memcg: charged pages always have valid per-memcg zone info
[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         /*
223           protect against reclaim related member.
224         */
225         spinlock_t reclaim_param_lock;
226
227         /*
228          * While reclaiming in a hierarchy, we cache the last child we
229          * reclaimed from.
230          */
231         int last_scanned_child;
232         /*
233          * Should the accounting and control be hierarchical, per subtree?
234          */
235         bool use_hierarchy;
236         atomic_t        oom_lock;
237         atomic_t        refcnt;
238
239         unsigned int    swappiness;
240         /* OOM-Killer disable */
241         int             oom_kill_disable;
242
243         /* set when res.limit == memsw.limit */
244         bool            memsw_is_minimum;
245
246         /* protect arrays of thresholds */
247         struct mutex thresholds_lock;
248
249         /* thresholds for memory usage. RCU-protected */
250         struct mem_cgroup_thresholds thresholds;
251
252         /* thresholds for mem+swap usage. RCU-protected */
253         struct mem_cgroup_thresholds memsw_thresholds;
254
255         /* For oom notifier event fd */
256         struct list_head oom_notify;
257
258         /*
259          * Should we move charges of a task when a task is moved into this
260          * mem_cgroup ? And what type of charges should we move ?
261          */
262         unsigned long   move_charge_at_immigrate;
263         /*
264          * percpu counter.
265          */
266         struct mem_cgroup_stat_cpu *stat;
267         /*
268          * used when a cpu is offlined or other synchronizations
269          * See mem_cgroup_read_stat().
270          */
271         struct mem_cgroup_stat_cpu nocpu_base;
272         spinlock_t pcp_counter_lock;
273 };
274
275 /* Stuffs for move charges at task migration. */
276 /*
277  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278  * left-shifted bitmap of these types.
279  */
280 enum move_type {
281         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
282         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
283         NR_MOVE_TYPE,
284 };
285
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct {
288         spinlock_t        lock; /* for from, to */
289         struct mem_cgroup *from;
290         struct mem_cgroup *to;
291         unsigned long precharge;
292         unsigned long moved_charge;
293         unsigned long moved_swap;
294         struct task_struct *moving_task;        /* a task moving charges */
295         wait_queue_head_t waitq;                /* a waitq for other context */
296 } mc = {
297         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
298         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
299 };
300
301 static bool move_anon(void)
302 {
303         return test_bit(MOVE_CHARGE_TYPE_ANON,
304                                         &mc.to->move_charge_at_immigrate);
305 }
306
307 static bool move_file(void)
308 {
309         return test_bit(MOVE_CHARGE_TYPE_FILE,
310                                         &mc.to->move_charge_at_immigrate);
311 }
312
313 /*
314  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315  * limit reclaim to prevent infinite loops, if they ever occur.
316  */
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
319
320 enum charge_type {
321         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
322         MEM_CGROUP_CHARGE_TYPE_MAPPED,
323         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
324         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
325         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
326         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
327         NR_CHARGE_TYPE,
328 };
329
330 /* for encoding cft->private value on file */
331 #define _MEM                    (0)
332 #define _MEMSWAP                (1)
333 #define _OOM_TYPE               (2)
334 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
335 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
336 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
337 /* Used for OOM nofiier */
338 #define OOM_CONTROL             (0)
339
340 /*
341  * Reclaim flags for mem_cgroup_hierarchical_reclaim
342  */
343 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
344 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
345 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
346 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
347 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
348 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
349
350 static void mem_cgroup_get(struct mem_cgroup *mem);
351 static void mem_cgroup_put(struct mem_cgroup *mem);
352 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
353 static void drain_all_stock_async(void);
354
355 static struct mem_cgroup_per_zone *
356 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
357 {
358         return &mem->info.nodeinfo[nid]->zoneinfo[zid];
359 }
360
361 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
362 {
363         return &mem->css;
364 }
365
366 static struct mem_cgroup_per_zone *
367 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
368 {
369         int nid = page_to_nid(page);
370         int zid = page_zonenum(page);
371
372         return mem_cgroup_zoneinfo(mem, nid, zid);
373 }
374
375 static struct mem_cgroup_tree_per_zone *
376 soft_limit_tree_node_zone(int nid, int zid)
377 {
378         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
379 }
380
381 static struct mem_cgroup_tree_per_zone *
382 soft_limit_tree_from_page(struct page *page)
383 {
384         int nid = page_to_nid(page);
385         int zid = page_zonenum(page);
386
387         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
388 }
389
390 static void
391 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
392                                 struct mem_cgroup_per_zone *mz,
393                                 struct mem_cgroup_tree_per_zone *mctz,
394                                 unsigned long long new_usage_in_excess)
395 {
396         struct rb_node **p = &mctz->rb_root.rb_node;
397         struct rb_node *parent = NULL;
398         struct mem_cgroup_per_zone *mz_node;
399
400         if (mz->on_tree)
401                 return;
402
403         mz->usage_in_excess = new_usage_in_excess;
404         if (!mz->usage_in_excess)
405                 return;
406         while (*p) {
407                 parent = *p;
408                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
409                                         tree_node);
410                 if (mz->usage_in_excess < mz_node->usage_in_excess)
411                         p = &(*p)->rb_left;
412                 /*
413                  * We can't avoid mem cgroups that are over their soft
414                  * limit by the same amount
415                  */
416                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
417                         p = &(*p)->rb_right;
418         }
419         rb_link_node(&mz->tree_node, parent, p);
420         rb_insert_color(&mz->tree_node, &mctz->rb_root);
421         mz->on_tree = true;
422 }
423
424 static void
425 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
426                                 struct mem_cgroup_per_zone *mz,
427                                 struct mem_cgroup_tree_per_zone *mctz)
428 {
429         if (!mz->on_tree)
430                 return;
431         rb_erase(&mz->tree_node, &mctz->rb_root);
432         mz->on_tree = false;
433 }
434
435 static void
436 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
437                                 struct mem_cgroup_per_zone *mz,
438                                 struct mem_cgroup_tree_per_zone *mctz)
439 {
440         spin_lock(&mctz->lock);
441         __mem_cgroup_remove_exceeded(mem, mz, mctz);
442         spin_unlock(&mctz->lock);
443 }
444
445
446 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
447 {
448         unsigned long long excess;
449         struct mem_cgroup_per_zone *mz;
450         struct mem_cgroup_tree_per_zone *mctz;
451         int nid = page_to_nid(page);
452         int zid = page_zonenum(page);
453         mctz = soft_limit_tree_from_page(page);
454
455         /*
456          * Necessary to update all ancestors when hierarchy is used.
457          * because their event counter is not touched.
458          */
459         for (; mem; mem = parent_mem_cgroup(mem)) {
460                 mz = mem_cgroup_zoneinfo(mem, nid, zid);
461                 excess = res_counter_soft_limit_excess(&mem->res);
462                 /*
463                  * We have to update the tree if mz is on RB-tree or
464                  * mem is over its softlimit.
465                  */
466                 if (excess || mz->on_tree) {
467                         spin_lock(&mctz->lock);
468                         /* if on-tree, remove it */
469                         if (mz->on_tree)
470                                 __mem_cgroup_remove_exceeded(mem, mz, mctz);
471                         /*
472                          * Insert again. mz->usage_in_excess will be updated.
473                          * If excess is 0, no tree ops.
474                          */
475                         __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
476                         spin_unlock(&mctz->lock);
477                 }
478         }
479 }
480
481 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
482 {
483         int node, zone;
484         struct mem_cgroup_per_zone *mz;
485         struct mem_cgroup_tree_per_zone *mctz;
486
487         for_each_node_state(node, N_POSSIBLE) {
488                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
489                         mz = mem_cgroup_zoneinfo(mem, node, zone);
490                         mctz = soft_limit_tree_node_zone(node, zone);
491                         mem_cgroup_remove_exceeded(mem, mz, mctz);
492                 }
493         }
494 }
495
496 static struct mem_cgroup_per_zone *
497 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
498 {
499         struct rb_node *rightmost = NULL;
500         struct mem_cgroup_per_zone *mz;
501
502 retry:
503         mz = NULL;
504         rightmost = rb_last(&mctz->rb_root);
505         if (!rightmost)
506                 goto done;              /* Nothing to reclaim from */
507
508         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
509         /*
510          * Remove the node now but someone else can add it back,
511          * we will to add it back at the end of reclaim to its correct
512          * position in the tree.
513          */
514         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
515         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
516                 !css_tryget(&mz->mem->css))
517                 goto retry;
518 done:
519         return mz;
520 }
521
522 static struct mem_cgroup_per_zone *
523 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
524 {
525         struct mem_cgroup_per_zone *mz;
526
527         spin_lock(&mctz->lock);
528         mz = __mem_cgroup_largest_soft_limit_node(mctz);
529         spin_unlock(&mctz->lock);
530         return mz;
531 }
532
533 /*
534  * Implementation Note: reading percpu statistics for memcg.
535  *
536  * Both of vmstat[] and percpu_counter has threshold and do periodic
537  * synchronization to implement "quick" read. There are trade-off between
538  * reading cost and precision of value. Then, we may have a chance to implement
539  * a periodic synchronizion of counter in memcg's counter.
540  *
541  * But this _read() function is used for user interface now. The user accounts
542  * memory usage by memory cgroup and he _always_ requires exact value because
543  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
544  * have to visit all online cpus and make sum. So, for now, unnecessary
545  * synchronization is not implemented. (just implemented for cpu hotplug)
546  *
547  * If there are kernel internal actions which can make use of some not-exact
548  * value, and reading all cpu value can be performance bottleneck in some
549  * common workload, threashold and synchonization as vmstat[] should be
550  * implemented.
551  */
552 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
553                 enum mem_cgroup_stat_index idx)
554 {
555         int cpu;
556         s64 val = 0;
557
558         get_online_cpus();
559         for_each_online_cpu(cpu)
560                 val += per_cpu(mem->stat->count[idx], cpu);
561 #ifdef CONFIG_HOTPLUG_CPU
562         spin_lock(&mem->pcp_counter_lock);
563         val += mem->nocpu_base.count[idx];
564         spin_unlock(&mem->pcp_counter_lock);
565 #endif
566         put_online_cpus();
567         return val;
568 }
569
570 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
571 {
572         s64 ret;
573
574         ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
575         ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
576         return ret;
577 }
578
579 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
580                                          bool charge)
581 {
582         int val = (charge) ? 1 : -1;
583         this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
584 }
585
586 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
587                                          bool file, int nr_pages)
588 {
589         preempt_disable();
590
591         if (file)
592                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
593         else
594                 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
595
596         /* pagein of a big page is an event. So, ignore page size */
597         if (nr_pages > 0)
598                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
599         else {
600                 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
601                 nr_pages = -nr_pages; /* for event */
602         }
603
604         __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages);
605
606         preempt_enable();
607 }
608
609 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
610                                         enum lru_list idx)
611 {
612         int nid, zid;
613         struct mem_cgroup_per_zone *mz;
614         u64 total = 0;
615
616         for_each_online_node(nid)
617                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
618                         mz = mem_cgroup_zoneinfo(mem, nid, zid);
619                         total += MEM_CGROUP_ZSTAT(mz, idx);
620                 }
621         return total;
622 }
623
624 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
625 {
626         s64 val;
627
628         val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
629
630         return !(val & ((1 << event_mask_shift) - 1));
631 }
632
633 /*
634  * Check events in order.
635  *
636  */
637 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
638 {
639         /* threshold event is triggered in finer grain than soft limit */
640         if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
641                 mem_cgroup_threshold(mem);
642                 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
643                         mem_cgroup_update_tree(mem, page);
644         }
645 }
646
647 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
648 {
649         return container_of(cgroup_subsys_state(cont,
650                                 mem_cgroup_subsys_id), struct mem_cgroup,
651                                 css);
652 }
653
654 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
655 {
656         /*
657          * mm_update_next_owner() may clear mm->owner to NULL
658          * if it races with swapoff, page migration, etc.
659          * So this can be called with p == NULL.
660          */
661         if (unlikely(!p))
662                 return NULL;
663
664         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
665                                 struct mem_cgroup, css);
666 }
667
668 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
669 {
670         struct mem_cgroup *mem = NULL;
671
672         if (!mm)
673                 return NULL;
674         /*
675          * Because we have no locks, mm->owner's may be being moved to other
676          * cgroup. We use css_tryget() here even if this looks
677          * pessimistic (rather than adding locks here).
678          */
679         rcu_read_lock();
680         do {
681                 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
682                 if (unlikely(!mem))
683                         break;
684         } while (!css_tryget(&mem->css));
685         rcu_read_unlock();
686         return mem;
687 }
688
689 /* The caller has to guarantee "mem" exists before calling this */
690 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
691 {
692         struct cgroup_subsys_state *css;
693         int found;
694
695         if (!mem) /* ROOT cgroup has the smallest ID */
696                 return root_mem_cgroup; /*css_put/get against root is ignored*/
697         if (!mem->use_hierarchy) {
698                 if (css_tryget(&mem->css))
699                         return mem;
700                 return NULL;
701         }
702         rcu_read_lock();
703         /*
704          * searching a memory cgroup which has the smallest ID under given
705          * ROOT cgroup. (ID >= 1)
706          */
707         css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
708         if (css && css_tryget(css))
709                 mem = container_of(css, struct mem_cgroup, css);
710         else
711                 mem = NULL;
712         rcu_read_unlock();
713         return mem;
714 }
715
716 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
717                                         struct mem_cgroup *root,
718                                         bool cond)
719 {
720         int nextid = css_id(&iter->css) + 1;
721         int found;
722         int hierarchy_used;
723         struct cgroup_subsys_state *css;
724
725         hierarchy_used = iter->use_hierarchy;
726
727         css_put(&iter->css);
728         /* If no ROOT, walk all, ignore hierarchy */
729         if (!cond || (root && !hierarchy_used))
730                 return NULL;
731
732         if (!root)
733                 root = root_mem_cgroup;
734
735         do {
736                 iter = NULL;
737                 rcu_read_lock();
738
739                 css = css_get_next(&mem_cgroup_subsys, nextid,
740                                 &root->css, &found);
741                 if (css && css_tryget(css))
742                         iter = container_of(css, struct mem_cgroup, css);
743                 rcu_read_unlock();
744                 /* If css is NULL, no more cgroups will be found */
745                 nextid = found + 1;
746         } while (css && !iter);
747
748         return iter;
749 }
750 /*
751  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
752  * be careful that "break" loop is not allowed. We have reference count.
