim simple, how long we should retry
3083          * is depends on callers. We set our retry-count to be function
3084          * of # of children which we should visit in this loop.
3085          */
3086         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3087
3088         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3089
3090         enlarge = 0;
3091         while (retry_count) {
3092                 if (signal_pending(current)) {
3093                         ret = -EINTR;
3094                         break;
3095                 }
3096                 /*
3097                  * Rather than hide all in some function, I do this in
3098                  * open coded manner. You see what this really does.
3099                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3100                  */
3101                 mutex_lock(&set_limit_mutex);
3102                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3103                 if (memswlimit < val) {
3104                         ret = -EINVAL;
3105                         mutex_unlock(&set_limit_mutex);
3106                         break;
3107                 }
3108
3109                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3110                 if (memlimit < val)
3111                         enlarge = 1;
3112
3113                 ret = res_counter_set_limit(&memcg->res, val);
3114                 if (!ret) {
3115                         if (memswlimit == val)
3116                                 memcg->memsw_is_minimum = true;
3117                         else
3118                                 memcg->memsw_is_minimum = false;
3119                 }
3120                 mutex_unlock(&set_limit_mutex);
3121
3122                 if (!ret)
3123                         break;
3124
3125                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3126                                                 MEM_CGROUP_RECLAIM_SHRINK);
3127                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3128                 /* Usage is reduced ? */
3129                 if (curusage >= oldusage)
3130                         retry_count--;
3131                 else
3132                         oldusage = curusage;
3133         }
3134         if (!ret && enlarge)
3135                 memcg_oom_recover(memcg);
3136
3137         return ret;
3138 }
3139
3140 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3141                                         unsigned long long val)
3142 {
3143         int retry_count;
3144         u64 memlimit, memswlimit, oldusage, curusage;
3145         int children = mem_cgroup_count_children(memcg);
3146         int ret = -EBUSY;
3147         int enlarge = 0;
3148
3149         /* see mem_cgroup_resize_res_limit */
3150         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3151         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3152         while (retry_count) {
3153                 if (signal_pending(current)) {
3154                         ret = -EINTR;
3155                         break;
3156                 }
3157                 /*
3158                  * Rather than hide all in some function, I do this in
3159                  * open coded manner. You see what this really does.
3160                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3161                  */
3162                 mutex_lock(&set_limit_mutex);
3163                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3164                 if (memlimit > val) {
3165                         ret = -EINVAL;
3166                         mutex_unlock(&set_limit_mutex);
3167                         break;
3168                 }
3169                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3170                 if (memswlimit < val)
3171                         enlarge = 1;
3172                 ret = res_counter_set_limit(&memcg->memsw, val);
3173                 if (!ret) {
3174                         if (memlimit == val)
3175                                 memcg->memsw_is_minimum = true;
3176                         else
3177                                 memcg->memsw_is_minimum = false;
3178                 }
3179                 mutex_unlock(&set_limit_mutex);
3180
3181                 if (!ret)
3182                         break;
3183
3184                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3185                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3186                                                 MEM_CGROUP_RECLAIM_SHRINK);
3187                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3188                 /* Usage is reduced ? */
3189                 if (curusage >= oldusage)
3190                         retry_count--;
3191                 else
3192                         oldusage = curusage;
3193         }
3194         if (!ret && enlarge)
3195                 memcg_oom_recover(memcg);
3196         return ret;
3197 }
3198
3199 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3200                                             gfp_t gfp_mask)
3201 {
3202         unsigned long nr_reclaimed = 0;
3203         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3204         unsigned long reclaimed;
3205         int loop = 0;
3206         struct mem_cgroup_tree_per_zone *mctz;
3207         unsigned long long excess;
3208
3209         if (order > 0)
3210                 return 0;
3211
3212         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3213         /*
3214          * This loop can run a while, specially if mem_cgroup's continuously
3215          * keep exceeding their soft limit and putting the system under
3216          * pressure
3217          */
3218         do {
3219                 if (next_mz)
3220                         mz = next_mz;
3221                 else
3222                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3223                 if (!mz)
3224                         break;
3225
3226                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3227                                                 gfp_mask,
3228                                                 MEM_CGROUP_RECLAIM_SOFT);
3229                 nr_reclaimed += reclaimed;
3230                 spin_lock(&mctz->lock);
3231
3232                 /*
3233                  * If we failed to reclaim anything from this memory cgroup
3234                  * it is time to move on to the next cgroup
3235                  */
3236                 next_mz = NULL;
3237                 if (!reclaimed) {
3238                         do {
3239                                 /*
3240                                  * Loop until we find yet another one.
3241                                  *
3242                                  * By the time we get the soft_limit lock
3243                                  * again, someone might have aded the
3244                                  * group back on the RB tree. Iterate to
3245                                  * make sure we get a different mem.
3246                                  * mem_cgroup_largest_soft_limit_node returns
3247                                  * NULL if no other cgroup is present on
3248                                  * the tree
3249                                  */
3250                                 next_mz =
3251                                 __mem_cgroup_largest_soft_limit_node(mctz);
3252                                 if (next_mz == mz) {
3253                                         css_put(&next_mz->mem->css);
3254                                         next_mz = NULL;
3255                                 } else /* next_mz == NULL or other memcg */
3256                                         break;
3257                         } while (1);
3258                 }
3259                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3260                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3261                 /*
3262                  * One school of thought says that we should not add
3263                  * back the node to the tree if reclaim returns 0.
3264                  * But our reclaim could return 0, simply because due
3265                  * to priority we are exposing a smaller subset of
3266                  * memory to reclaim from. Consider this as a longer
3267                  * term TODO.
3268                  */
3269                 /* If excess == 0, no tree ops */
3270                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3271                 spin_unlock(&mctz->lock);
3272                 css_put(&mz->mem->css);
3273                 loop++;
3274                 /*
3275                  * Could not reclaim anything and there are no more
3276                  * mem cgroups to try or we seem to be looping without
3277                  * reclaiming anything.
3278                  */
3279                 if (!nr_reclaimed &&
3280                         (next_mz == NULL ||
3281                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3282                         break;
3283         } while (!nr_reclaimed);
3284         if (next_mz)
3285                 css_put(&next_mz->mem->css);
3286         return nr_reclaimed;
3287 }
3288
3289 /*
3290  * This routine traverse page_cgroup in given list and drop them all.
3291  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3292  */
3293 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3294                                 int node, int zid, enum lru_list lru)
3295 {
3296         struct zone *zone;
3297         struct mem_cgroup_per_zone *mz;
3298         struct page_cgroup *pc, *busy;
3299         unsigned long flags, loop;
3300         struct list_head *list;
3301         int ret = 0;
3302
3303         zone = &NODE_DATA(node)->node_zones[zid];
3304         mz = mem_cgroup_zoneinfo(mem, node, zid);
3305         list = &mz->lists[lru];
3306
3307         loop = MEM_CGROUP_ZSTAT(mz, lru);
3308         /* give some margin against EBUSY etc...*/
3309         loop += 256;
3310         busy = NULL;
3311         while (loop--) {
3312                 struct page *page;
3313
3314                 ret = 0;
3315                 spin_lock_irqsave(&zone->lru_lock, flags);
3316                 if (list_empty(list)) {
3317                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3318                         break;
3319                 }
3320                 pc = list_entry(list->prev, struct page_cgroup, lru);
3321                 if (busy == pc) {
3322                         list_move(&pc->lru, list);
3323                         busy = NULL;
3324                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3325                         continue;
3326                 }
3327                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3328
3329                 page = lookup_cgroup_page(pc);
3330
3331                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3332                 if (ret == -ENOMEM)
3333                         break;
3334
3335                 if (ret == -EBUSY || ret == -EINVAL) {
3336                         /* found lock contention or "pc" is obsolete. */
3337                         busy = pc;
3338                         cond_resched();
3339                 } else
3340                         busy = NULL;
3341         }
3342
3343         if (!ret && !list_empty(list))
3344                 return -EBUSY;
3345         return ret;
3346 }
3347
3348 /*
3349  * make mem_cgroup's charge to be 0 if there is no task.