753  * Instead of that modify "cond" to be false and "continue" to exit the loop.
754  */
755 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
756         for (iter = mem_cgroup_start_loop(root);\
757              iter != NULL;\
758              iter = mem_cgroup_get_next(iter, root, cond))
759
760 #define for_each_mem_cgroup_tree(iter, root) \
761         for_each_mem_cgroup_tree_cond(iter, root, true)
762
763 #define for_each_mem_cgroup_all(iter) \
764         for_each_mem_cgroup_tree_cond(iter, NULL, true)
765
766
767 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
768 {
769         return (mem == root_mem_cgroup);
770 }
771
772 /*
773  * Following LRU functions are allowed to be used without PCG_LOCK.
774  * Operations are called by routine of global LRU independently from memcg.
775  * What we have to take care of here is validness of pc->mem_cgroup.
776  *
777  * Changes to pc->mem_cgroup happens when
778  * 1. charge
779  * 2. moving account
780  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
781  * It is added to LRU before charge.
782  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
783  * When moving account, the page is not on LRU. It's isolated.
784  */
785
786 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
787 {
788         struct page_cgroup *pc;
789         struct mem_cgroup_per_zone *mz;
790
791         if (mem_cgroup_disabled())
792                 return;
793         pc = lookup_page_cgroup(page);
794         /* can happen while we handle swapcache. */
795         if (!TestClearPageCgroupAcctLRU(pc))
796                 return;
797         VM_BUG_ON(!pc->mem_cgroup);
798         /*
799          * We don't check PCG_USED bit. It's cleared when the "page" is finally
800          * removed from global LRU.
801          */
802         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
803         /* huge page split is done under lru_lock. so, we have no races. */
804         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
805         if (mem_cgroup_is_root(pc->mem_cgroup))
806                 return;
807         VM_BUG_ON(list_empty(&pc->lru));
808         list_del_init(&pc->lru);
809 }
810
811 void mem_cgroup_del_lru(struct page *page)
812 {
813         mem_cgroup_del_lru_list(page, page_lru(page));
814 }
815
816 /*
817  * Writeback is about to end against a page which has been marked for immediate
818  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
819  * inactive list.
820  */
821 void mem_cgroup_rotate_reclaimable_page(struct page *page)
822 {
823         struct mem_cgroup_per_zone *mz;
824         struct page_cgroup *pc;
825         enum lru_list lru = page_lru(page);
826
827         if (mem_cgroup_disabled())
828                 return;
829
830         pc = lookup_page_cgroup(page);
831         /* unused or root page is not rotated. */
832         if (!PageCgroupUsed(pc))
833                 return;
834         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
835         smp_rmb();
836         if (mem_cgroup_is_root(pc->mem_cgroup))
837                 return;
838         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
839         list_move_tail(&pc->lru, &mz->lists[lru]);
840 }
841
842 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
843 {
844         struct mem_cgroup_per_zone *mz;
845         struct page_cgroup *pc;
846
847         if (mem_cgroup_disabled())
848                 return;
849
850         pc = lookup_page_cgroup(page);
851         /* unused or root page is not rotated. */
852         if (!PageCgroupUsed(pc))
853                 return;
854         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
855         smp_rmb();
856         if (mem_cgroup_is_root(pc->mem_cgroup))
857                 return;
858         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
859         list_move(&pc->lru, &mz->lists[lru]);
860 }
861
862 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
863 {
864         struct page_cgroup *pc;
865         struct mem_cgroup_per_zone *mz;
866
867         if (mem_cgroup_disabled())
868                 return;
869         pc = lookup_page_cgroup(page);
870         VM_BUG_ON(PageCgroupAcctLRU(pc));
871         if (!PageCgroupUsed(pc))
872                 return;
873         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
874         smp_rmb();
875         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
876         /* huge page split is done under lru_lock. so, we have no races. */
877         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
878         SetPageCgroupAcctLRU(pc);
879         if (mem_cgroup_is_root(pc->mem_cgroup))
880                 return;
881         list_add(&pc->lru, &mz->lists[lru]);
882 }
883
884 /*
885  * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
886  * lru because the page may.be reused after it's fully uncharged (because of
887  * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
888  * it again. This function is only used to charge SwapCache. It's done under
889  * lock_page and expected that zone->lru_lock is never held.
890  */
891 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
892 {
893         unsigned long flags;
894         struct zone *zone = page_zone(page);
895         struct page_cgroup *pc = lookup_page_cgroup(page);
896
897         spin_lock_irqsave(&zone->lru_lock, flags);
898         /*
899          * Forget old LRU when this page_cgroup is *not* used. This Used bit
900          * is guarded by lock_page() because the page is SwapCache.
901          */
902         if (!PageCgroupUsed(pc))
903                 mem_cgroup_del_lru_list(page, page_lru(page));
904         spin_unlock_irqrestore(&zone->lru_lock, flags);
905 }
906
907 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
908 {
909         unsigned long flags;
910         struct zone *zone = page_zone(page);
911         struct page_cgroup *pc = lookup_page_cgroup(page);
912
913         spin_lock_irqsave(&zone->lru_lock, flags);
914         /* link when the page is linked to LRU but page_cgroup isn't */
915         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
916                 mem_cgroup_add_lru_list(page, page_lru(page));
917         spin_unlock_irqrestore(&zone->lru_lock, flags);
918 }
919
920
921 void mem_cgroup_move_lists(struct page *page,
922                            enum lru_list from, enum lru_list to)
923 {
924         if (mem_cgroup_disabled())
925                 return;
926         mem_cgroup_del_lru_list(page, from);
927         mem_cgroup_add_lru_list(page, to);
928 }
929
930 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
931 {
932         int ret;
933         struct mem_cgroup *curr = NULL;
934         struct task_struct *p;
935
936         p = find_lock_task_mm(task);
937         if (!p)
938                 return 0;
939         curr = try_get_mem_cgroup_from_mm(p->mm);
940         task_unlock(p);
941         if (!curr)
942                 return 0;
943         /*
944          * We should check use_hierarchy of "mem" not "curr". Because checking
945          * use_hierarchy of "curr" here make this function true if hierarchy is
946          * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
947          * hierarchy(even if use_hierarchy is disabled in "mem").
948          */
949         if (mem->use_hierarchy)
950                 ret = css_is_ancestor(&curr->css, &mem->css);
951         else
952                 ret = (curr == mem);
953         css_put(&curr->css);
954         return ret;
955 }
956
957 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
958 {
959         unsigned long active;
960         unsigned long inactive;
961         unsigned long gb;
962         unsigned long inactive_ratio;
963
964         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
965         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
966
967         gb = (inactive + active) >> (30 - PAGE_SHIFT);
968         if (gb)
969                 inactive_ratio = int_sqrt(10 * gb);
970         else
971                 inactive_ratio = 1;
972
973         if (present_pages) {
974                 present_pages[0] = inactive;
975                 present_pages[1] = active;
976         }
977
978         return inactive_ratio;
979 }
980
981 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
982 {
983         unsigned long active;
984         unsigned long inactive;
985         unsigned long present_pages[2];
986         unsigned long inactive_ratio;
987
988         inactive_ratio = calc_inactive_ratio(memcg, present_pages);
989
990         inactive = present_pages[0];
991         active = present_pages[1];
992
993         if (inactive * inactive_ratio < active)
994                 return 1;
995
996         return 0;
997 }
998
999 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1000 {
1001         unsigned long active;
1002         unsigned long inactive;
1003
1004         inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1005         active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1006
1007         return (active > inactive);
1008 }
1009
1010 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
1011                                        struct zone *zone,
1012                                        enum lru_list lru)
1013 {
1014         int nid = zone_to_nid(zone);
1015         int zid = zone_idx(zone);
1016         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1017
1018         return MEM_CGROUP_ZSTAT(mz, lru);
1019 }
1020
1021 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1022                                                       struct zone *zone)
1023 {
1024         int nid = zone_to_nid(zone);
1025         int zid = zone_idx(zone);
1026         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1027
1028         return &mz->reclaim_stat;
1029 }
1030
1031 struct zone_reclaim_stat *
1032 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1033 {
1034         struct page_cgroup *pc;
1035         struct mem_cgroup_per_zone *mz;
1036
1037         if (mem_cgroup_disabled())
1038                 return NULL;
1039
1040         pc = lookup_page_cgroup(page);
1041         if (!PageCgroupUsed(pc))
1042                 return NULL;
1043         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1044         smp_rmb();
1045         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1046         return &mz->reclaim_stat;
1047 }
1048
1049 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1050                                         struct list_head *dst,
1051                                         unsigned long *scanned, int order,
1052                                         int mode, struct zone *z,
1053                                         struct mem_cgroup *mem_cont,
1054                                         int active, int file)
1055 {
1056         unsigned long nr_taken = 0;
1057         struct page *page;
1058         unsigned long scan;
1059         LIST_HEAD(pc_list);
1060         struct list_head *src;
1061         struct page_cgroup *pc, *tmp;
1062         int nid = zone_to_nid(z);
1063         int zid = zone_idx(z);
1064         struct mem_cgroup_per_zone *mz;
1065         int lru = LRU_FILE * file + active;
1066         int ret;
1067
1068         BUG_ON(!mem_cont);
1069         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1070         src = &mz->lists[lru];
1071
1072         scan = 0;
1073         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1074                 if (scan >= nr_to_scan)
1075                         break;
1076
1077                 if (unlikely(!PageCgroupUsed(pc)))
1078                         continue;
1079
1080                 page = lookup_cgroup_page(pc);
1081
1082                 if (unlikely(!PageLRU(page)))
1083                         continue;
1084
1085                 scan++;
1086                 ret = __isolate_lru_page(page, mode, file);
1087                 switch (ret) {
1088                 case 0:
1089                         list_move(&page->lru, dst);
1090                         mem_cgroup_del_lru(page);
1091                         nr_taken += hpage_nr_pages(page);
1092                         break;
1093                 case -EBUSY:
1094                         /* we don't affect global LRU but rotate in our LRU */
1095                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1096                         break;
1097                 default:
1098                         break;
1099                 }
1100         }
1101
1102         *scanned = scan;
1103
1104         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1105                                       0, 0, 0, mode);
1106
1107         return nr_taken;
1108 }
1109
1110 #define mem_cgroup_from_res_counter(counter, member)    \
1111         container_of(counter, struct mem_cgroup, member)
1112
1113 /**
1114  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1115  * @mem: the memory cgroup
1116  *
1117  * Returns the maximum amount of memory @mem can be charged with, in
1118  * bytes.
1119  */
1120 static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem)
1121 {
1122         unsigned long long margin;
1123
1124         margin = res_counter_margin(&mem->res);
1125         if (do_swap_account)
1126                 margin = min(margin, res_counter_margin(&mem->memsw));
1127         return margin;
1128 }
1129
1130 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1131 {
1132         struct cgroup *cgrp = memcg->css.cgroup;
1133         unsigned int swappiness;
1134
1135         /* root ? */
1136         if (cgrp->parent == NULL)
1137                 return vm_swappiness;
1138
1139         spin_lock(&memcg->reclaim_param_lock);
1140         swappiness = memcg->swappiness;
1141         spin_unlock(&memcg->reclaim_param_lock);
1142
1143         return swappiness;
1144 }
1145
1146 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1147 {
1148         int cpu;
1149
1150         get_online_cpus();
1151         spin_lock(&mem->pcp_counter_lock);
1152         for_each_online_cpu(cpu)
1153                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1154         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1155         spin_unlock(&mem->pcp_counter_lock);
1156         put_online_cpus();
1157
1158         synchronize_rcu();
1159 }
1160
1161 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1162 {
1163         int cpu;
1164
1165         if (!mem)
1166                 return;
1167         get_online_cpus();
1168         spin_lock(&mem->pcp_counter_lock);
1169         for_each_online_cpu(cpu)
1170                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1171         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1172         spin_unlock(&mem->pcp_counter_lock);
1173         put_online_cpus();
1174 }
1175 /*
1176  * 2 routines for checking "mem" is under move_account() or not.
1177  *
1178  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1179  *                        for avoiding race in accounting. If true,
1180  *                        pc->mem_cgroup may be overwritten.
1181  *
1182  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1183  *                        under hierarchy of moving cgroups. This is for
1184  *                        waiting at hith-memory prressure caused by "move".
1185  */
1186
1187 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1188 {
1189         VM_BUG_ON(!rcu_read_lock_held());
1190         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1191 }
1192
1193 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1194 {
1195         struct mem_cgroup *from;
1196         struct mem_cgroup *to;
1197         bool ret = false;
1198         /*
1199          * Unlike task_move routines, we access mc.to, mc.from not under
1200          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1201          */
1202         spin_lock(&mc.lock);
1203         from = mc.from;
1204         to = mc.to;
1205         if (!from)
1206                 goto unlock;
1207         if (from == mem || to == mem
1208             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1209             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1210                 ret = true;
1211 unlock:
1212         spin_unlock(&mc.lock);
1213         return ret;
1214 }
1215
1216 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1217 {
1218         if (mc.moving_task && current != mc.moving_task) {
1219                 if (mem_cgroup_under_move(mem)) {
1220                         DEFINE_WAIT(wait);
1221                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1222                         /* moving charge context might have finished. */
1223                         if (mc.moving_task)
1224                                 schedule();
1225                         finish_wait(&mc.waitq, &wait);
1226                         return true;
1227                 }
1228         }
1229         return false;
1230 }
1231
1232 /**
1233  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1234  * @memcg: The memory cgroup that went over limit
1235  * @p: Task that is going to be killed
1236  *
1237  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1238  * enabled
1239  */
1240 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1241 {
1242         struct cgroup *task_cgrp;
1243         struct cgroup *mem_cgrp;
1244         /*
1245          * Need a buffer in BSS, can't rely on allocations. The code relies
1246          * on the assumption that OOM is serialized for memory controller.