3350  * This enables deleting this mem_cgroup.
3351  */
3352 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3353 {
3354         int ret;
3355         int node, zid, shrink;
3356         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3357         struct cgroup *cgrp = mem->css.cgroup;
3358
3359         css_get(&mem->css);
3360
3361         shrink = 0;
3362         /* should free all ? */
3363         if (free_all)
3364                 goto try_to_free;
3365 move_account:
3366         do {
3367                 ret = -EBUSY;
3368                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3369                         goto out;
3370                 ret = -EINTR;
3371                 if (signal_pending(current))
3372                         goto out;
3373                 /* This is for making all *used* pages to be on LRU. */
3374                 lru_add_drain_all();
3375                 drain_all_stock_sync();
3376                 ret = 0;
3377                 mem_cgroup_start_move(mem);
3378                 for_each_node_state(node, N_HIGH_MEMORY) {
3379                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3380                                 enum lru_list l;
3381                                 for_each_lru(l) {
3382                                         ret = mem_cgroup_force_empty_list(mem,
3383                                                         node, zid, l);
3384                                         if (ret)
3385                                                 break;
3386                                 }
3387                         }
3388                         if (ret)
3389                                 break;
3390                 }
3391                 mem_cgroup_end_move(mem);
3392                 memcg_oom_recover(mem);
3393                 /* it seems parent cgroup doesn't have enough mem */
3394                 if (ret == -ENOMEM)
3395                         goto try_to_free;
3396                 cond_resched();
3397         /* "ret" should also be checked to ensure all lists are empty. */
3398         } while (mem->res.usage > 0 || ret);
3399 out:
3400         css_put(&mem->css);
3401         return ret;
3402
3403 try_to_free:
3404         /* returns EBUSY if there is a task or if we come here twice. */
3405         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3406                 ret = -EBUSY;
3407                 goto out;
3408         }
3409         /* we call try-to-free pages for make this cgroup empty */
3410         lru_add_drain_all();
3411         /* try to free all pages in this cgroup */
3412         shrink = 1;
3413         while (nr_retries && mem->res.usage > 0) {
3414                 int progress;
3415
3416                 if (signal_pending(current)) {
3417                         ret = -EINTR;
3418                         goto out;
3419                 }
3420                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3421                                                 false, get_swappiness(mem));
3422                 if (!progress) {
3423                         nr_retries--;
3424                         /* maybe some writeback is necessary */
3425                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3426                 }
3427
3428         }
3429         lru_add_drain();
3430         /* try move_account...there may be some *locked* pages. */
3431         goto move_account;
3432 }
3433
3434 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3435 {
3436         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3437 }
3438
3439
3440 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3441 {
3442         return mem_cgroup_from_cont(cont)->use_hierarchy;
3443 }
3444
3445 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3446                                         u64 val)
3447 {
3448         int retval = 0;
3449         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3450         struct cgroup *parent = cont->parent;
3451         struct mem_cgroup *parent_mem = NULL;
3452
3453         if (parent)
3454                 parent_mem = mem_cgroup_from_cont(parent);
3455
3456         cgroup_lock();
3457         /*
3458          * If parent's use_hierarchy is set, we can't make any modifications
3459          * in the child subtrees. If it is unset, then the change can
3460          * occur, provided the current cgroup has no children.
3461          *
3462          * For the root cgroup, parent_mem is NULL, we allow value to be
3463          * set if there are no children.
3464          */
3465         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3466                                 (val == 1 || val == 0)) {
3467                 if (list_empty(&cont->children))
3468                         mem->use_hierarchy = val;
3469                 else
3470                         retval = -EBUSY;
3471         } else
3472                 retval = -EINVAL;
3473         cgroup_unlock();
3474
3475         return retval;
3476 }
3477
3478
3479 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3480                                 enum mem_cgroup_stat_index idx)
3481 {
3482         struct mem_cgroup *iter;
3483         s64 val = 0;
3484
3485         /* each per cpu's value can be minus.Then, use s64 */
3486         for_each_mem_cgroup_tree(iter, mem)
3487                 val += mem_cgroup_read_stat(iter, idx);
3488
3489         if (val < 0) /* race ? */
3490                 val = 0;
3491         return val;
3492 }
3493
3494 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3495 {
3496         u64 val;
3497
3498         if (!mem_cgroup_is_root(mem)) {
3499                 if (!swap)
3500                         return res_counter_read_u64(&mem->res, RES_USAGE);
3501                 else
3502                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3503         }
3504
3505         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3506         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3507
3508         if (swap)
3509                 val += mem_cgroup_get_recursive_idx_stat(mem,
3510                                 MEM_CGROUP_STAT_SWAPOUT);
3511
3512         return val << PAGE_SHIFT;
3513 }
3514
3515 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3516 {
3517         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3518         u64 val;
3519         int type, name;
3520
3521         type = MEMFILE_TYPE(cft->private);
3522         name = MEMFILE_ATTR(cft->private);
3523         switch (type) {
3524         case _MEM:
3525                 if (name == RES_USAGE)
3526                         val = mem_cgroup_usage(mem, false);
3527                 else
3528                         val = res_counter_read_u64(&mem->res, name);
3529                 break;
3530         case _MEMSWAP:
3531                 if (name == RES_USAGE)
3532                         val = mem_cgroup_usage(mem, true);
3533                 else
3534                         val = res_counter_read_u64(&mem->memsw, name);
3535                 break;
3536         default:
3537                 BUG();
3538                 break;
3539         }
3540         return val;
3541 }
3542 /*
3543  * The user of this function is...
3544  * RES_LIMIT.
3545  */
3546 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3547                             const char *buffer)
3548 {
3549         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3550         int type, name;
3551         unsigned long long val;
3552         int ret;
3553
3554         type = MEMFILE_TYPE(cft->private);
3555         name = MEMFILE_ATTR(cft->private);
3556         switch (name) {
3557         case RES_LIMIT:
3558                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3559                         ret = -EINVAL;
3560                         break;
3561                 }
3562                 /* This function does all necessary parse...reuse it */
3563                 ret = res_counter_memparse_write_strategy(buffer, &val);
3564                 if (ret)
3565                         break;
3566                 if (type == _MEM)
3567                         ret = mem_cgroup_resize_limit(memcg, val);
3568                 else
3569                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3570                 break;
3571         case RES_SOFT_LIMIT:
3572                 ret = res_counter_memparse_write_strategy(buffer, &val);
3573                 if (ret)
3574                         break;
3575                 /*
3576                  * For memsw, soft limits are hard to implement in terms
3577                  * of semantics, for now, we support soft limits for
3578                  * control without swap
3579                  */
3580                 if (type == _MEM)
3581                         ret = res_counter_set_soft_limit(&memcg->res, val);
3582                 else
3583                         ret = -EINVAL;
3584                 break;
3585         default:
3586                 ret = -EINVAL; /* should be BUG() ? */
3587                 break;
3588         }
3589         return ret;
3590 }
3591
3592 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3593                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3594 {
3595         struct cgroup *cgroup;
3596         unsigned long long min_limit, min_memsw_limit, tmp;
3597
3598         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3599         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3600         cgroup = memcg->css.cgroup;
3601         if (!memcg->use_hierarchy)
3602                 goto out;
3603
3604         while (cgroup->parent) {
3605                 cgroup = cgroup->parent;
3606                 memcg = mem_cgroup_from_cont(cgroup);
3607                 if (!memcg->use_hierarchy)
3608                         break;
3609                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3610                 min_limit = min(min_limit, tmp);
3611                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3612                 min_memsw_limit = min(min_memsw_limit, tmp);
3613         }
3614 out:
3615         *mem_limit = min_limit;
3616         *memsw_limit = min_memsw_limit;
3617         return;
3618 }
3619
3620 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3621 {
3622         struct mem_cgroup *mem;
3623         int type, name;
3624
3625         mem = mem_cgroup_from_cont(cont);
3626         type = MEMFILE_TYPE(event);
3627         name = MEMFILE_ATTR(event);
3628         switch (name) {
3629         case RES_MAX_USAGE:
3630                 if (type == _MEM)
3631                         res_counter_reset_max(&mem->res);
3632                 else
3633                         res_counter_reset_max(&mem->memsw);
3634                 break;
3635         case RES_FAILCNT:
3636                 if (type == _MEM)
3637                         res_counter_reset_failcnt(&mem->res);
3638                 else
3639                         res_counter_reset_failcnt(&mem->memsw);
3640                 break;
3641         }
3642
3643         return 0;
3644 }
3645
3646 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3647                                         struct cftype *cft)
3648 {
3649         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3650 }
3651
3652 #ifdef CONFIG_MMU
3653 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3654                                         struct cftype *cft, u64 val)
3655 {
3656         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3657
3658         if (val >= (1 << NR_MOVE_TYPE))
3659                 return -EINVAL;
3660         /*
3661          * We check this value several times in both in can_attach() and
3662          * attach(), so we need cgroup lock to prevent this value from being
3663          * inconsistent.
3664          */
3665         cgroup_lock();
3666         mem->move_charge_at_immigrate = val;
3667         cgroup_unlock();
3668
3669         return 0;
3670 }
3671 #else
3672 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3673                                         struct cftype *cft, u64 val)
3674 {
3675         return -ENOSYS;
3676 }
3677 #endif
3678
3679
3680 /* For read statistics */
3681 enum {
3682         MCS_CACHE,
3683         MCS_RSS,
3684         MCS_FILE_MAPPED,
3685         MCS_PGPGIN,
3686         MCS_PGPGOUT,
3687         MCS_SWAP,
3688         MCS_INACTIVE_ANON,
3689         MCS_ACTIVE_ANON,
3690         MCS_INACTIVE_FILE,
3691         MCS_ACTIVE_FILE,
3692         MCS_UNEVICTABLE,
3693         NR_MCS_STAT,
3694 };
3695
3696 struct mcs_total_stat {
3697         s64 stat[NR_MCS_STAT];
3698 };
3699
3700 struct {
3701         char *local_name;
3702         char *total_name;
3703 } memcg_stat_strings[NR_MCS_STAT] = {
3704         {"cache", "total_cache"},
3705         {"rss", "total_rss"},
3706         {"mapped_file", "total_mapped_file"},
3707         {"pgpgin", "total_pgpgin"},
3708         {"pgpgout", "total_pgpgout"},
3709         {"swap", "total_swap"},
3710         {"inactive_anon", "total_inactive_anon"},
3711         {"active_anon", "total_active_anon"},
3712         {"inactive_file", "total_inactive_file"},
3713         {"active_file", "total_active_file"},
3714         {"unevictable", "total_unevictable"}
3715 };
3716
3717
3718 static void
3719 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3720 {
3721         s64 val;
3722
3723         /* per cpu stat */
3724         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3725         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3726         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3727         s->stat[MCS_RSS] += val * PAGE_SIZE;
3728         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3729         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3730         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3731         s->stat[MCS_PGPGIN] += val;
3732         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3733         s->stat[MCS_PGPGOUT] += val;
3734         if (do_swap_account) {
3735                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3736                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3737         }
3738
3739         /* per zone stat */
3740         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3741         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3742         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3743         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3744         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3745         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3746         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3747         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3748         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3749         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3750 }
3751
3752 static void
3753 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3754 {
3755         struct mem_cgroup *iter;
3756
3757         for_each_mem_cgroup_tree(iter, mem)
3758                 mem_cgroup_get_local_stat(iter, s);
3759 }
3760
3761 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3762                                  struct cgroup_map_cb *cb)
3763 {
3764         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3765         struct mcs_total_stat mystat;
3766         int i;
3767
3768         memset(&mystat, 0, sizeof(mystat));
3769         mem_cgroup_get_local_stat(mem_cont, &mystat);
3770
3771         for (i = 0; i < NR_MCS_STAT; i++) {
3772                 if (i == MCS_SWAP && !do_swap_account)
3773                         continue;
3774                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3775         }
3776
3777         /* Hierarchical information */
3778         {
3779                 unsigned long long limit, memsw_limit;
3780                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3781                 cb->fill(cb, "hierarchical_memory_limit", limit);
3782                 if (do_swap_account)
3783                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3784         }
3785
3786         memset(&mystat, 0, sizeof(mystat));
3787         mem_cgroup_get_total_stat(mem_cont, &mystat);
3788         for (i = 0; i < NR_MCS_STAT; i++) {
3789                 if (i == MCS_SWAP && !do_swap_account)
3790                         continue;
3791                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3792         }
3793
3794 #ifdef CONFIG_DEBUG_VM
3795         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3796
3797         {
3798                 int nid, zid;
3799                 struct mem_cgroup_per_zone *mz;
3800                 unsigned long recent_rotated[2] = {0, 0};
3801                 unsigned long recent_scanned[2] = {0, 0};
3802
3803                 for_each_online_node(nid)
3804                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3805                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3806
3807                                 recent_rotated[0] +=
3808                                         mz->reclaim_stat.recent_rotated[0];
3809                                 recent_rotated[1] +=
3810                                         mz->reclaim_stat.recent_rotated[1];
3811                                 recent_scanned[0] +=
3812                                         mz->reclaim_stat.recent_scanned[0];
3813                                 recent_scanned[1] +=
3814                                         mz->reclaim_stat.recent_scanned[1];
3815                         }
3816                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3817                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3818                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3819                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3820         }
3821 #endif
3822
3823         return 0;
3824 }
3825
3826 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3827 {
3828         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3829
3830         return get_swappiness(memcg);
3831 }
3832
3833 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3834                                        u64 val)
3835 {
3836         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3837         struct mem_cgroup *parent;
3838
3839         if (val > 100)
3840                 return -EINVAL;
3841
3842         if (cgrp->parent == NULL)
3843                 return -EINVAL;
3844
3845         parent = mem_cgroup_from_cont(cgrp->parent);
3846
3847         cgroup_lock();
3848
3849         /* If under hierarchy, only empty-root can set this value */
3850         if ((parent->use_hierarchy) ||
3851             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3852                 cgroup_unlock();
3853                 return -EINVAL;
3854         }
3855
3856         memcg->swappiness = val;
3857
3858         cgroup_unlock();
3859
3860         return 0;
3861 }
3862
3863 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3864 {
3865         struct mem_cgroup_threshold_ary *t;
3866         u64 usage;
3867         int i;
3868
3869         rcu_read_lock();
3870         if (!swap)
3871                 t = rcu_dereference(memcg->thresholds.primary);
3872         else
3873                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3874
3875         if (!t)
3876                 goto unlock;
3877
3878         usage = mem_cgroup_usage(memcg, swap);
3879
3880         /*
3881          * current_threshold points to threshold just below usage.