1247          * If this assumption is broken, revisit this code.
1248          */
1249         static char memcg_name[PATH_MAX];
1250         int ret;
1251
1252         if (!memcg || !p)
1253                 return;
1254
1255
1256         rcu_read_lock();
1257
1258         mem_cgrp = memcg->css.cgroup;
1259         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1260
1261         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1262         if (ret < 0) {
1263                 /*
1264                  * Unfortunately, we are unable to convert to a useful name
1265                  * But we'll still print out the usage information
1266                  */
1267                 rcu_read_unlock();
1268                 goto done;
1269         }
1270         rcu_read_unlock();
1271
1272         printk(KERN_INFO "Task in %s killed", memcg_name);
1273
1274         rcu_read_lock();
1275         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1276         if (ret < 0) {
1277                 rcu_read_unlock();
1278                 goto done;
1279         }
1280         rcu_read_unlock();
1281
1282         /*
1283          * Continues from above, so we don't need an KERN_ level
1284          */
1285         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1286 done:
1287
1288         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1289                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1290                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1291                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1292         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1293                 "failcnt %llu\n",
1294                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1295                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1296                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1297 }
1298
1299 /*
1300  * This function returns the number of memcg under hierarchy tree. Returns
1301  * 1(self count) if no children.
1302  */
1303 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1304 {
1305         int num = 0;
1306         struct mem_cgroup *iter;
1307
1308         for_each_mem_cgroup_tree(iter, mem)
1309                 num++;
1310         return num;
1311 }
1312
1313 /*
1314  * Return the memory (and swap, if configured) limit for a memcg.
1315  */
1316 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1317 {
1318         u64 limit;
1319         u64 memsw;
1320
1321         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1322         limit += total_swap_pages << PAGE_SHIFT;
1323
1324         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1325         /*
1326          * If memsw is finite and limits the amount of swap space available
1327          * to this memcg, return that limit.
1328          */
1329         return min(limit, memsw);
1330 }
1331
1332 /*
1333  * Visit the first child (need not be the first child as per the ordering
1334  * of the cgroup list, since we track last_scanned_child) of @mem and use
1335  * that to reclaim free pages from.
1336  */
1337 static struct mem_cgroup *
1338 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1339 {
1340         struct mem_cgroup *ret = NULL;
1341         struct cgroup_subsys_state *css;
1342         int nextid, found;
1343
1344         if (!root_mem->use_hierarchy) {
1345                 css_get(&root_mem->css);
1346                 ret = root_mem;
1347         }
1348
1349         while (!ret) {
1350                 rcu_read_lock();
1351                 nextid = root_mem->last_scanned_child + 1;
1352                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1353                                    &found);
1354                 if (css && css_tryget(css))
1355                         ret = container_of(css, struct mem_cgroup, css);
1356
1357                 rcu_read_unlock();
1358                 /* Updates scanning parameter */
1359                 spin_lock(&root_mem->reclaim_param_lock);
1360                 if (!css) {
1361                         /* this means start scan from ID:1 */
1362                         root_mem->last_scanned_child = 0;
1363                 } else
1364                         root_mem->last_scanned_child = found;
1365                 spin_unlock(&root_mem->reclaim_param_lock);
1366         }
1367
1368         return ret;
1369 }
1370
1371 /*
1372  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1373  * we reclaimed from, so that we don't end up penalizing one child extensively
1374  * based on its position in the children list.
1375  *
1376  * root_mem is the original ancestor that we've been reclaim from.
1377  *
1378  * We give up and return to the caller when we visit root_mem twice.
1379  * (other groups can be removed while we're walking....)
1380  *
1381  * If shrink==true, for avoiding to free too much, this returns immedieately.
1382  */
1383 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1384                                                 struct zone *zone,
1385                                                 gfp_t gfp_mask,
1386                                                 unsigned long reclaim_options)
1387 {
1388         struct mem_cgroup *victim;
1389         int ret, total = 0;
1390         int loop = 0;
1391         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1392         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1393         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1394         unsigned long excess;
1395
1396         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1397
1398         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1399         if (root_mem->memsw_is_minimum)
1400                 noswap = true;
1401
1402         while (1) {
1403                 victim = mem_cgroup_select_victim(root_mem);
1404                 if (victim == root_mem) {
1405                         loop++;
1406                         if (loop >= 1)
1407                                 drain_all_stock_async();
1408                         if (loop >= 2) {
1409                                 /*
1410                                  * If we have not been able to reclaim
1411                                  * anything, it might because there are
1412                                  * no reclaimable pages under this hierarchy
1413                                  */
1414                                 if (!check_soft || !total) {
1415                                         css_put(&victim->css);
1416                                         break;
1417                                 }
1418                                 /*
1419                                  * We want to do more targetted reclaim.
1420                                  * excess >> 2 is not to excessive so as to
1421                                  * reclaim too much, nor too less that we keep
1422                                  * coming back to reclaim from this cgroup
1423                                  */
1424                                 if (total >= (excess >> 2) ||
1425                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1426                                         css_put(&victim->css);
1427                                         break;
1428                                 }
1429                         }
1430                 }
1431                 if (!mem_cgroup_local_usage(victim)) {
1432                         /* this cgroup's local usage == 0 */
1433                         css_put(&victim->css);
1434                         continue;
1435                 }
1436                 /* we use swappiness of local cgroup */
1437                 if (check_soft)
1438                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1439                                 noswap, get_swappiness(victim), zone);
1440                 else
1441                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1442                                                 noswap, get_swappiness(victim));
1443                 css_put(&victim->css);
1444                 /*
1445                  * At shrinking usage, we can't check we should stop here or
1446                  * reclaim more. It's depends on callers. last_scanned_child
1447                  * will work enough for keeping fairness under tree.
1448                  */
1449                 if (shrink)
1450                         return ret;
1451                 total += ret;
1452                 if (check_soft) {
1453                         if (!res_counter_soft_limit_excess(&root_mem->res))
1454                                 return total;
1455                 } else if (mem_cgroup_margin(root_mem))
1456                         return 1 + total;
1457         }
1458         return total;
1459 }
1460
1461 /*
1462  * Check OOM-Killer is already running under our hierarchy.
1463  * If someone is running, return false.
1464  */
1465 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1466 {
1467         int x, lock_count = 0;
1468         struct mem_cgroup *iter;
1469
1470         for_each_mem_cgroup_tree(iter, mem) {
1471                 x = atomic_inc_return(&iter->oom_lock);
1472                 lock_count = max(x, lock_count);
1473         }
1474
1475         if (lock_count == 1)
1476                 return true;
1477         return false;
1478 }
1479
1480 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1481 {
1482         struct mem_cgroup *iter;
1483
1484         /*
1485          * When a new child is created while the hierarchy is under oom,
1486          * mem_cgroup_oom_lock() may not be called. We have to use
1487          * atomic_add_unless() here.
1488          */
1489         for_each_mem_cgroup_tree(iter, mem)
1490                 atomic_add_unless(&iter->oom_lock, -1, 0);
1491         return 0;
1492 }
1493
1494
1495 static DEFINE_MUTEX(memcg_oom_mutex);
1496 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1497
1498 struct oom_wait_info {
1499         struct mem_cgroup *mem;
1500         wait_queue_t    wait;
1501 };
1502
1503 static int memcg_oom_wake_function(wait_queue_t *wait,
1504         unsigned mode, int sync, void *arg)
1505 {
1506         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1507         struct oom_wait_info *oom_wait_info;
1508
1509         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1510
1511         if (oom_wait_info->mem == wake_mem)
1512                 goto wakeup;
1513         /* if no hierarchy, no match */
1514         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1515                 return 0;
1516         /*
1517          * Both of oom_wait_info->mem and wake_mem are stable under us.
1518          * Then we can use css_is_ancestor without taking care of RCU.
1519          */
1520         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1521             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1522                 return 0;
1523
1524 wakeup:
1525         return autoremove_wake_function(wait, mode, sync, arg);
1526 }
1527
1528 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1529 {
1530         /* for filtering, pass "mem" as argument. */
1531         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1532 }
1533
1534 static void memcg_oom_recover(struct mem_cgroup *mem)
1535 {
1536         if (mem && atomic_read(&mem->oom_lock))
1537                 memcg_wakeup_oom(mem);
1538 }
1539
1540 /*
1541  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1542  */
1543 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1544 {
1545         struct oom_wait_info owait;
1546         bool locked, need_to_kill;
1547
1548         owait.mem = mem;
1549         owait.wait.flags = 0;
1550         owait.wait.func = memcg_oom_wake_function;
1551         owait.wait.private = current;
1552         INIT_LIST_HEAD(&owait.wait.task_list);
1553         need_to_kill = true;
1554         /* At first, try to OOM lock hierarchy under mem.*/
1555         mutex_lock(&memcg_oom_mutex);
1556         locked = mem_cgroup_oom_lock(mem);
1557         /*
1558          * Even if signal_pending(), we can't quit charge() loop without
1559          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1560          * under OOM is always welcomed, use TASK_KILLABLE here.
1561          */
1562         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1563         if (!locked || mem->oom_kill_disable)
1564                 need_to_kill = false;
1565         if (locked)
1566                 mem_cgroup_oom_notify(mem);
1567         mutex_unlock(&memcg_oom_mutex);
1568
1569         if (need_to_kill) {
1570                 finish_wait(&memcg_oom_waitq, &owait.wait);
1571                 mem_cgroup_out_of_memory(mem, mask);
1572         } else {
1573                 schedule();
1574                 finish_wait(&memcg_oom_waitq, &owait.wait);
1575         }
1576         mutex_lock(&memcg_oom_mutex);
1577         mem_cgroup_oom_unlock(mem);
1578         memcg_wakeup_oom(mem);
1579         mutex_unlock(&memcg_oom_mutex);
1580
1581         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1582                 return false;
1583         /* Give chance to dying process */
1584         schedule_timeout(1);
1585         return true;
1586 }
1587
1588 /*
1589  * Currently used to update mapped file statistics, but the routine can be
1590  * generalized to update other statistics as well.
1591  *
1592  * Notes: Race condition
1593  *
1594  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1595  * it tends to be costly. But considering some conditions, we doesn't need
1596  * to do so _always_.
1597  *
1598  * Considering "charge", lock_page_cgroup() is not required because all
1599  * file-stat operations happen after a page is attached to radix-tree. There
1600  * are no race with "charge".
1601  *
1602  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1603  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1604  * if there are race with "uncharge". Statistics itself is properly handled
1605  * by flags.
1606  *
1607  * Considering "move", this is an only case we see a race. To make the race
1608  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1609  * possibility of race condition. If there is, we take a lock.
1610  */
1611
1612 void mem_cgroup_update_page_stat(struct page *page,
1613                                  enum mem_cgroup_page_stat_item idx, int val)
1614 {
1615         struct mem_cgroup *mem;
1616         struct page_cgroup *pc = lookup_page_cgroup(page);
1617         bool need_unlock = false;
1618         unsigned long uninitialized_var(flags);
1619
1620         if (unlikely(!pc))
1621                 return;
1622
1623         rcu_read_lock();
1624         mem = pc->mem_cgroup;
1625         if (unlikely(!mem || !PageCgroupUsed(pc)))
1626                 goto out;
1627         /* pc->mem_cgroup is unstable ? */
1628         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1629                 /* take a lock against to access pc->mem_cgroup */
1630                 move_lock_page_cgroup(pc, &flags);
1631                 need_unlock = true;
1632                 mem = pc->mem_cgroup;
1633                 if (!mem || !PageCgroupUsed(pc))
1634                         goto out;
1635         }
1636
1637         switch (idx) {
1638         case MEMCG_NR_FILE_MAPPED:
1639                 if (val > 0)
1640                         SetPageCgroupFileMapped(pc);
1641                 else if (!page_mapped(page))
1642                         ClearPageCgroupFileMapped(pc);
1643                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1644                 break;
1645         default:
1646                 BUG();
1647         }
1648
1649         this_cpu_add(mem->stat->count[idx], val);
1650
1651 out:
1652         if (unlikely(need_unlock))
1653                 move_unlock_page_cgroup(pc, &flags);
1654         rcu_read_unlock();
1655         return;
1656 }
1657 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1658
1659 /*
1660  * size of first charge trial. "32" comes from vmscan.c's magic value.
1661  * TODO: maybe necessary to use big numbers in big irons.
1662  */
1663 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1664 struct memcg_stock_pcp {
1665         struct mem_cgroup *cached; /* this never be root cgroup */
1666         int charge;
1667         struct work_struct work;
1668 };
1669 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1670 static atomic_t memcg_drain_count;
1671
1672 /*
1673  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1674  * from local stock and true is returned. If the stock is 0 or charges from a
1675  * cgroup which is not current target, returns false. This stock will be
1676  * refilled.
1677  */
1678 static bool consume_stock(struct mem_cgroup *mem)
1679 {
1680         struct memcg_stock_pcp *stock;
1681         bool ret = true;
1682
1683         stock = &get_cpu_var(memcg_stock);
1684         if (mem == stock->cached && stock->charge)
1685                 stock->charge -= PAGE_SIZE;
1686         else /* need to call res_counter_charge */
1687                 ret = false;
1688         put_cpu_var(memcg_stock);
1689         return ret;
1690 }
1691
1692 /*
1693  * Returns stocks cached in percpu to res_counter and reset cached information.
1694  */
1695 static void drain_stock(struct memcg_stock_pcp *stock)
1696 {
1697         struct mem_cgroup *old = stock->cached;
1698
1699         if (stock->charge) {
1700                 res_counter_uncharge(&old->res, stock->charge);
1701                 if (do_swap_account)
1702                         res_counter_uncharge(&old->memsw, stock->charge);
1703         }
1704         stock->cached = NULL;
1705         stock->charge = 0;
1706 }
1707
1708 /*
1709  * This must be called under preempt disabled or must be called by
1710  * a thread which is pinned to local cpu.
1711  */
1712 static void drain_local_stock(struct work_struct *dummy)
1713 {
1714         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1715         drain_stock(stock);
1716 }
1717
1718 /*
1719  * Cache charges(val) which is from res_counter, to local per_cpu area.
1720  * This will be consumed by consume_stock() function, later.
1721  */
1722 static void refill_stock(struct mem_cgroup *mem, int val)
1723 {
1724         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1725
1726         if (stock->cached != mem) { /* reset if necessary */
1727                 drain_stock(stock);
1728                 stock->cached = mem;
1729         }
1730         stock->charge += val;
1731         put_cpu_var(memcg_stock);
1732 }
1733
1734 /*
1735  * Tries to drain stocked charges in other cpus. This function is asynchronous
1736  * and just put a work per cpu for draining localy on each cpu. Caller can
1737  * expects some charges will be back to res_counter later but cannot wait for
1738  * it.
1739  */
1740 static void drain_all_stock_async(void)
1741 {
1742         int cpu;
1743         /* This function is for scheduling "drain" in asynchronous way.
1744          * The result of "drain" is not directly handled by callers. Then,
1745          * if someone is calling drain, we don't have to call drain more.
1746          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1747          * there is a race. We just do loose check here.
1748          */
1749         if (atomic_read(&memcg_drain_count))
1750                 return;
1751         /* Notify other cpus that system-wide "drain" is running */
1752         atomic_inc(&memcg_drain_count);
1753         get_online_cpus();
1754         for_each_online_cpu(cpu) {
1755                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1756                 schedule_work_on(cpu, &stock->work);
1757         }
1758         put_online_cpus();
1759         atomic_dec(&memcg_drain_count);
1760         /* We don't wait for flush_work */
1761 }
1762
1763 /* This is a synchronous drain interface. */
1764 static void drain_all_stock_sync(void)
1765 {
1766         /* called when force_empty is called */
1767         atomic_inc(&memcg_drain_count);
1768         schedule_on_each_cpu(drain_local_stock);
1769         atomic_dec(&memcg_drain_count);
1770 }
1771
1772 /*
1773  * This function drains percpu counter value from DEAD cpu and
1774  * move it to local cpu. Note that this function can be preempted.
1775  */
1776 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1777 {
1778         int i;
1779
1780         spin_lock(&mem->pcp_counter_lock);
1781         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1782                 s64 x = per_cpu(mem->stat->count[i], cpu);
1783
1784                 per_cpu(mem->stat->count[i], cpu) = 0;
1785                 mem->nocpu_base.count[i] += x;
1786         }
1787         /* need to clear ON_MOVE value, works as a kind of lock. */
1788         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1789         spin_unlock(&mem->pcp_counter_lock);
1790 }
1791
1792 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1793 {
1794         int idx = MEM_CGROUP_ON_MOVE;
1795
1796         spin_lock(&mem->pcp_counter_lock);
1797         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1798         spin_unlock(&mem->pcp_counter_lock);
1799 }
1800
1801 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1802                                         unsigned long action,
1803                                         void *hcpu)
1804 {
1805         int cpu = (unsigned long)hcpu;
1806         struct memcg_stock_pcp *stock;
1807         struct mem_cgroup *iter;
1808
1809         if ((action == CPU_ONLINE)) {
1810                 for_each_mem_cgroup_all(iter)
1811                         synchronize_mem_cgroup_on_move(iter, cpu);
1812                 return NOTIFY_OK;
1813         }
1814
1815         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1816                 return NOTIFY_OK;
1817
1818         for_each_mem_cgroup_all(iter)
1819                 mem_cgroup_drain_pcp_counter(iter, cpu);
1820
1821         stock = &per_cpu(memcg_stock, cpu);
1822         drain_stock(stock);
1823         return NOTIFY_OK;
1824 }
1825
1826
1827 /* See __mem_cgroup_try_charge() for details */
1828 enum {
1829         CHARGE_OK,              /* success */
1830         CHARGE_RETRY,           /* need to retry but retry is not bad */
1831         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1832         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1833         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1834 };
1835
1836 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1837                                 int csize, bool oom_check)
1838 {
1839         struct mem_cgroup *mem_over_limit;
1840         struct res_counter *fail_res;
1841         unsigned long flags = 0;
1842         int ret;
1843
1844         ret = res_counter_charge(&mem->res, csize, &fail_res);
1845
1846         if (likely(!ret)) {
1847                 if (!do_swap_account)
1848                         return CHARGE_OK;
1849                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1850                 if (likely(!ret))
1851                         return CHARGE_OK;
1852
1853                 res_counter_uncharge(&mem->res, csize);
1854                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1855                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1856         } else
1857                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1858         /*
1859          * csize can be either a huge page (HPAGE_SIZE), a batch of
1860          * regular pages (CHARGE_SIZE), or a single regular page
1861          * (PAGE_SIZE).
1862          *
1863          * Never reclaim on behalf of optional batching, retry with a
1864          * single page instead.
1865          */
1866         if (csize == CHARGE_SIZE)
1867                 return CHARGE_RETRY;
1868
1869         if (!(gfp_mask & __GFP_WAIT))
1870                 return CHARGE_WOULDBLOCK;
1871
1872         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1873                                               gfp_mask, flags);
1874         if (mem_cgroup_margin(mem_over_limit) >= csize)
1875                 return CHARGE_RETRY;
1876         /*
1877          * Even though the limit is exceeded at this point, reclaim
1878          * may have been able to free some pages.  Retry the charge
1879          * before killing the task.
1880          *
1881          * Only for regular pages, though: huge pages are rather
1882          * unlikely to succeed so close to the limit, and we fall back
1883          * to regular pages anyway in case of failure.
1884          */
1885         if (csize == PAGE_SIZE && ret)
1886                 return CHARGE_RETRY;
1887
1888         /*
1889          * At task move, charge accounts can be doubly counted. So, it's
1890          * better to wait until the end of task_move if something is going on.
1891          */
1892         if (mem_cgroup_wait_acct_move(mem_over_limit))
1893                 return CHARGE_RETRY;
1894
1895         /* If we don't need to call oom-killer at el, return immediately */
1896         if (!oom_check)
1897                 return CHARGE_NOMEM;
1898         /* check OOM */
1899         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1900                 return CHARGE_OOM_DIE;
1901
1902         return CHARGE_RETRY;
1903 }
1904
1905 /*
1906  * Unlike exported interface, "oom" parameter is added. if oom==true,
1907  * oom-killer can be invoked.
1908  */
1909 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1910                                    gfp_t gfp_mask,
1911                                    struct mem_cgroup **memcg, bool oom,
1912                                    int page_size)
1913 {
1914         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1915         struct mem_cgroup *mem = NULL;
1916         int ret;
1917         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1918
1919         /*
1920          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1921          * in system level. So, allow to go ahead dying process in addition to
1922          * MEMDIE process.
1923          */
1924         if (unlikely(test_thread_flag(TIF_MEMDIE)
1925                      || fatal_signal_pending(current)))
1926                 goto bypass;
1927
1928         /*
1929          * We always charge the cgroup the mm_struct belongs to.
1930          * The mm_struct's mem_cgroup changes on task migration if the
1931          * thread group leader migrates. It's possible that mm is not
1932          * set, if so charge the init_mm (happens for pagecache usage).
1933          */
1934         if (!*memcg && !mm)
1935                 goto bypass;
1936 again:
1937         if (*memcg) { /* css should be a valid one */
1938                 mem = *memcg;
1939                 VM_BUG_ON(css_is_removed(&mem->css));
1940                 if (mem_cgroup_is_root(mem))
1941                         goto done;
1942                 if (page_size == PAGE_SIZE && consume_stock(mem))
1943                         goto done;
1944                 css_get(&mem->css);
1945         } else {
1946                 struct task_struct *p;
1947
1948                 rcu_read_lock();
1949                 p = rcu_dereference(mm->owner);
1950                 /*
1951                  * Because we don't have task_lock(), "p" can exit.
1952                  * In that case, "mem" can point to root or p can be NULL with
1953                  * race with swapoff. Then, we have small risk of mis-accouning.
1954                  * But such kind of mis-account by race always happens because
1955                  * we don't have cgroup_mutex(). It's overkill and we allo that
1956                  * small race, here.
1957                  * (*) swapoff at el will charge against mm-struct not against
1958                  * task-struct. So, mm->owner can be NULL.
1959                  */
1960                 mem = mem_cgroup_from_task(p);
1961                 if (!mem || mem_cgroup_is_root(mem)) {
1962                         rcu_read_unlock();
1963                         goto done;
1964                 }
1965                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1966                         /*
1967                          * It seems dagerous to access memcg without css_get().
1968                          * But considering how consume_stok works, it's not
1969                          * necessary. If consume_stock success, some charges
1970                          * from this memcg are cached on this cpu. So, we
1971                          * don't need to call css_get()/css_tryget() before
1972                          * calling consume_stock().
1973                          */
1974                         rcu_read_unlock();
1975                         goto done;
1976                 }
1977                 /* after here, we may be blocked. we need to get refcnt */
1978                 if (!css_tryget(&mem->css)) {
1979                         rcu_read_unlock();
1980                         goto again;
1981                 }
1982                 rcu_read_unlock();
1983         }
1984
1985         do {
1986                 bool oom_check;
1987
1988                 /* If killed, bypass charge */
1989                 if (fatal_signal_pending(current)) {
1990                         css_put(&mem->css);
1991                         goto bypass;
1992                 }
1993
1994                 oom_check = false;
1995                 if (oom && !nr_oom_retries) {
1996                         oom_check = true;
1997                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1998                 }
1999
2000                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
2001
2002                 switch (ret) {
2003                 case CHARGE_OK:
2004                         break;
2005                 case CHARGE_RETRY: /* not in OOM situation but retry */
2006                         csize = page_size;
2007                         css_put(&mem->css);
2008                         mem = NULL;
2009                         goto again;
2010                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2011                         css_put(&mem->css);
2012                         goto nomem;
2013                 case CHARGE_NOMEM: /* OOM routine works */
2014                         if (!oom) {
2015                                 css_put(&mem->css);
2016                                 goto nomem;
2017                         }
2018                         /* If oom, we never return -ENOMEM */
2019                         nr_oom_retries--;
2020                         break;
2021                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2022                         css_put(&mem->css);
2023                         goto bypass;
2024                 }
2025         } while (ret != CHARGE_OK);
2026
2027         if (csize > page_size)
2028                 refill_stock(mem, csize - page_size);
2029         css_put(&mem->css);
2030 done:
2031         *memcg = mem;
2032         return 0;
2033 nomem:
2034         *memcg = NULL;
2035         return -ENOMEM;
2036 bypass:
2037         *memcg = NULL;
2038         return 0;
2039 }
2040
2041 /*
2042  * Somemtimes we have to undo a charge we got by try_charge().
2043  * This function is for that and do uncharge, put css's refcnt.
2044  * gotten by try_charge().
2045  */
2046 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2047                                                         unsigned long count)
2048 {
2049         if (!mem_cgroup_is_root(mem)) {
2050                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2051                 if (do_swap_account)
2052                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2053         }
2054 }
2055
2056 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2057                                      int page_size)
2058 {
2059         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2060 }
2061
2062 /*
2063  * A helper function to get mem_cgroup from ID. must be called under
2064  * rcu_read_lock(). The caller must check css_is_removed() or some if
2065  * it's concern. (dropping refcnt from swap can be called against removed
2066  * memcg.)
2067  */
2068 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2069 {
2070         struct cgroup_subsys_state *css;
2071
2072         /* ID 0 is unused ID */
2073         if (!id)
2074                 return NULL;
2075         css = css_lookup(&mem_cgroup_subsys, id);
2076         if (!css)
2077                 return NULL;
2078         return container_of(css, struct mem_cgroup, css);
2079 }
2080
2081 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2082 {
2083         struct mem_cgroup *mem = NULL;
2084         struct page_cgroup *pc;
2085         unsigned short id;
2086         swp_entry_t ent;
2087
2088         VM_BUG_ON(!PageLocked(page));
2089
2090         pc = lookup_page_cgroup(page);
2091         lock_page_cgroup(pc);
2092         if (PageCgroupUsed(pc)) {
2093                 mem = pc->mem_cgroup;
2094                 if (mem && !css_tryget(&mem->css))
2095                         mem = NULL;
2096         } else if (PageSwapCache(page)) {
2097                 ent.val = page_private(page);
2098                 id = lookup_swap_cgroup(ent);
2099                 rcu_read_lock();
2100                 mem = mem_cgroup_lookup(id);
2101                 if (mem && !css_tryget(&mem->css))
2102                         mem = NULL;
2103                 rcu_read_unlock();
2104         }
2105         unlock_page_cgroup(pc);
2106         return mem;
2107 }
2108
2109 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2110                                        struct page *page,
2111                                        struct page_cgroup *pc,
2112                                        enum charge_type ctype,
2113                                        int page_size)
2114 {
2115         int nr_pages = page_size >> PAGE_SHIFT;
2116
2117         lock_page_cgroup(pc);
2118         if (unlikely(PageCgroupUsed(pc))) {
2119                 unlock_page_cgroup(pc);
2120                 mem_cgroup_cancel_charge(mem, page_size);
2121                 return;
2122         }
2123         /*
2124          * we don't need page_cgroup_lock about tail pages, becase they are not
2125          * accessed by any other context at this point.