3882          * If it's not true, a threshold was crossed after last
3883          * call of __mem_cgroup_threshold().
3884          */
3885         i = t->current_threshold;
3886
3887         /*
3888          * Iterate backward over array of thresholds starting from
3889          * current_threshold and check if a threshold is crossed.
3890          * If none of thresholds below usage is crossed, we read
3891          * only one element of the array here.
3892          */
3893         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3894                 eventfd_signal(t->entries[i].eventfd, 1);
3895
3896         /* i = current_threshold + 1 */
3897         i++;
3898
3899         /*
3900          * Iterate forward over array of thresholds starting from
3901          * current_threshold+1 and check if a threshold is crossed.
3902          * If none of thresholds above usage is crossed, we read
3903          * only one element of the array here.
3904          */
3905         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3906                 eventfd_signal(t->entries[i].eventfd, 1);
3907
3908         /* Update current_threshold */
3909         t->current_threshold = i - 1;
3910 unlock:
3911         rcu_read_unlock();
3912 }
3913
3914 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3915 {
3916         while (memcg) {
3917                 __mem_cgroup_threshold(memcg, false);
3918                 if (do_swap_account)
3919                         __mem_cgroup_threshold(memcg, true);
3920
3921                 memcg = parent_mem_cgroup(memcg);
3922         }
3923 }
3924
3925 static int compare_thresholds(const void *a, const void *b)
3926 {
3927         const struct mem_cgroup_threshold *_a = a;
3928         const struct mem_cgroup_threshold *_b = b;
3929
3930         return _a->threshold - _b->threshold;
3931 }
3932
3933 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3934 {
3935         struct mem_cgroup_eventfd_list *ev;
3936
3937         list_for_each_entry(ev, &mem->oom_notify, list)
3938                 eventfd_signal(ev->eventfd, 1);
3939         return 0;
3940 }
3941
3942 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3943 {
3944         struct mem_cgroup *iter;
3945
3946         for_each_mem_cgroup_tree(iter, mem)
3947                 mem_cgroup_oom_notify_cb(iter);
3948 }
3949
3950 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3951         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3952 {
3953         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3954         struct mem_cgroup_thresholds *thresholds;
3955         struct mem_cgroup_threshold_ary *new;
3956         int type = MEMFILE_TYPE(cft->private);
3957         u64 threshold, usage;
3958         int i, size, ret;
3959
3960         ret = res_counter_memparse_write_strategy(args, &threshold);
3961         if (ret)
3962                 return ret;
3963
3964         mutex_lock(&memcg->thresholds_lock);
3965
3966         if (type == _MEM)
3967                 thresholds = &memcg->thresholds;
3968         else if (type == _MEMSWAP)
3969                 thresholds = &memcg->memsw_thresholds;
3970         else
3971                 BUG();
3972
3973         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3974
3975         /* Check if a threshold crossed before adding a new one */
3976         if (thresholds->primary)
3977                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3978
3979         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3980
3981         /* Allocate memory for new array of thresholds */
3982         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3983                         GFP_KERNEL);
3984         if (!new) {
3985                 ret = -ENOMEM;
3986                 goto unlock;
3987         }
3988         new->size = size;
3989
3990         /* Copy thresholds (if any) to new array */
3991         if (thresholds->primary) {
3992                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3993                                 sizeof(struct mem_cgroup_threshold));
3994         }
3995
3996         /* Add new threshold */
3997         new->entries[size - 1].eventfd = eventfd;
3998         new->entries[size - 1].threshold = threshold;
3999
4000         /* Sort thresholds. Registering of new threshold isn't time-critical */
4001         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4002                         compare_thresholds, NULL);
4003
4004         /* Find current threshold */
4005         new->current_threshold = -1;
4006         for (i = 0; i < size; i++) {
4007                 if (new->entries[i].threshold < usage) {
4008                         /*
4009                          * new->current_threshold will not be used until
4010                          * rcu_assign_pointer(), so it's safe to increment
4011                          * it here.
4012                          */
4013                         ++new->current_threshold;
4014                 }
4015         }
4016
4017         /* Free old spare buffer and save old primary buffer as spare */
4018         kfree(thresholds->spare);
4019         thresholds->spare = thresholds->primary;
4020
4021         rcu_assign_pointer(thresholds->primary, new);
4022
4023         /* To be sure that nobody uses thresholds */
4024         synchronize_rcu();
4025
4026 unlock:
4027         mutex_unlock(&memcg->thresholds_lock);
4028
4029         return ret;
4030 }
4031
4032 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4033         struct cftype *cft, struct eventfd_ctx *eventfd)
4034 {
4035         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4036         struct mem_cgroup_thresholds *thresholds;
4037         struct mem_cgroup_threshold_ary *new;
4038         int type = MEMFILE_TYPE(cft->private);
4039         u64 usage;
4040         int i, j, size;
4041
4042         mutex_lock(&memcg->thresholds_lock);
4043         if (type == _MEM)
4044                 thresholds = &memcg->thresholds;
4045         else if (type == _MEMSWAP)
4046                 thresholds = &memcg->memsw_thresholds;
4047         else
4048                 BUG();
4049
4050         /*
4051          * Something went wrong if we trying to unregister a threshold
4052          * if we don't have thresholds
4053          */
4054         BUG_ON(!thresholds);
4055
4056         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4057
4058         /* Check if a threshold crossed before removing */
4059         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4060
4061         /* Calculate new number of threshold */
4062         size = 0;
4063         for (i = 0; i < thresholds->primary->size; i++) {
4064                 if (thresholds->primary->entries[i].eventfd != eventfd)
4065                         size++;
4066         }
4067
4068         new = thresholds->spare;
4069
4070         /* Set thresholds array to NULL if we don't have thresholds */
4071         if (!size) {
4072                 kfree(new);
4073                 new = NULL;
4074                 goto swap_buffers;
4075         }
4076
4077         new->size = size;
4078
4079         /* Copy thresholds and find current threshold */
4080         new->current_threshold = -1;
4081         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4082                 if (thresholds->primary->entries[i].eventfd == eventfd)
4083                         continue;
4084
4085                 new->entries[j] = thresholds->primary->entries[i];
4086                 if (new->entries[j].threshold < usage) {
4087                         /*
4088                          * new->current_threshold will not be used
4089                          * until rcu_assign_pointer(), so it's safe to increment
4090                          * it here.