2126          */
2127         pc->mem_cgroup = mem;
2128         /*
2129          * We access a page_cgroup asynchronously without lock_page_cgroup().
2130          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2131          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2132          * before USED bit, we need memory barrier here.
2133          * See mem_cgroup_add_lru_list(), etc.
2134          */
2135         smp_wmb();
2136         switch (ctype) {
2137         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2138         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2139                 SetPageCgroupCache(pc);
2140                 SetPageCgroupUsed(pc);
2141                 break;
2142         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2143                 ClearPageCgroupCache(pc);
2144                 SetPageCgroupUsed(pc);
2145                 break;
2146         default:
2147                 break;
2148         }
2149
2150         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2151         unlock_page_cgroup(pc);
2152         /*
2153          * "charge_statistics" updated event counter. Then, check it.
2154          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2155          * if they exceeds softlimit.
2156          */
2157         memcg_check_events(mem, page);
2158 }
2159
2160 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2161
2162 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2163                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2164 /*
2165  * Because tail pages are not marked as "used", set it. We're under
2166  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2167  */
2168 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2169 {
2170         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2171         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2172         unsigned long flags;
2173
2174         if (mem_cgroup_disabled())
2175                 return;
2176         /*
2177          * We have no races with charge/uncharge but will have races with
2178          * page state accounting.
2179          */
2180         move_lock_page_cgroup(head_pc, &flags);
2181
2182         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2183         smp_wmb(); /* see __commit_charge() */
2184         if (PageCgroupAcctLRU(head_pc)) {
2185                 enum lru_list lru;
2186                 struct mem_cgroup_per_zone *mz;
2187
2188                 /*
2189                  * LRU flags cannot be copied because we need to add tail
2190                  *.page to LRU by generic call and our hook will be called.
2191                  * We hold lru_lock, then, reduce counter directly.
2192                  */
2193                 lru = page_lru(head);
2194                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2195                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2196         }
2197         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2198         move_unlock_page_cgroup(head_pc, &flags);
2199 }
2200 #endif
2201
2202 /**
2203  * mem_cgroup_move_account - move account of the page
2204  * @page: the page
2205  * @pc: page_cgroup of the page.
2206  * @from: mem_cgroup which the page is moved from.
2207  * @to: mem_cgroup which the page is moved to. @from != @to.
2208  * @uncharge: whether we should call uncharge and css_put against @from.
2209  * @charge_size: number of bytes to charge (regular or huge page)
2210  *
2211  * The caller must confirm following.
2212  * - page is not on LRU (isolate_page() is useful.)
2213  * - compound_lock is held when charge_size > PAGE_SIZE
2214  *
2215  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2216  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2217  * true, this function does "uncharge" from old cgroup, but it doesn't if
2218  * @uncharge is false, so a caller should do "uncharge".
2219  */
2220 static int mem_cgroup_move_account(struct page *page, struct page_cgroup *pc,
2221                                    struct mem_cgroup *from, struct mem_cgroup *to,
2222                                    bool uncharge, int charge_size)
2223 {
2224         int nr_pages = charge_size >> PAGE_SHIFT;
2225         unsigned long flags;
2226         int ret;
2227
2228         VM_BUG_ON(from == to);
2229         VM_BUG_ON(PageLRU(page));
2230         /*
2231          * The page is isolated from LRU. So, collapse function
2232          * will not handle this page. But page splitting can happen.
2233          * Do this check under compound_page_lock(). The caller should
2234          * hold it.
2235          */
2236         ret = -EBUSY;
2237         if (charge_size > PAGE_SIZE && !PageTransHuge(page))
2238                 goto out;
2239
2240         lock_page_cgroup(pc);
2241
2242         ret = -EINVAL;
2243         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2244                 goto unlock;
2245
2246         move_lock_page_cgroup(pc, &flags);
2247
2248         if (PageCgroupFileMapped(pc)) {
2249                 /* Update mapped_file data for mem_cgroup */
2250                 preempt_disable();
2251                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2252                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2253                 preempt_enable();
2254         }
2255         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2256         if (uncharge)
2257                 /* This is not "cancel", but cancel_charge does all we need. */
2258                 mem_cgroup_cancel_charge(from, charge_size);
2259
2260         /* caller should have done css_get */
2261         pc->mem_cgroup = to;
2262         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2263         /*
2264          * We charges against "to" which may not have any tasks. Then, "to"
2265          * can be under rmdir(). But in current implementation, caller of
2266          * this function is just force_empty() and move charge, so it's
2267          * garanteed that "to" is never removed. So, we don't check rmdir
2268          * status here.
2269          */
2270         move_unlock_page_cgroup(pc, &flags);
2271         ret = 0;
2272 unlock:
2273         unlock_page_cgroup(pc);
2274         /*
2275          * check events
2276          */
2277         memcg_check_events(to, page);
2278         memcg_check_events(from, page);
2279 out:
2280         return ret;
2281 }
2282
2283 /*
2284  * move charges to its parent.
2285  */
2286
2287 static int mem_cgroup_move_parent(struct page *page,
2288                                   struct page_cgroup *pc,
2289                                   struct mem_cgroup *child,
2290                                   gfp_t gfp_mask)
2291 {
2292         struct cgroup *cg = child->css.cgroup;
2293         struct cgroup *pcg = cg->parent;
2294         struct mem_cgroup *parent;
2295         int page_size = PAGE_SIZE;
2296         unsigned long flags;
2297         int ret;
2298
2299         /* Is ROOT ? */
2300         if (!pcg)
2301                 return -EINVAL;
2302
2303         ret = -EBUSY;
2304         if (!get_page_unless_zero(page))
2305                 goto out;
2306         if (isolate_lru_page(page))
2307                 goto put;
2308
2309         if (PageTransHuge(page))
2310                 page_size = HPAGE_SIZE;
2311
2312         parent = mem_cgroup_from_cont(pcg);
2313         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2314                                 &parent, false, page_size);
2315         if (ret || !parent)
2316                 goto put_back;
2317
2318         if (page_size > PAGE_SIZE)
2319                 flags = compound_lock_irqsave(page);
2320
2321         ret = mem_cgroup_move_account(page, pc, child, parent, true, page_size);
2322         if (ret)
2323                 mem_cgroup_cancel_charge(parent, page_size);
2324
2325         if (page_size > PAGE_SIZE)
2326                 compound_unlock_irqrestore(page, flags);
2327 put_back:
2328         putback_lru_page(page);
2329 put:
2330         put_page(page);
2331 out:
2332         return ret;
2333 }
2334
2335 /*
2336  * Charge the memory controller for page usage.
2337  * Return
2338  * 0 if the charge was successful
2339  * < 0 if the cgroup is over its limit
2340  */
2341 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2342                                 gfp_t gfp_mask, enum charge_type ctype)
2343 {
2344         struct mem_cgroup *mem = NULL;
2345         int page_size = PAGE_SIZE;
2346         struct page_cgroup *pc;
2347         bool oom = true;
2348         int ret;
2349
2350         if (PageTransHuge(page)) {
2351                 page_size <<= compound_order(page);
2352                 VM_BUG_ON(!PageTransHuge(page));
2353                 /*
2354                  * Never OOM-kill a process for a huge page.  The
2355                  * fault handler will fall back to regular pages.
2356                  */
2357                 oom = false;
2358         }
2359
2360         pc = lookup_page_cgroup(page);
2361         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2362
2363         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2364         if (ret || !mem)
2365                 return ret;
2366
2367         __mem_cgroup_commit_charge(mem, page, pc, ctype, page_size);
2368         return 0;
2369 }
2370
2371 int mem_cgroup_newpage_charge(struct page *page,
2372                               struct mm_struct *mm, gfp_t gfp_mask)
2373 {
2374         if (mem_cgroup_disabled())
2375                 return 0;
2376         /*
2377          * If already mapped, we don't have to account.
2378          * If page cache, page->mapping has address_space.
2379          * But page->mapping may have out-of-use anon_vma pointer,
2380          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2381          * is NULL.
2382          */
2383         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2384                 return 0;
2385         if (unlikely(!mm))
2386                 mm = &init_mm;
2387         return mem_cgroup_charge_common(page, mm, gfp_mask,
2388                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2389 }
2390
2391 static void
2392 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2393                                         enum charge_type ctype);
2394
2395 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2396                                 gfp_t gfp_mask)
2397 {
2398         int ret;
2399
2400         if (mem_cgroup_disabled())
2401                 return 0;
2402         if (PageCompound(page))
2403                 return 0;
2404         /*
2405          * Corner case handling. This is called from add_to_page_cache()
2406          * in usual. But some FS (shmem) precharges this page before calling it
2407          * and call add_to_page_cache() with GFP_NOWAIT.
2408          *
2409          * For GFP_NOWAIT case, the page may be pre-charged before calling
2410          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2411          * charge twice. (It works but has to pay a bit larger cost.)
2412          * And when the page is SwapCache, it should take swap information
2413          * into account. This is under lock_page() now.
2414          */
2415         if (!(gfp_mask & __GFP_WAIT)) {
2416                 struct page_cgroup *pc;
2417
2418                 pc = lookup_page_cgroup(page);
2419                 if (!pc)
2420                         return 0;
2421                 lock_page_cgroup(pc);
2422                 if (PageCgroupUsed(pc)) {
2423                         unlock_page_cgroup(pc);
2424                         return 0;
2425                 }
2426                 unlock_page_cgroup(pc);
2427         }
2428
2429         if (unlikely(!mm))
2430                 mm = &init_mm;
2431
2432         if (page_is_file_cache(page))
2433                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2434                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2435
2436         /* shmem */
2437         if (PageSwapCache(page)) {
2438                 struct mem_cgroup *mem;
2439
2440                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2441                 if (!ret)
2442                         __mem_cgroup_commit_charge_swapin(page, mem,
2443                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2444         } else
2445                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2446                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2447
2448         return ret;
2449 }
2450
2451 /*
2452  * While swap-in, try_charge -> commit or cancel, the page is locked.
2453  * And when try_charge() successfully returns, one refcnt to memcg without
2454  * struct page_cgroup is acquired. This refcnt will be consumed by
2455  * "commit()" or removed by "cancel()"
2456  */
2457 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2458                                  struct page *page,
2459                                  gfp_t mask, struct mem_cgroup **ptr)
2460 {
2461         struct mem_cgroup *mem;
2462         int ret;
2463
2464         *ptr = NULL;
2465
2466         if (mem_cgroup_disabled())
2467                 return 0;
2468
2469         if (!do_swap_account)
2470                 goto charge_cur_mm;
2471         /*
2472          * A racing thread's fault, or swapoff, may have already updated
2473          * the pte, and even removed page from swap cache: in those cases
2474          * do_swap_page()'s pte_same() test will fail; but there's also a
2475          * KSM case which does need to charge the page.
2476          */
2477         if (!PageSwapCache(page))
2478                 goto charge_cur_mm;
2479         mem = try_get_mem_cgroup_from_page(page);
2480         if (!mem)
2481                 goto charge_cur_mm;
2482         *ptr = mem;
2483         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2484         css_put(&mem->css);
2485         return ret;
2486 charge_cur_mm:
2487         if (unlikely(!mm))
2488                 mm = &init_mm;
2489         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2490 }
2491
2492 static void
2493 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2494                                         enum charge_type ctype)
2495 {
2496         struct page_cgroup *pc;
2497
2498         if (mem_cgroup_disabled())
2499                 return;
2500         if (!ptr)
2501                 return;
2502         cgroup_exclude_rmdir(&ptr->css);
2503         pc = lookup_page_cgroup(page);
2504         mem_cgroup_lru_del_before_commit_swapcache(page);
2505         __mem_cgroup_commit_charge(ptr, page, pc, ctype, PAGE_SIZE);
2506         mem_cgroup_lru_add_after_commit_swapcache(page);
2507         /*
2508          * Now swap is on-memory. This means this page may be
2509          * counted both as mem and swap....double count.
2510          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2511          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2512          * may call delete_from_swap_cache() before reach here.
2513          */
2514         if (do_swap_account && PageSwapCache(page)) {
2515                 swp_entry_t ent = {.val = page_private(page)};
2516                 unsigned short id;
2517                 struct mem_cgroup *memcg;
2518
2519                 id = swap_cgroup_record(ent, 0);
2520                 rcu_read_lock();
2521                 memcg = mem_cgroup_lookup(id);
2522                 if (memcg) {
2523                         /*
2524                          * This recorded memcg can be obsolete one. So, avoid
2525                          * calling css_tryget
2526                          */
2527                         if (!mem_cgroup_is_root(memcg))
2528                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2529                         mem_cgroup_swap_statistics(memcg, false);
2530                         mem_cgroup_put(memcg);
2531                 }
2532                 rcu_read_unlock();
2533         }
2534         /*
2535          * At swapin, we may charge account against cgroup which has no tasks.
2536          * So, rmdir()->pre_destroy() can be called while we do this charge.
2537          * In that case, we need to call pre_destroy() again. check it here.
2538          */
2539         cgroup_release_and_wakeup_rmdir(&ptr->css);
2540 }
2541
2542 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2543 {
2544         __mem_cgroup_commit_charge_swapin(page, ptr,
2545                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2546 }
2547
2548 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2549 {
2550         if (mem_cgroup_disabled())
2551                 return;
2552         if (!mem)
2553                 return;
2554         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2555 }
2556
2557 static void
2558 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2559               int page_size)
2560 {
2561         struct memcg_batch_info *batch = NULL;
2562         bool uncharge_memsw = true;
2563         /* If swapout, usage of swap doesn't decrease */
2564         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2565                 uncharge_memsw = false;
2566
2567         batch = &current->memcg_batch;
2568         /*
2569          * In usual, we do css_get() when we remember memcg pointer.
2570          * But in this case, we keep res->usage until end of a series of
2571          * uncharges. Then, it's ok to ignore memcg's refcnt.
2572          */
2573         if (!batch->memcg)
2574                 batch->memcg = mem;
2575         /*
2576          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2577          * In those cases, all pages freed continously can be expected to be in
2578          * the same cgroup and we have chance to coalesce uncharges.
2579          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2580          * because we want to do uncharge as soon as possible.
2581          */
2582
2583         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2584                 goto direct_uncharge;
2585
2586         if (page_size != PAGE_SIZE)
2587                 goto direct_uncharge;
2588
2589         /*
2590          * In typical case, batch->memcg == mem. This means we can
2591          * merge a series of uncharges to an uncharge of res_counter.
2592          * If not, we uncharge res_counter ony by one.
2593          */
2594         if (batch->memcg != mem)
2595                 goto direct_uncharge;
2596         /* remember freed charge and uncharge it later */
2597         batch->bytes += PAGE_SIZE;
2598         if (uncharge_memsw)
2599                 batch->memsw_bytes += PAGE_SIZE;
2600         return;
2601 direct_uncharge:
2602         res_counter_uncharge(&mem->res, page_size);
2603         if (uncharge_memsw)
2604                 res_counter_uncharge(&mem->memsw, page_size);
2605         if (unlikely(batch->memcg != mem))
2606                 memcg_oom_recover(mem);
2607         return;
2608 }
2609
2610 /*
2611  * uncharge if !page_mapped(page)
2612  */
2613 static struct mem_cgroup *
2614 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2615 {
2616         int count;
2617         struct page_cgroup *pc;
2618         struct mem_cgroup *mem = NULL;
2619         int page_size = PAGE_SIZE;
2620
2621         if (mem_cgroup_disabled())
2622                 return NULL;
2623
2624         if (PageSwapCache(page))
2625                 return NULL;
2626
2627         if (PageTransHuge(page)) {
2628                 page_size <<= compound_order(page);
2629                 VM_BUG_ON(!PageTransHuge(page));
2630         }
2631
2632         count = page_size >> PAGE_SHIFT;
2633         /*
2634          * Check if our page_cgroup is valid
2635          */
2636         pc = lookup_page_cgroup(page);
2637         if (unlikely(!pc || !PageCgroupUsed(pc)))
2638                 return NULL;
2639
2640         lock_page_cgroup(pc);
2641
2642         mem = pc->mem_cgroup;
2643
2644         if (!PageCgroupUsed(pc))
2645                 goto unlock_out;
2646
2647         switch (ctype) {
2648         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2649         case MEM_CGROUP_CHARGE_TYPE_DROP:
2650                 /* See mem_cgroup_prepare_migration() */
2651                 if (page_mapped(page) || PageCgroupMigration(pc))
2652                         goto unlock_out;
2653                 break;
2654         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2655                 if (!PageAnon(page)) {  /* Shared memory */
2656                         if (page->mapping && !page_is_file_cache(page))
2657                                 goto unlock_out;
2658                 } else if (page_mapped(page)) /* Anon */
2659                                 goto unlock_out;
2660                 break;
2661         default:
2662                 break;
2663         }
2664
2665         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2666
2667         ClearPageCgroupUsed(pc);
2668         /*
2669          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2670          * freed from LRU. This is safe because uncharged page is expected not
2671          * to be reused (freed soon). Exception is SwapCache, it's handled by
2672          * special functions.
2673          */
2674
2675         unlock_page_cgroup(pc);
2676         /*
2677          * even after unlock, we have mem->res.usage here and this memcg
2678          * will never be freed.
2679          */
2680         memcg_check_events(mem, page);
2681         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2682                 mem_cgroup_swap_statistics(mem, true);
2683                 mem_cgroup_get(mem);
2684         }
2685         if (!mem_cgroup_is_root(mem))
2686                 __do_uncharge(mem, ctype, page_size);
2687
2688         return mem;
2689
2690 unlock_out:
2691         unlock_page_cgroup(pc);
2692         return NULL;
2693 }
2694
2695 void mem_cgroup_uncharge_page(struct page *page)
2696 {
2697         /* early check. */
2698         if (page_mapped(page))
2699                 return;
2700         if (page->mapping && !PageAnon(page))
2701                 return;
2702         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2703 }
2704
2705 void mem_cgroup_uncharge_cache_page(struct page *page)
2706 {
2707         VM_BUG_ON(page_mapped(page));
2708         VM_BUG_ON(page->mapping);
2709         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2710 }
2711
2712 /*
2713  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2714  * In that cases, pages are freed continuously and we can expect pages
2715  * are in the same memcg. All these calls itself limits the number of
2716  * pages freed at once, then uncharge_start/end() is called properly.
2717  * This may be called prural(2) times in a context,
2718  */
2719
2720 void mem_cgroup_uncharge_start(void)
2721 {
2722         current->memcg_batch.do_batch++;
2723         /* We can do nest. */
2724         if (current->memcg_batch.do_batch == 1) {
2725                 current->memcg_batch.memcg = NULL;
2726                 current->memcg_batch.bytes = 0;
2727                 current->memcg_batch.memsw_bytes = 0;
2728         }
2729 }
2730
2731 void mem_cgroup_uncharge_end(void)
2732 {
2733         struct memcg_batch_info *batch = &current->memcg_batch;
2734
2735         if (!batch->do_batch)
2736                 return;
2737
2738         batch->do_batch--;
2739         if (batch->do_batch) /* If stacked, do nothing. */
2740                 return;
2741
2742         if (!batch->memcg)
2743                 return;
2744         /*
2745          * This "batch->memcg" is valid without any css_get/put etc...
2746          * bacause we hide charges behind us.
2747          */
2748         if (batch->bytes)
2749                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2750         if (batch->memsw_bytes)
2751                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2752         memcg_oom_recover(batch->memcg);
2753         /* forget this pointer (for sanity check) */
2754         batch->memcg = NULL;
2755 }
2756
2757 #ifdef CONFIG_SWAP
2758 /*
2759  * called after __delete_from_swap_cache() and drop "page" account.
2760  * memcg information is recorded to swap_cgroup of "ent"
2761  */
2762 void
2763 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2764 {
2765         struct mem_cgroup *memcg;
2766         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2767
2768         if (!swapout) /* this was a swap cache but the swap is unused ! */
2769                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2770
2771         memcg = __mem_cgroup_uncharge_common(page, ctype);
2772
2773         /*
2774          * record memcg information,  if swapout && memcg != NULL,
2775          * mem_cgroup_get() was called in uncharge().
2776          */
2777         if (do_swap_account && swapout && memcg)
2778                 swap_cgroup_record(ent, css_id(&memcg->css));
2779 }
2780 #endif
2781
2782 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2783 /*
2784  * called from swap_entry_free(). remove record in swap_cgroup and
2785  * uncharge "memsw" account.
2786  */
2787 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2788 {
2789         struct mem_cgroup *memcg;
2790         unsigned short id;
2791
2792         if (!do_swap_account)
2793                 return;
2794
2795         id = swap_cgroup_record(ent, 0);
2796         rcu_read_lock();
2797         memcg = mem_cgroup_lookup(id);
2798         if (memcg) {
2799                 /*
2800                  * We uncharge this because swap is freed.
2801                  * This memcg can be obsolete one. We avoid calling css_tryget
2802                  */
2803                 if (!mem_cgroup_is_root(memcg))
2804                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2805                 mem_cgroup_swap_statistics(memcg, false);
2806                 mem_cgroup_put(memcg);
2807         }
2808         rcu_read_unlock();
2809 }
2810
2811 /**
2812  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2813  * @entry: swap entry to be moved
2814  * @from:  mem_cgroup which the entry is moved from
2815  * @to:  mem_cgroup which the entry is moved to
2816  * @need_fixup: whether we should fixup res_counters and refcounts.
2817  *
2818  * It succeeds only when the swap_cgroup's record for this entry is the same
2819  * as the mem_cgroup's id of @from.
2820  *
2821  * Returns 0 on success, -EINVAL on failure.
2822  *
2823  * The caller must have charged to @to, IOW, called res_counter_charge() about
2824  * both res and memsw, and called css_get().
2825  */
2826 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2827                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2828 {
2829         unsigned short old_id, new_id;
2830
2831         old_id = css_id(&from->css);
2832         new_id = css_id(&to->css);
2833
2834         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2835                 mem_cgroup_swap_statistics(from, false);
2836                 mem_cgroup_swap_statistics(to, true);
2837                 /*
2838                  * This function is only called from task migration context now.
2839                  * It postpones res_counter and refcount handling till the end
2840                  * of task migration(mem_cgroup_clear_mc()) for performance
2841                  * improvement. But we cannot postpone mem_cgroup_get(to)
2842                  * because if the process that has been moved to @to does
2843                  * swap-in, the refcount of @to might be decreased to 0.
2844                  */
2845                 mem_cgroup_get(to);
2846                 if (need_fixup) {
2847                         if (!mem_cgroup_is_root(from))
2848                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2849                         mem_cgroup_put(from);
2850                         /*
2851                          * we charged both to->res and to->memsw, so we should
2852                          * uncharge to->res.
2853                          */
2854                         if (!mem_cgroup_is_root(to))
2855                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2856                 }
2857                 return 0;
2858         }
2859         return -EINVAL;
2860 }
2861 #else
2862 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2863                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2864 {
2865         return -EINVAL;
2866 }
2867 #endif
2868
2869 /*
2870  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2871  * page belongs to.
2872  */
2873 int mem_cgroup_prepare_migration(struct page *page,
2874         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2875 {
2876         struct page_cgroup *pc;
2877         struct mem_cgroup *mem = NULL;
2878         enum charge_type ctype;
2879         int ret = 0;
2880
2881         *ptr = NULL;
2882
2883         VM_BUG_ON(PageTransHuge(page));
2884         if (mem_cgroup_disabled())
2885                 return 0;
2886
2887         pc = lookup_page_cgroup(page);
2888         lock_page_cgroup(pc);
2889         if (PageCgroupUsed(pc)) {
2890                 mem = pc->mem_cgroup;
2891                 css_get(&mem->css);
2892                 /*
2893                  * At migrating an anonymous page, its mapcount goes down
2894                  * to 0 and uncharge() will be called. But, even if it's fully
2895                  * unmapped, migration may fail and this page has to be
2896                  * charged again. We set MIGRATION flag here and delay uncharge
2897                  * until end_migration() is called
2898                  *
2899                  * Corner Case Thinking
2900                  * A)
2901                  * When the old page was mapped as Anon and it's unmap-and-freed
2902                  * while migration was ongoing.
2903                  * If unmap finds the old page, uncharge() of it will be delayed
2904                  * until end_migration(). If unmap finds a new page, it's
2905                  * uncharged when it make mapcount to be 1->0. If unmap code
2906                  * finds swap_migration_entry, the new page will not be mapped
2907                  * and end_migration() will find it(mapcount==0).
2908                  *
2909                  * B)
2910                  * When the old page was mapped but migraion fails, the kernel
2911                  * remaps it. A charge for it is kept by MIGRATION flag even
2912                  * if mapcount goes down to 0. We can do remap successfully
2913                  * without charging it again.
2914                  *
2915                  * C)
2916                  * The "old" page is under lock_page() until the end of
2917                  * migration, so, the old page itself will not be swapped-out.
2918                  * If the new page is swapped out before end_migraton, our
2919                  * hook to usual swap-out path will catch the event.
2920                  */
2921                 if (PageAnon(page))
2922                         SetPageCgroupMigration(pc);
2923         }
2924         unlock_page_cgroup(pc);
2925         /*
2926          * If the page is not charged at this point,
2927          * we return here.
2928          */
2929         if (!mem)
2930                 return 0;
2931
2932         *ptr = mem;
2933         ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2934         css_put(&mem->css);/* drop extra refcnt */
2935         if (ret || *ptr == NULL) {
2936                 if (PageAnon(page)) {
2937                         lock_page_cgroup(pc);
2938                         ClearPageCgroupMigration(pc);
2939                         unlock_page_cgroup(pc);
2940                         /*
2941                          * The old page may be fully unmapped while we kept it.
2942                          */
2943                         mem_cgroup_uncharge_page(page);
2944                 }
2945                 return -ENOMEM;
2946         }
2947         /*
2948          * We charge new page before it's used/mapped. So, even if unlock_page()
2949          * is called before end_migration, we can catch all events on this new
2950          * page. In the case new page is migrated but not remapped, new page's
2951          * mapcount will be finally 0 and we call uncharge in end_migration().
2952          */
2953         pc = lookup_page_cgroup(newpage);
2954         if (PageAnon(page))
2955                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2956         else if (page_is_file_cache(page))
2957                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2958         else
2959                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2960         __mem_cgroup_commit_charge(mem, page, pc, ctype, PAGE_SIZE);
2961         return ret;
2962 }
2963
2964 /* remove redundant charge if migration failed*/
2965 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2966         struct page *oldpage, struct page *newpage, bool migration_ok)
2967 {
2968         struct page *used, *unused;
2969         struct page_cgroup *pc;
2970
2971         if (!mem)
2972                 return;
2973         /* blocks rmdir() */
2974         cgroup_exclude_rmdir(&mem->css);
2975         if (!migration_ok) {
2976                 used = oldpage;
2977                 unused = newpage;
2978         } else {
2979                 used = newpage;
2980                 unused = oldpage;
2981         }
2982         /*
2983          * We disallowed uncharge of pages under migration because mapcount
2984          * of the page goes down to zero, temporarly.
2985          * Clear the flag and check the page should be charged.