4091                          */
4092                         ++new->current_threshold;
4093                 }
4094                 j++;
4095         }
4096
4097 swap_buffers:
4098         /* Swap primary and spare array */
4099         thresholds->spare = thresholds->primary;
4100         rcu_assign_pointer(thresholds->primary, new);
4101
4102         /* To be sure that nobody uses thresholds */
4103         synchronize_rcu();
4104
4105         mutex_unlock(&memcg->thresholds_lock);
4106 }
4107
4108 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4109         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4110 {
4111         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4112         struct mem_cgroup_eventfd_list *event;
4113         int type = MEMFILE_TYPE(cft->private);
4114
4115         BUG_ON(type != _OOM_TYPE);
4116         event = kmalloc(sizeof(*event), GFP_KERNEL);
4117         if (!event)
4118                 return -ENOMEM;
4119
4120         mutex_lock(&memcg_oom_mutex);
4121
4122         event->eventfd = eventfd;
4123         list_add(&event->list, &memcg->oom_notify);
4124
4125         /* already in OOM ? */
4126         if (atomic_read(&memcg->oom_lock))
4127                 eventfd_signal(eventfd, 1);
4128         mutex_unlock(&memcg_oom_mutex);
4129
4130         return 0;
4131 }
4132
4133 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4134         struct cftype *cft, struct eventfd_ctx *eventfd)
4135 {
4136         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4137         struct mem_cgroup_eventfd_list *ev, *tmp;
4138         int type = MEMFILE_TYPE(cft->private);
4139
4140         BUG_ON(type != _OOM_TYPE);
4141
4142         mutex_lock(&memcg_oom_mutex);
4143
4144         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4145                 if (ev->eventfd == eventfd) {
4146                         list_del(&ev->list);
4147                         kfree(ev);
4148                 }
4149         }
4150
4151         mutex_unlock(&memcg_oom_mutex);
4152 }
4153
4154 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4155         struct cftype *cft,  struct cgroup_map_cb *cb)
4156 {
4157         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4158
4159         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4160
4161         if (atomic_read(&mem->oom_lock))
4162                 cb->fill(cb, "under_oom", 1);
4163         else
4164                 cb->fill(cb, "under_oom", 0);
4165         return 0;
4166 }
4167
4168 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4169         struct cftype *cft, u64 val)
4170 {
4171         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4172         struct mem_cgroup *parent;
4173
4174         /* cannot set to root cgroup and only 0 and 1 are allowed */
4175         if (!cgrp->parent || !((val == 0) || (val == 1)))
4176                 return -EINVAL;
4177
4178         parent = mem_cgroup_from_cont(cgrp->parent);
4179
4180         cgroup_lock();
4181         /* oom-kill-disable is a flag for subhierarchy. */
4182         if ((parent->use_hierarchy) ||
4183             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4184                 cgroup_unlock();
4185                 return -EINVAL;
4186         }
4187         mem->oom_kill_disable = val;
4188         if (!val)
4189                 memcg_oom_recover(mem);
4190         cgroup_unlock();
4191         return 0;
4192 }
4193
4194 static struct cftype mem_cgroup_files[] = {
4195         {
4196                 .name = "usage_in_bytes",
4197                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4198                 .read_u64 = mem_cgroup_read,
4199                 .register_event = mem_cgroup_usage_register_event,
4200                 .unregister_event = mem_cgroup_usage_unregister_event,
4201         },
4202         {
4203                 .name = "max_usage_in_bytes",
4204                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4205                 .trigger = mem_cgroup_reset,
4206                 .read_u64 = mem_cgroup_read,
4207         },
4208         {
4209                 .name = "limit_in_bytes",
4210                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4211                 .write_string = mem_cgroup_write,
4212                 .read_u64 = mem_cgroup_read,
4213         },
4214         {
4215                 .name = "soft_limit_in_bytes",
4216                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4217                 .write_string = mem_cgroup_write,
4218                 .read_u64 = mem_cgroup_read,
4219         },
4220         {
4221                 .name = "failcnt",
4222                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4223                 .trigger = mem_cgroup_reset,
4224                 .read_u64 = mem_cgroup_read,
4225         },
4226         {
4227                 .name = "stat",
4228                 .read_map = mem_control_stat_show,
4229         },
4230         {
4231                 .name = "force_empty",
4232                 .trigger = mem_cgroup_force_empty_write,
4233         },
4234         {
4235                 .name = "use_hierarchy",
4236                 .write_u64 = mem_cgroup_hierarchy_write,
4237                 .read_u64 = mem_cgroup_hierarchy_read,
4238         },
4239         {
4240                 .name = "swappiness",
4241                 .read_u64 = mem_cgroup_swappiness_read,
4242                 .write_u64 = mem_cgroup_swappiness_write,
4243         },
4244         {
4245                 .name = "move_charge_at_immigrate",
4246                 .read_u64 = mem_cgroup_move_charge_read,
4247                 .write_u64 = mem_cgroup_move_charge_write,
4248         },
4249         {
4250                 .name = "oom_control",
4251                 .read_map = mem_cgroup_oom_control_read,
4252                 .write_u64 = mem_cgroup_oom_control_write,
4253                 .register_event = mem_cgroup_oom_register_event,
4254                 .unregister_event = mem_cgroup_oom_unregister_event,
4255                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4256         },
4257 };
4258
4259 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4260 static struct cftype memsw_cgroup_files[] = {
4261         {
4262                 .name = "memsw.usage_in_bytes",
4263                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4264                 .read_u64 = mem_cgroup_read,
4265                 .register_event = mem_cgroup_usage_register_event,
4266                 .unregister_event = mem_cgroup_usage_unregister_event,
4267         },
4268         {
4269                 .name = "memsw.max_usage_in_bytes",
4270                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4271                 .trigger = mem_cgroup_reset,
4272                 .read_u64 = mem_cgroup_read,
4273         },
4274         {
4275                 .name = "memsw.limit_in_bytes",
4276                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4277                 .write_string = mem_cgroup_write,
4278                 .read_u64 = mem_cgroup_read,
4279         },
4280         {
4281                 .name = "memsw.failcnt",
4282                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4283                 .trigger = mem_cgroup_reset,
4284                 .read_u64 = mem_cgroup_read,
4285         },
4286 };
4287
4288 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4289 {
4290         if (!do_swap_account)
4291                 return 0;
4292         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4293                                 ARRAY_SIZE(memsw_cgroup_files));
4294 };
4295 #else
4296 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4297 {
4298         return 0;
4299 }
4300 #endif
4301
4302 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4303 {
4304         struct mem_cgroup_per_node *pn;
4305         struct mem_cgroup_per_zone *mz;
4306         enum lru_list l;
4307         int zone, tmp = node;
4308         /*
4309          * This routine is called against possible nodes.
4310          * But it's BUG to call kmalloc() against offline node.
4311          *
4312          * TODO: this routine can waste much memory for nodes which will
4313          *       never be onlined. It's better to use memory hotplug callback
4314          *       function.
4315          */
4316         if (!node_state(node, N_NORMAL_MEMORY))
4317                 tmp = -1;
4318         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4319         if (!pn)
4320                 return 1;
4321
4322         mem->info.nodeinfo[node] = pn;
4323         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4324                 mz = &pn->zoneinfo[zone];
4325                 for_each_lru(l)
4326                         INIT_LIST_HEAD(&mz->lists[l]);
4327                 mz->usage_in_excess = 0;
4328                 mz->on_tree = false;
4329                 mz->mem = mem;
4330         }
4331         return 0;
4332 }
4333
4334 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4335 {
4336         kfree(mem->info.nodeinfo[node]);
4337 }
4338
4339 static struct mem_cgroup *mem_cgroup_alloc(void)
4340 {
4341         struct mem_cgroup *mem;
4342         int size = sizeof(struct mem_cgroup);
4343
4344         /* Can be very big if MAX_NUMNODES is very big */
4345         if (size < PAGE_SIZE)
4346                 mem = kzalloc(size, GFP_KERNEL);
4347         else
4348                 mem = vzalloc(size);
4349
4350         if (!mem)
4351                 return NULL;
4352
4353         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4354         if (!mem->stat)
4355                 goto out_free;
4356         spin_lock_init(&mem->pcp_counter_lock);
4357         return mem;
4358
4359 out_free:
4360         if (size < PAGE_SIZE)
4361                 kfree(mem);
4362         else
4363                 vfree(mem);
4364         return NULL;
4365 }
4366
4367 /*
4368  * At destroying mem_cgroup, references from swap_cgroup can remain.
4369  * (scanning all at force_empty is too costly...)
4370  *
4371  * Instead of clearing all references at force_empty, we remember
4372  * the number of reference from swap_cgroup and free mem_cgroup when
4373  * it goes down to 0.
4374  *
4375  * Removal of cgroup itself succeeds regardless of refs from swap.
4376  */
4377
4378 static void __mem_cgroup_free(struct mem_cgroup *mem)
4379 {
4380         int node;
4381
4382         mem_cgroup_remove_from_trees(mem);
4383         free_css_id(&mem_cgroup_subsys, &mem->css);
4384
4385         for_each_node_state(node, N_POSSIBLE)
4386                 free_mem_cgroup_per_zone_info(mem, node);
4387
4388         free_percpu(mem->stat);
4389         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4390                 kfree(mem);
4391         else
4392                 vfree(mem);
4393 }
4394
4395 static void mem_cgroup_get(struct mem_cgroup *mem)
4396 {
4397         atomic_inc(&mem->refcnt);
4398 }
4399
4400 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4401 {
4402         if (atomic_sub_and_test(count, &mem->refcnt)) {
4403                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4404                 __mem_cgroup_free(mem);
4405                 if (parent)
4406                         mem_cgroup_put(parent);
4407         }
4408 }
4409
4410 static void mem_cgroup_put(struct mem_cgroup *mem)
4411 {
4412         __mem_cgroup_put(mem, 1);
4413 }
4414
4415 /*
4416  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4417  */
4418 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4419 {
4420         if (!mem->res.parent)
4421                 return NULL;
4422         return mem_cgroup_from_res_counter(mem->res.parent, res);
4423 }
4424
4425 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4426 static void __init enable_swap_cgroup(void)
4427 {
4428         if (!mem_cgroup_disabled() && really_do_swap_account)
4429                 do_swap_account = 1;
4430 }
4431 #else
4432 static void __init enable_swap_cgroup(void)
4433 {
4434 }
4435 #endif
4436
4437 static int mem_cgroup_soft_limit_tree_init(void)
4438 {
4439         struct mem_cgroup_tree_per_node *rtpn;
4440         struct mem_cgroup_tree_per_zone *rtpz;
4441         int tmp, node, zone;
4442
4443         for_each_node_state(node, N_POSSIBLE) {
4444                 tmp = node;
4445                 if (!node_state(node, N_NORMAL_MEMORY))
4446                         tmp = -1;
4447                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4448                 if (!rtpn)
4449                         return 1;
4450
4451                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4452
4453                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4454                         rtpz = &rtpn->rb_tree_per_zone[zone];
4455                         rtpz->rb_root = RB_ROOT;
4456                         spin_lock_init(&rtpz->lock);
4457                 }
4458         }
4459         return 0;
4460 }
4461
4462 static struct cgroup_subsys_state * __ref
4463 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4464 {
4465         struct mem_cgroup *mem, *parent;
4466         long error = -ENOMEM;
4467         int node;
4468
4469         mem = mem_cgroup_alloc();
4470         if (!mem)
4471                 return ERR_PTR(error);
4472
4473         for_each_node_state(node, N_POSSIBLE)
4474                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4475                         goto free_out;
4476
4477         /* root ? */
4478         if (cont->parent == NULL) {
4479                 int cpu;
4480                 enable_swap_cgroup();
4481                 parent = NULL;
4482                 root_mem_cgroup = mem;
4483                 if (mem_cgroup_soft_limit_tree_init())
4484                         goto free_out;
4485                 for_each_possible_cpu(cpu) {
4486                         struct memcg_stock_pcp *stock =
4487                                                 &per_cpu(memcg_stock, cpu);
4488                         INIT_WORK(&stock->work, drain_local_stock);
4489                 }
4490                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4491         } else {
4492                 parent = mem_cgroup_from_cont(cont->parent);
4493                 mem->use_hierarchy = parent->use_hierarchy;
4494                 mem->oom_kill_disable = parent->oom_kill_disable;
4495         }
4496
4497         if (parent && parent->use_hierarchy) {
4498                 res_counter_init(&mem->res, &parent->res);
4499                 res_counter_init(&mem->memsw, &parent->memsw);
4500                 /*
4501                  * We increment refcnt of the parent to ensure that we can
4502                  * safely access it on res_counter_charge/uncharge.
4503                  * This refcnt will be decremented when freeing this
4504                  * mem_cgroup(see mem_cgroup_put).
4505                  */
4506                 mem_cgroup_get(parent);
4507         } else {
4508                 res_counter_init(&mem->res, NULL);
4509                 res_counter_init(&mem->memsw, NULL);
4510         }
4511         mem->last_scanned_child = 0;
4512         INIT_LIST_HEAD(&mem->oom_notify);
4513
4514         if (parent)
4515                 mem->swappiness = get_swappiness(parent);
4516         atomic_set(&mem->refcnt, 1);
4517         mem->move_charge_at_immigrate = 0;
4518         mutex_init(&mem->thresholds_lock);
4519         return &mem->css;
4520 free_out:
4521         __mem_cgroup_free(mem);
4522         root_mem_cgroup = NULL;
4523         return ERR_PTR(error);
4524 }
4525
4526 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4527                                         struct cgroup *cont)
4528 {
4529         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4530
4531         return mem_cgroup_force_empty(mem, false);
4532 }
4533
4534 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4535                                 struct cgroup *cont)
4536 {
4537         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4538
4539         mem_cgroup_put(mem);
4540 }
4541
4542 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4543                                 struct cgroup *cont)
4544 {
4545         int ret;
4546
4547         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4548                                 ARRAY_SIZE(mem_cgroup_files));
4549
4550         if (!ret)
4551                 ret = register_memsw_files(cont, ss);
4552         return ret;
4553 }
4554
4555 #ifdef CONFIG_MMU
4556 /* Handlers for move charge at task migration. */
4557 #define PRECHARGE_COUNT_AT_ONCE 256
4558 static int mem_cgroup_do_precharge(unsigned long count)
4559 {
4560         int ret = 0;
4561         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4562         struct mem_cgroup *mem = mc.to;
4563
4564         if (mem_cgroup_is_root(mem)) {
4565                 mc.precharge += count;
4566                 /* we don't need css_get for root */
4567                 return ret;
4568         }
4569         /* try to charge at once */
4570         if (count > 1) {
4571                 struct res_counter *dummy;
4572                 /*
4573                  * "mem" cannot be under rmdir() because we've already checked
4574                  * by cgroup_lock_live_cgroup() that it is not removed and we
4575                  * are still under the same cgroup_mutex. So we can postpone
4576                  * css_get().
4577                  */
4578                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4579                         goto one_by_one;
4580                 if (do_swap_account && res_counter_charge(&mem->memsw,
4581                                                 PAGE_SIZE * count, &dummy)) {
4582                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4583                         goto one_by_one;
4584                 }
4585                 mc.precharge += count;
4586                 return ret;
4587         }
4588 one_by_one:
4589         /* fall back to one by one charge */
4590         while (count--) {
4591                 if (signal_pending(current)) {
4592                         ret = -EINTR;
4593                         break;
4594                 }
4595                 if (!batch_count--) {
4596                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4597                         cond_resched();
4598                 }
4599                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4600                                               PAGE_SIZE);
4601                 if (ret || !mem)
4602                         /* mem_cgroup_clear_mc() will do uncharge later */
4603                         return -ENOMEM;
4604                 mc.precharge++;
4605         }
4606         return ret;
4607 }
4608
4609 /**
4610  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4611  * @vma: the vma the pte to be checked belongs
4612  * @addr: the address corresponding to the pte to be checked
4613  * @ptent: the pte to be checked
4614  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4615  *
4616  * Returns
4617  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4618  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4619  *     move charge. if @target is not NULL, the page is stored in target->page
4620  *     with extra refcnt got(Callers should handle it).