2986          */
2987         pc = lookup_page_cgroup(oldpage);
2988         lock_page_cgroup(pc);
2989         ClearPageCgroupMigration(pc);
2990         unlock_page_cgroup(pc);
2991
2992         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2993
2994         /*
2995          * If a page is a file cache, radix-tree replacement is very atomic
2996          * and we can skip this check. When it was an Anon page, its mapcount
2997          * goes down to 0. But because we added MIGRATION flage, it's not
2998          * uncharged yet. There are several case but page->mapcount check
2999          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3000          * check. (see prepare_charge() also)
3001          */
3002         if (PageAnon(used))
3003                 mem_cgroup_uncharge_page(used);
3004         /*
3005          * At migration, we may charge account against cgroup which has no
3006          * tasks.
3007          * So, rmdir()->pre_destroy() can be called while we do this charge.
3008          * In that case, we need to call pre_destroy() again. check it here.
3009          */
3010         cgroup_release_and_wakeup_rmdir(&mem->css);
3011 }
3012
3013 /*
3014  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3015  * Calling hierarchical_reclaim is not enough because we should update
3016  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3017  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3018  * not from the memcg which this page would be charged to.
3019  * try_charge_swapin does all of these works properly.
3020  */
3021 int mem_cgroup_shmem_charge_fallback(struct page *page,
3022                             struct mm_struct *mm,
3023                             gfp_t gfp_mask)
3024 {
3025         struct mem_cgroup *mem;
3026         int ret;
3027
3028         if (mem_cgroup_disabled())
3029                 return 0;
3030
3031         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3032         if (!ret)
3033                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3034
3035         return ret;
3036 }
3037
3038 #ifdef CONFIG_DEBUG_VM
3039 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3040 {
3041         struct page_cgroup *pc;
3042
3043         pc = lookup_page_cgroup(page);
3044         if (likely(pc) && PageCgroupUsed(pc))
3045                 return pc;
3046         return NULL;
3047 }
3048
3049 bool mem_cgroup_bad_page_check(struct page *page)
3050 {
3051         if (mem_cgroup_disabled())
3052                 return false;
3053
3054         return lookup_page_cgroup_used(page) != NULL;
3055 }
3056
3057 void mem_cgroup_print_bad_page(struct page *page)
3058 {
3059         struct page_cgroup *pc;
3060
3061         pc = lookup_page_cgroup_used(page);
3062         if (pc) {
3063                 int ret = -1;
3064                 char *path;
3065
3066                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3067                        pc, pc->flags, pc->mem_cgroup);
3068
3069                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3070                 if (path) {
3071                         rcu_read_lock();
3072                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3073                                                         path, PATH_MAX);
3074                         rcu_read_unlock();
3075                 }
3076
3077                 printk(KERN_CONT "(%s)\n",
3078                                 (ret < 0) ? "cannot get the path" : path);
3079                 kfree(path);
3080         }
3081 }
3082 #endif
3083
3084 static DEFINE_MUTEX(set_limit_mutex);
3085
3086 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3087                                 unsigned long long val)
3088 {
3089         int retry_count;
3090         u64 memswlimit, memlimit;
3091         int ret = 0;
3092         int children = mem_cgroup_count_children(memcg);
3093         u64 curusage, oldusage;
3094         int enlarge;
3095
3096         /*
3097          * For keeping hierarchical_reclaim simple, how long we should retry
3098          * is depends on callers. We set our retry-count to be function
3099          * of # of children which we should visit in this loop.
3100          */
3101         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3102
3103         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3104
3105         enlarge = 0;
3106         while (retry_count) {
3107                 if (signal_pending(current)) {
3108                         ret = -EINTR;
3109                         break;
3110                 }
3111                 /*
3112                  * Rather than hide all in some function, I do this in
3113                  * open coded manner. You see what this really does.
3114                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3115                  */
3116                 mutex_lock(&set_limit_mutex);
3117                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3118                 if (memswlimit < val) {
3119                         ret = -EINVAL;
3120                         mutex_unlock(&set_limit_mutex);
3121                         break;
3122                 }
3123
3124                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3125                 if (memlimit < val)
3126                         enlarge = 1;
3127
3128                 ret = res_counter_set_limit(&memcg->res, val);
3129                 if (!ret) {
3130                         if (memswlimit == val)
3131                                 memcg->memsw_is_minimum = true;
3132                         else
3133                                 memcg->memsw_is_minimum = false;
3134                 }
3135                 mutex_unlock(&set_limit_mutex);
3136
3137                 if (!ret)
3138                         break;
3139
3140                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3141                                                 MEM_CGROUP_RECLAIM_SHRINK);
3142                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3143                 /* Usage is reduced ? */
3144                 if (curusage >= oldusage)
3145                         retry_count--;
3146                 else
3147                         oldusage = curusage;
3148         }
3149         if (!ret && enlarge)
3150                 memcg_oom_recover(memcg);
3151
3152         return ret;
3153 }
3154
3155 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3156                                         unsigned long long val)
3157 {
3158         int retry_count;
3159         u64 memlimit, memswlimit, oldusage, curusage;
3160         int children = mem_cgroup_count_children(memcg);
3161         int ret = -EBUSY;
3162         int enlarge = 0;
3163
3164         /* see mem_cgroup_resize_res_limit */
3165         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3166         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3167         while (retry_count) {
3168                 if (signal_pending(current)) {
3169                         ret = -EINTR;
3170                         break;
3171                 }
3172                 /*
3173                  * Rather than hide all in some function, I do this in
3174                  * open coded manner. You see what this really does.
3175                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3176                  */
3177                 mutex_lock(&set_limit_mutex);
3178                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3179                 if (memlimit > val) {
3180                         ret = -EINVAL;
3181                         mutex_unlock(&set_limit_mutex);
3182                         break;
3183                 }
3184                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3185                 if (memswlimit < val)
3186                         enlarge = 1;
3187                 ret = res_counter_set_limit(&memcg->memsw, val);
3188                 if (!ret) {
3189                         if (memlimit == val)
3190                                 memcg->memsw_is_minimum = true;
3191                         else
3192                                 memcg->memsw_is_minimum = false;
3193                 }
3194                 mutex_unlock(&set_limit_mutex);
3195
3196                 if (!ret)
3197                         break;
3198
3199                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3200                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3201                                                 MEM_CGROUP_RECLAIM_SHRINK);
3202                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3203                 /* Usage is reduced ? */
3204                 if (curusage >= oldusage)
3205                         retry_count--;
3206                 else
3207                         oldusage = curusage;
3208         }
3209         if (!ret && enlarge)
3210                 memcg_oom_recover(memcg);
3211         return ret;
3212 }
3213
3214 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3215                                             gfp_t gfp_mask)
3216 {
3217         unsigned long nr_reclaimed = 0;
3218         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3219         unsigned long reclaimed;
3220         int loop = 0;
3221         struct mem_cgroup_tree_per_zone *mctz;
3222         unsigned long long excess;
3223
3224         if (order > 0)
3225                 return 0;
3226
3227         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3228         /*
3229          * This loop can run a while, specially if mem_cgroup's continuously
3230          * keep exceeding their soft limit and putting the system under
3231          * pressure
3232          */
3233         do {
3234                 if (next_mz)
3235                         mz = next_mz;
3236                 else
3237                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3238                 if (!mz)
3239                         break;
3240
3241                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3242                                                 gfp_mask,
3243                                                 MEM_CGROUP_RECLAIM_SOFT);
3244                 nr_reclaimed += reclaimed;
3245                 spin_lock(&mctz->lock);
3246
3247                 /*
3248                  * If we failed to reclaim anything from this memory cgroup
3249                  * it is time to move on to the next cgroup
3250                  */
3251                 next_mz = NULL;
3252                 if (!reclaimed) {
3253                         do {
3254                                 /*
3255                                  * Loop until we find yet another one.
3256                                  *
3257                                  * By the time we get the soft_limit lock
3258                                  * again, someone might have aded the
3259                                  * group back on the RB tree. Iterate to
3260                                  * make sure we get a different mem.
3261                                  * mem_cgroup_largest_soft_limit_node returns
3262                                  * NULL if no other cgroup is present on
3263                                  * the tree
3264                                  */
3265                                 next_mz =
3266                                 __mem_cgroup_largest_soft_limit_node(mctz);
3267                                 if (next_mz == mz) {
3268                                         css_put(&next_mz->mem->css);
3269                                         next_mz = NULL;
3270                                 } else /* next_mz == NULL or other memcg */
3271                                         break;
3272                         } while (1);
3273                 }
3274                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3275                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3276                 /*
3277                  * One school of thought says that we should not add
3278                  * back the node to the tree if reclaim returns 0.
3279                  * But our reclaim could return 0, simply because due
3280                  * to priority we are exposing a smaller subset of
3281                  * memory to reclaim from. Consider this as a longer
3282                  * term TODO.
3283                  */
3284                 /* If excess == 0, no tree ops */
3285                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3286                 spin_unlock(&mctz->lock);
3287                 css_put(&mz->mem->css);
3288                 loop++;
3289                 /*
3290                  * Could not reclaim anything and there are no more
3291                  * mem cgroups to try or we seem to be looping without
3292                  * reclaiming anything.
3293                  */
3294                 if (!nr_reclaimed &&
3295                         (next_mz == NULL ||
3296                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3297                         break;
3298         } while (!nr_reclaimed);
3299         if (next_mz)
3300                 css_put(&next_mz->mem->css);
3301         return nr_reclaimed;
3302 }
3303
3304 /*
3305  * This routine traverse page_cgroup in given list and drop them all.
3306  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3307  */
3308 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3309                                 int node, int zid, enum lru_list lru)
3310 {
3311         struct zone *zone;
3312         struct mem_cgroup_per_zone *mz;
3313         struct page_cgroup *pc, *busy;
3314         unsigned long flags, loop;
3315         struct list_head *list;
3316         int ret = 0;
3317
3318         zone = &NODE_DATA(node)->node_zones[zid];
3319         mz = mem_cgroup_zoneinfo(mem, node, zid);
3320         list = &mz->lists[lru];
3321
3322         loop = MEM_CGROUP_ZSTAT(mz, lru);
3323         /* give some margin against EBUSY etc...*/
3324         loop += 256;
3325         busy = NULL;
3326         while (loop--) {
3327                 struct page *page;
3328
3329                 ret = 0;
3330                 spin_lock_irqsave(&zone->lru_lock, flags);
3331                 if (list_empty(list)) {
3332                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3333                         break;
3334                 }
3335                 pc = list_entry(list->prev, struct page_cgroup, lru);
3336                 if (busy == pc) {
3337                         list_move(&pc->lru, list);
3338                         busy = NULL;
3339                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3340                         continue;
3341                 }
3342                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3343
3344                 page = lookup_cgroup_page(pc);
3345
3346                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3347                 if (ret == -ENOMEM)
3348                         break;
3349
3350                 if (ret == -EBUSY || ret == -EINVAL) {
3351                         /* found lock contention or "pc" is obsolete. */
3352                         busy = pc;
3353                         cond_resched();
3354                 } else
3355                         busy = NULL;
3356         }
3357
3358         if (!ret && !list_empty(list))
3359                 return -EBUSY;
3360         return ret;
3361 }
3362
3363 /*
3364  * make mem_cgroup's charge to be 0 if there is no task.
3365  * This enables deleting this mem_cgroup.
3366  */
3367 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3368 {
3369         int ret;
3370         int node, zid, shrink;
3371         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3372         struct cgroup *cgrp = mem->css.cgroup;
3373
3374         css_get(&mem->css);
3375
3376         shrink = 0;
3377         /* should free all ? */
3378         if (free_all)
3379                 goto try_to_free;
3380 move_account:
3381         do {
3382                 ret = -EBUSY;
3383                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3384                         goto out;
3385                 ret = -EINTR;
3386                 if (signal_pending(current))
3387                         goto out;
3388                 /* This is for making all *used* pages to be on LRU. */
3389                 lru_add_drain_all();
3390                 drain_all_stock_sync();
3391                 ret = 0;
3392                 mem_cgroup_start_move(mem);
3393                 for_each_node_state(node, N_HIGH_MEMORY) {
3394                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3395                                 enum lru_list l;
3396                                 for_each_lru(l) {
3397                                         ret = mem_cgroup_force_empty_list(mem,
3398                                                         node, zid, l);
3399                                         if (ret)
3400                                                 break;
3401                                 }
3402                         }
3403                         if (ret)
3404                                 break;
3405                 }
3406                 mem_cgroup_end_move(mem);
3407                 memcg_oom_recover(mem);
3408                 /* it seems parent cgroup doesn't have enough mem */
3409                 if (ret == -ENOMEM)
3410                         goto try_to_free;
3411                 cond_resched();
3412         /* "ret" should also be checked to ensure all lists are empty. */
3413         } while (mem->res.usage > 0 || ret);
3414 out:
3415         css_put(&mem->css);
3416         return ret;
3417
3418 try_to_free:
3419         /* returns EBUSY if there is a task or if we come here twice. */
3420         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3421                 ret = -EBUSY;
3422                 goto out;
3423         }
3424         /* we call try-to-free pages for make this cgroup empty */
3425         lru_add_drain_all();
3426         /* try to free all pages in this cgroup */
3427         shrink = 1;
3428         while (nr_retries && mem->res.usage > 0) {
3429                 int progress;
3430
3431                 if (signal_pending(current)) {
3432                         ret = -EINTR;
3433                         goto out;
3434                 }
3435                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3436                                                 false, get_swappiness(mem));
3437                 if (!progress) {
3438                         nr_retries--;
3439                         /* maybe some writeback is necessary */
3440                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3441                 }
3442
3443         }
3444         lru_add_drain();
3445         /* try move_account...there may be some *locked* pages. */
3446         goto move_account;
3447 }
3448
3449 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3450 {
3451         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3452 }
3453
3454
3455 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3456 {
3457         return mem_cgroup_from_cont(cont)->use_hierarchy;
3458 }
3459
3460 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3461                                         u64 val)
3462 {
3463         int retval = 0;
3464         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3465         struct cgroup *parent = cont->parent;
3466         struct mem_cgroup *parent_mem = NULL;
3467
3468         if (parent)
3469                 parent_mem = mem_cgroup_from_cont(parent);
3470
3471         cgroup_lock();
3472         /*
3473          * If parent's use_hierarchy is set, we can't make any modifications
3474          * in the child subtrees. If it is unset, then the change can
3475          * occur, provided the current cgroup has no children.