4621  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4622  *     target for charge migration. if @target is not NULL, the entry is stored
4623  *     in target->ent.
4624  *
4625  * Called with pte lock held.
4626  */
4627 union mc_target {
4628         struct page     *page;
4629         swp_entry_t     ent;
4630 };
4631
4632 enum mc_target_type {
4633         MC_TARGET_NONE, /* not used */
4634         MC_TARGET_PAGE,
4635         MC_TARGET_SWAP,
4636 };
4637
4638 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4639                                                 unsigned long addr, pte_t ptent)
4640 {
4641         struct page *page = vm_normal_page(vma, addr, ptent);
4642
4643         if (!page || !page_mapped(page))
4644                 return NULL;
4645         if (PageAnon(page)) {
4646                 /* we don't move shared anon */
4647                 if (!move_anon() || page_mapcount(page) > 2)
4648                         return NULL;
4649         } else if (!move_file())
4650                 /* we ignore mapcount for file pages */
4651                 return NULL;
4652         if (!get_page_unless_zero(page))
4653                 return NULL;
4654
4655         return page;
4656 }
4657
4658 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4659                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4660 {
4661         int usage_count;
4662         struct page *page = NULL;
4663         swp_entry_t ent = pte_to_swp_entry(ptent);
4664
4665         if (!move_anon() || non_swap_entry(ent))
4666                 return NULL;
4667         usage_count = mem_cgroup_count_swap_user(ent, &page);
4668         if (usage_count > 1) { /* we don't move shared anon */
4669                 if (page)
4670                         put_page(page);
4671                 return NULL;
4672         }
4673         if (do_swap_account)
4674                 entry->val = ent.val;
4675
4676         return page;
4677 }
4678
4679 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4680                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4681 {
4682         struct page *page = NULL;
4683         struct inode *inode;
4684         struct address_space *mapping;
4685         pgoff_t pgoff;
4686
4687         if (!vma->vm_file) /* anonymous vma */
4688                 return NULL;
4689         if (!move_file())
4690                 return NULL;
4691
4692         inode = vma->vm_file->f_path.dentry->d_inode;
4693         mapping = vma->vm_file->f_mapping;
4694         if (pte_none(ptent))
4695                 pgoff = linear_page_index(vma, addr);
4696         else /* pte_file(ptent) is true */
4697                 pgoff = pte_to_pgoff(ptent);
4698
4699         /* page is moved even if it's not RSS of this task(page-faulted). */
4700         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4701                 page = find_get_page(mapping, pgoff);
4702         } else { /* shmem/tmpfs file. we should take account of swap too. */
4703                 swp_entry_t ent;
4704                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4705                 if (do_swap_account)
4706                         entry->val = ent.val;
4707         }
4708
4709         return page;
4710 }
4711
4712 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4713                 unsigned long addr, pte_t ptent, union mc_target *target)
4714 {
4715         struct page *page = NULL;
4716         struct page_cgroup *pc;
4717         int ret = 0;
4718         swp_entry_t ent = { .val = 0 };
4719
4720         if (pte_present(ptent))
4721                 page = mc_handle_present_pte(vma, addr, ptent);
4722         else if (is_swap_pte(ptent))
4723                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4724         else if (pte_none(ptent) || pte_file(ptent))
4725                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4726
4727         if (!page && !ent.val)
4728                 return 0;
4729         if (page) {
4730                 pc = lookup_page_cgroup(page);
4731                 /*
4732                  * Do only loose check w/o page_cgroup lock.
4733                  * mem_cgroup_move_account() checks the pc is valid or not under
4734                  * the lock.
4735                  */
4736                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4737                         ret = MC_TARGET_PAGE;
4738                         if (target)
4739                                 target->page = page;
4740                 }
4741                 if (!ret || !target)
4742                         put_page(page);
4743         }
4744         /* There is a swap entry and a page doesn't exist or isn't charged */
4745         if (ent.val && !ret &&
4746                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4747                 ret = MC_TARGET_SWAP;
4748                 if (target)
4749                         target->ent = ent;
4750         }
4751         return ret;
4752 }
4753
4754 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4755                                         unsigned long addr, unsigned long end,
4756                                         struct mm_walk *walk)
4757 {
4758         struct vm_area_struct *vma = walk->private;
4759         pte_t *pte;
4760         spinlock_t *ptl;
4761
4762         split_huge_page_pmd(walk->mm, pmd);
4763
4764         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4765         for (; addr != end; pte++, addr += PAGE_SIZE)
4766                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4767                         mc.precharge++; /* increment precharge temporarily */
4768         pte_unmap_unlock(pte - 1, ptl);
4769         cond_resched();
4770
4771         return 0;
4772 }
4773
4774 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4775 {
4776         unsigned long precharge;
4777         struct vm_area_struct *vma;
4778
4779         down_read(&mm->mmap_sem);
4780         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4781                 struct mm_walk mem_cgroup_count_precharge_walk = {
4782                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4783                         .mm = mm,
4784                         .private = vma,
4785                 };
4786                 if (is_vm_hugetlb_page(vma))
4787                         continue;
4788                 walk_page_range(vma->vm_start, vma->vm_end,
4789                                         &mem_cgroup_count_precharge_walk);
4790         }
4791         up_read(&mm->mmap_sem);
4792
4793         precharge = mc.precharge;
4794         mc.precharge = 0;
4795
4796         return precharge;
4797 }
4798
4799 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4800 {
4801         unsigned long precharge = mem_cgroup_count_precharge(mm);
4802
4803         VM_BUG_ON(mc.moving_task);
4804         mc.moving_task = current;
4805         return mem_cgroup_do_precharge(precharge);
4806 }
4807
4808 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4809 static void __mem_cgroup_clear_mc(void)
4810 {
4811         struct mem_cgroup *from = mc.from;
4812         struct mem_cgroup *to = mc.to;
4813
4814         /* we must uncharge all the leftover precharges from mc.to */
4815         if (mc.precharge) {
4816                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4817                 mc.precharge = 0;
4818         }
4819         /*
4820          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4821          * we must uncharge here.
4822          */
4823         if (mc.moved_charge) {
4824                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4825                 mc.moved_charge = 0;
4826         }
4827         /* we must fixup refcnts and charges */
4828         if (mc.moved_swap) {
4829                 /* uncharge swap account from the old cgroup */
4830                 if (!mem_cgroup_is_root(mc.from))
4831                         res_counter_uncharge(&mc.from->memsw,
4832                                                 PAGE_SIZE * mc.moved_swap);
4833                 __mem_cgroup_put(mc.from, mc.moved_swap);
4834
4835                 if (!mem_cgroup_is_root(mc.to)) {
4836                         /*
4837                          * we charged both to->res and to->memsw, so we should
4838                          * uncharge to->res.
4839                          */
4840                         res_counter_uncharge(&mc.to->res,
4841                                                 PAGE_SIZE * mc.moved_swap);
4842                 }
4843                 /* we've already done mem_cgroup_get(mc.to) */
4844                 mc.moved_swap = 0;
4845         }
4846         memcg_oom_recover(from);
4847         memcg_oom_recover(to);
4848         wake_up_all(&mc.waitq);
4849 }
4850
4851 static void mem_cgroup_clear_mc(void)
4852 {
4853         struct mem_cgroup *from = mc.from;
4854
4855         /*
4856          * we must clear moving_task before waking up waiters at the end of
4857          * task migration.
4858          */
4859         mc.moving_task = NULL;
4860         __mem_cgroup_clear_mc();
4861         spin_lock(&mc.lock);
4862         mc.from = NULL;
4863         mc.to = NULL;
4864         spin_unlock(&mc.lock);
4865         mem_cgroup_end_move(from);
4866 }
4867
4868 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4869                                 struct cgroup *cgroup,
4870                                 struct task_struct *p,
4871                                 bool threadgroup)
4872 {
4873         int ret = 0;
4874         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4875
4876         if (mem->move_charge_at_immigrate) {
4877                 struct mm_struct *mm;
4878                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4879
4880                 VM_BUG_ON(from == mem);
4881
4882                 mm = get_task_mm(p);
4883                 if (!mm)
4884                         return 0;
4885                 /* We move charges only when we move a owner of the mm */
4886                 if (mm->owner == p) {
4887                         VM_BUG_ON(mc.from);
4888                         VM_BUG_ON(mc.to);
4889                         VM_BUG_ON(mc.precharge);
4890                         VM_BUG_ON(mc.moved_charge);
4891                         VM_BUG_ON(mc.moved_swap);
4892                         mem_cgroup_start_move(from);
4893                         spin_lock(&mc.lock);
4894                         mc.from = from;
4895                         mc.to = mem;
4896                         spin_unlock(&mc.lock);
4897                         /* We set mc.moving_task later */
4898
4899                         ret = mem_cgroup_precharge_mc(mm);
4900                         if (ret)
4901                                 mem_cgroup_clear_mc();
4902                 }
4903                 mmput(mm);
4904         }
4905         return ret;
4906 }
4907
4908 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4909                                 struct cgroup *cgroup,
4910                                 struct task_struct *p,
4911                                 bool threadgroup)
4912 {
4913         mem_cgroup_clear_mc();
4914 }
4915
4916 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4917                                 unsigned long addr, unsigned long end,
4918                                 struct mm_walk *walk)
4919 {
4920         int ret = 0;
4921         struct vm_area_struct *vma = walk->private;
4922         pte_t *pte;
4923         spinlock_t *ptl;
4924
4925         split_huge_page_pmd(walk->mm, pmd);
4926 retry:
4927         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4928         for (; addr != end; addr += PAGE_SIZE) {
4929                 pte_t ptent = *(pte++);
4930                 union mc_target target;
4931                 int type;
4932                 struct page *page;
4933                 struct page_cgroup *pc;
4934                 swp_entry_t ent;
4935
4936                 if (!mc.precharge)
4937                         break;
4938
4939                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4940                 switch (type) {
4941                 case MC_TARGET_PAGE:
4942                         page = target.page;
4943                         if (isolate_lru_page(page))
4944                                 goto put;
4945                         pc = lookup_page_cgroup(page);
4946                         if (!mem_cgroup_move_account(page, pc,
4947                                         mc.from, mc.to, false, PAGE_SIZE)) {
4948                                 mc.precharge--;
4949                                 /* we uncharge from mc.from later. */
4950                                 mc.moved_charge++;
4951                         }
4952                         putback_lru_page(page);
4953 put:                    /* is_target_pte_for_mc() gets the page */
4954                         put_page(page);
4955                         break;
4956                 case MC_TARGET_SWAP:
4957                         ent = target.ent;
4958                         if (!mem_cgroup_move_swap_account(ent,
4959                                                 mc.from, mc.to, false)) {
4960                                 mc.precharge--;
4961                                 /* we fixup refcnts and charges later. */
4962                                 mc.moved_swap++;
4963                         }
4964                         break;
4965                 default:
4966                         break;
4967                 }
4968         }
4969         pte_unmap_unlock(pte - 1, ptl);
4970         cond_resched();
4971
4972         if (addr != end) {
4973                 /*
4974                  * We have consumed all precharges we got in can_attach().
4975                  * We try charge one by one, but don't do any additional
4976                  * charges to mc.to if we have failed in charge once in attach()
4977                  * phase.
4978                  */
4979                 ret = mem_cgroup_do_precharge(1);
4980                 if (!ret)
4981                         goto retry;
4982         }
4983
4984         return ret;
4985 }
4986
4987 static void mem_cgroup_move_charge(struct mm_struct *mm)
4988 {
4989         struct vm_area_struct *vma;
4990
4991         lru_add_drain_all();
4992 retry:
4993         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4994                 /*
4995                  * Someone who are holding the mmap_sem might be waiting in
4996                  * waitq. So we cancel all extra charges, wake up all waiters,
4997                  * and retry. Because we cancel precharges, we might not be able
4998                  * to move enough charges, but moving charge is a best-effort
4999                  * feature anyway, so it wouldn't be a big problem.
5000                  */
5001                 __mem_cgroup_clear_mc();
5002                 cond_resched();
5003                 goto retry;
5004         }
5005         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5006                 int ret;
5007                 struct mm_walk mem_cgroup_move_charge_walk = {
5008                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5009                         .mm = mm,
5010                         .private = vma,
5011                 };
5012                 if (is_vm_hugetlb_page(vma))
5013                         continue;
5014                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5015                                                 &mem_cgroup_move_charge_walk);
5016                 if (ret)
5017                         /*
5018                          * means we have consumed all precharges and failed in
5019                          * doing additional charge. Just abandon here.
5020                          */
5021                         break;
5022         }
5023         up_read(&mm->mmap_sem);
5024 }
5025
5026 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5027                                 struct cgroup *cont,
5028                                 struct cgroup *old_cont,
5029                                 struct task_struct *p,
5030                                 bool threadgroup)
5031 {
5032         struct mm_struct *mm;
5033
5034         if (!mc.to)
5035                 /* no need to move charge */
5036                 return;
5037
5038         mm = get_task_mm(p);
5039         if (mm) {
5040                 mem_cgroup_move_charge(mm);
5041                 mmput(mm);
5042         }
5043         mem_cgroup_clear_mc();
5044 }
5045 #else   /* !CONFIG_MMU */
5046 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5047                                 struct cgroup *cgroup,
5048                                 struct task_struct *p,
5049                                 bool threadgroup)
5050 {
5051         return 0;
5052 }
5053 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5054                                 struct cgroup *cgroup,
5055                                 struct task_struct *p,
5056                                 bool threadgroup)
5057 {
5058 }
5059 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5060                                 struct cgroup *cont,
5061                                 struct cgroup *old_cont,
5062                                 struct task_struct *p,
5063                                 bool threadgroup)
5064 {
5065 }
5066 #endif
5067
5068 struct cgroup_subsys mem_cgroup_subsys = {
5069         .name = "memory",
5070         .subsys_id = mem_cgroup_subsys_id,
5071         .create = mem_cgroup_create,
5072         .pre_destroy = mem_cgroup_pre_destroy,
5073         .destroy = mem_cgroup_destroy,
5074         .populate = mem_cgroup_populate,
5075         .can_attach = mem_cgroup_can_attach,
5076         .cancel_attach = mem_cgroup_cancel_attach,
5077         .attach = mem_cgroup_move_task,
5078         .early_init = 0,
5079         .use_id = 1,
5080 };
5081
5082 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5083 static int __init enable_swap_account(char *s)
5084 {
5085         /* consider enabled if no parameter or 1 is given */
5086         if (!(*s) || !strcmp(s, "=1"))
5087                 really_do_swap_account = 1;
5088         else if (!strcmp(s, "=0"))
5089                 really_do_swap_account = 0;
5090         return 1;
5091 }
5092 __setup("swapaccount", enable_swap_account);
5093
5094 static int __init disable_swap_account(char *s)
5095 {
5096         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5097         enable_swap_account("=0");
5098         return 1;
5099 }
5100 __setup("noswapaccount", disable_swap_account);
5101 #endif