3476          *
3477          * For the root cgroup, parent_mem is NULL, we allow value to be
3478          * set if there are no children.
3479          */
3480         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3481                                 (val == 1 || val == 0)) {
3482                 if (list_empty(&cont->children))
3483                         mem->use_hierarchy = val;
3484                 else
3485                         retval = -EBUSY;
3486         } else
3487                 retval = -EINVAL;
3488         cgroup_unlock();
3489
3490         return retval;
3491 }
3492
3493
3494 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3495                                 enum mem_cgroup_stat_index idx)
3496 {
3497         struct mem_cgroup *iter;
3498         s64 val = 0;
3499
3500         /* each per cpu's value can be minus.Then, use s64 */
3501         for_each_mem_cgroup_tree(iter, mem)
3502                 val += mem_cgroup_read_stat(iter, idx);
3503
3504         if (val < 0) /* race ? */
3505                 val = 0;
3506         return val;
3507 }
3508
3509 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3510 {
3511         u64 val;
3512
3513         if (!mem_cgroup_is_root(mem)) {
3514                 if (!swap)
3515                         return res_counter_read_u64(&mem->res, RES_USAGE);
3516                 else
3517                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3518         }
3519
3520         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3521         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3522
3523         if (swap)
3524                 val += mem_cgroup_get_recursive_idx_stat(mem,
3525                                 MEM_CGROUP_STAT_SWAPOUT);
3526
3527         return val << PAGE_SHIFT;
3528 }
3529
3530 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3531 {
3532         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3533         u64 val;
3534         int type, name;
3535
3536         type = MEMFILE_TYPE(cft->private);
3537         name = MEMFILE_ATTR(cft->private);
3538         switch (type) {
3539         case _MEM:
3540                 if (name == RES_USAGE)
3541                         val = mem_cgroup_usage(mem, false);
3542                 else
3543                         val = res_counter_read_u64(&mem->res, name);
3544                 break;
3545         case _MEMSWAP:
3546                 if (name == RES_USAGE)
3547                         val = mem_cgroup_usage(mem, true);
3548                 else
3549                         val = res_counter_read_u64(&mem->memsw, name);
3550                 break;
3551         default:
3552                 BUG();
3553                 break;
3554         }
3555         return val;
3556 }
3557 /*
3558  * The user of this function is...
3559  * RES_LIMIT.
3560  */
3561 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3562                             const char *buffer)
3563 {
3564         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3565         int type, name;
3566         unsigned long long val;
3567         int ret;
3568
3569         type = MEMFILE_TYPE(cft->private);
3570         name = MEMFILE_ATTR(cft->private);
3571         switch (name) {
3572         case RES_LIMIT:
3573                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3574                         ret = -EINVAL;
3575                         break;
3576                 }
3577                 /* This function does all necessary parse...reuse it */
3578                 ret = res_counter_memparse_write_strategy(buffer, &val);
3579                 if (ret)
3580                         break;
3581                 if (type == _MEM)
3582                         ret = mem_cgroup_resize_limit(memcg, val);
3583                 else
3584                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3585                 break;
3586         case RES_SOFT_LIMIT:
3587                 ret = res_counter_memparse_write_strategy(buffer, &val);
3588                 if (ret)
3589                         break;
3590                 /*
3591                  * For memsw, soft limits are hard to implement in terms
3592                  * of semantics, for now, we support soft limits for
3593                  * control without swap
3594                  */
3595                 if (type == _MEM)
3596                         ret = res_counter_set_soft_limit(&memcg->res, val);
3597                 else
3598                         ret = -EINVAL;
3599                 break;
3600         default:
3601                 ret = -EINVAL; /* should be BUG() ? */
3602                 break;
3603         }
3604         return ret;
3605 }
3606
3607 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3608                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3609 {
3610         struct cgroup *cgroup;
3611         unsigned long long min_limit, min_memsw_limit, tmp;
3612
3613         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3614         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3615         cgroup = memcg->css.cgroup;
3616         if (!memcg->use_hierarchy)
3617                 goto out;
3618
3619         while (cgroup->parent) {
3620                 cgroup = cgroup->parent;
3621                 memcg = mem_cgroup_from_cont(cgroup);
3622                 if (!memcg->use_hierarchy)
3623                         break;
3624                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3625                 min_limit = min(min_limit, tmp);
3626                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3627                 min_memsw_limit = min(min_memsw_limit, tmp);
3628         }
3629 out:
3630         *mem_limit = min_limit;
3631         *memsw_limit = min_memsw_limit;
3632         return;
3633 }
3634
3635 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3636 {
3637         struct mem_cgroup *mem;
3638         int type, name;
3639
3640         mem = mem_cgroup_from_cont(cont);
3641         type = MEMFILE_TYPE(event);
3642         name = MEMFILE_ATTR(event);
3643         switch (name) {
3644         case RES_MAX_USAGE:
3645                 if (type == _MEM)
3646                         res_counter_reset_max(&mem->res);
3647                 else
3648                         res_counter_reset_max(&mem->memsw);
3649                 break;
3650         case RES_FAILCNT:
3651                 if (type == _MEM)
3652                         res_counter_reset_failcnt(&mem->res);
3653                 else
3654                         res_counter_reset_failcnt(&mem->memsw);
3655                 break;
3656         }
3657
3658         return 0;
3659 }
3660
3661 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3662                                         struct cftype *cft)
3663 {
3664         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3665 }
3666
3667 #ifdef CONFIG_MMU
3668 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3669                                         struct cftype *cft, u64 val)
3670 {
3671         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3672
3673         if (val >= (1 << NR_MOVE_TYPE))
3674                 return -EINVAL;
3675         /*
3676          * We check this value several times in both in can_attach() and
3677          * attach(), so we need cgroup lock to prevent this value from being
3678          * inconsistent.
3679          */
3680         cgroup_lock();
3681         mem->move_charge_at_immigrate = val;
3682         cgroup_unlock();
3683
3684         return 0;
3685 }
3686 #else
3687 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3688                                         struct cftype *cft, u64 val)
3689 {
3690         return -ENOSYS;
3691 }
3692 #endif
3693
3694
3695 /* For read statistics */
3696 enum {
3697         MCS_CACHE,
3698         MCS_RSS,
3699         MCS_FILE_MAPPED,
3700         MCS_PGPGIN,
3701         MCS_PGPGOUT,
3702         MCS_SWAP,
3703         MCS_INACTIVE_ANON,
3704         MCS_ACTIVE_ANON,
3705         MCS_INACTIVE_FILE,
3706         MCS_ACTIVE_FILE,
3707         MCS_UNEVICTABLE,
3708         NR_MCS_STAT,
3709 };
3710
3711 struct mcs_total_stat {
3712         s64 stat[NR_MCS_STAT];
3713 };
3714
3715 struct {
3716         char *local_name;
3717         char *total_name;
3718 } memcg_stat_strings[NR_MCS_STAT] = {
3719         {"cache", "total_cache"},
3720         {"rss", "total_rss"},
3721         {"mapped_file", "total_mapped_file"},
3722         {"pgpgin", "total_pgpgin"},
3723         {"pgpgout", "total_pgpgout"},
3724         {"swap", "total_swap"},
3725         {"inactive_anon", "total_inactive_anon"},
3726         {"active_anon", "total_active_anon"},
3727         {"inactive_file", "total_inactive_file"},
3728         {"active_file", "total_active_file"},
3729         {"unevictable", "total_unevictable"}
3730 };
3731
3732
3733 static void
3734 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3735 {
3736         s64 val;
3737
3738         /* per cpu stat */
3739         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3740         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3741         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3742         s->stat[MCS_RSS] += val * PAGE_SIZE;
3743         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3744         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3745         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3746         s->stat[MCS_PGPGIN] += val;
3747         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3748         s->stat[MCS_PGPGOUT] += val;
3749         if (do_swap_account) {
3750                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3751                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3752         }
3753
3754         /* per zone stat */
3755         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3756         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3757         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3758         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3759         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3760         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3761         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3762         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3763         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3764         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3765 }
3766
3767 static void
3768 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3769 {
3770         struct mem_cgroup *iter;
3771
3772         for_each_mem_cgroup_tree(iter, mem)
3773                 mem_cgroup_get_local_stat(iter, s);
3774 }
3775
3776 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3777                                  struct cgroup_map_cb *cb)
3778 {
3779         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3780         struct mcs_total_stat mystat;
3781         int i;
3782
3783         memset(&mystat, 0, sizeof(mystat));
3784         mem_cgroup_get_local_stat(mem_cont, &mystat);
3785
3786         for (i = 0; i < NR_MCS_STAT; i++) {
3787                 if (i == MCS_SWAP && !do_swap_account)
3788                         continue;
3789                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3790         }
3791
3792         /* Hierarchical information */
3793         {
3794                 unsigned long long limit, memsw_limit;
3795                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3796                 cb->fill(cb, "hierarchical_memory_limit", limit);
3797                 if (do_swap_account)
3798                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3799         }
3800
3801         memset(&mystat, 0, sizeof(mystat));
3802         mem_cgroup_get_total_stat(mem_cont, &mystat);
3803         for (i = 0; i < NR_MCS_STAT; i++) {
3804                 if (i == MCS_SWAP && !do_swap_account)
3805                         continue;
3806                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3807         }
3808
3809 #ifdef CONFIG_DEBUG_VM
3810         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3811
3812         {
3813                 int nid, zid;
3814                 struct mem_cgroup_per_zone *mz;
3815                 unsigned long recent_rotated[2] = {0, 0};
3816                 unsigned long recent_scanned[2] = {0, 0};
3817
3818                 for_each_online_node(nid)
3819                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3820                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3821
3822                                 recent_rotated[0] +=
3823                                         mz->reclaim_stat.recent_rotated[0];
3824                                 recent_rotated[1] +=
3825                                         mz->reclaim_stat.recent_rotated[1];
3826                                 recent_scanned[0] +=
3827                                         mz->reclaim_stat.recent_scanned[0];
3828                                 recent_scanned[1] +=
3829                                         mz->reclaim_stat.recent_scanned[1];
3830                         }
3831                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3832                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3833                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3834                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3835         }
3836 #endif
3837
3838         return 0;
3839 }
3840
3841 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3842 {
3843         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3844
3845         return get_swappiness(memcg);
3846 }
3847
3848 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3849                                        u64 val)
3850 {
3851         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3852         struct mem_cgroup *parent;
3853
3854         if (val > 100)
3855                 return -EINVAL;
3856
3857         if (cgrp->parent == NULL)
3858                 return -EINVAL;
3859
3860         parent = mem_cgroup_from_cont(cgrp->parent);
3861
3862         cgroup_lock();
3863
3864         /* If under hierarchy, only empty-root can set this value */
3865         if ((parent->use_hierarchy) ||
3866             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3867                 cgroup_unlock();
3868                 return -EINVAL;
3869         }
3870
3871         spin_lock(&memcg->reclaim_param_lock);
3872         memcg->swappiness = val;
3873         spin_unlock(&memcg->reclaim_param_lock);
3874
3875         cgroup_unlock();
3876
3877         return 0;
3878 }
3879
3880 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3881 {
3882         struct mem_cgroup_threshold_ary *t;
3883         u64 usage;
3884         int i;
3885
3886         rcu_read_lock();
3887         if (!swap)
3888                 t = rcu_dereference(memcg->thresholds.primary);
3889         else
3890                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3891
3892         if (!t)
3893                 goto unlock;
3894
3895         usage = mem_cgroup_usage(memcg, swap);
3896
3897         /*
3898          * current_threshold points to threshold just below usage.
3899          * If it's not true, a threshold was crossed after last
3900          * call of __mem_cgroup_threshold().
3901          */
3902         i = t->current_threshold;
3903
3904         /*
3905          * Iterate backward over array of thresholds starting from
3906          * current_threshold and check if a threshold is crossed.
3907          * If none of thresholds below usage is crossed, we read
3908          * only one element of the array here.
3909          */
3910         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3911                 eventfd_signal(t->entries[i].eventfd, 1);
3912
3913         /* i = current_threshold + 1 */
3914         i++;
3915
3916         /*
3917          * Iterate forward over array of thresholds starting from
3918          * current_threshold+1 and check if a threshold is crossed.
3919          * If none of thresholds above usage is crossed, we read
3920          * only one element of the array here.
3921          */
3922         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3923                 eventfd_signal(t->entries[i].eventfd, 1);
3924
3925         /* Update current_threshold */
3926         t->current_threshold = i - 1;
3927 unlock:
3928         rcu_read_unlock();
3929 }
3930
3931 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3932 {
3933         while (memcg) {
3934                 __mem_cgroup_threshold(memcg, false);
3935                 if (do_swap_account)
3936                         __mem_cgroup_threshold(memcg, true);
3937
3938                 memcg = parent_mem_cgroup(memcg);
3939         }
3940 }
3941
3942 static int compare_thresholds(const void *a, const void *b)
3943 {
3944         const struct mem_cgroup_threshold *_a = a;
3945         const struct mem_cgroup_threshold *_b = b;
3946
3947         return _a->threshold - _b->threshold;
3948 }
3949
3950 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3951 {
3952         struct mem_cgroup_eventfd_list *ev;
3953
3954         list_for_each_entry(ev, &mem->oom_notify, list)
3955                 eventfd_signal(ev->eventfd, 1);
3956         return 0;
3957 }
3958
3959 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3960 {
3961         struct mem_cgroup *iter;
3962
3963         for_each_mem_cgroup_tree(iter, mem)
3964                 mem_cgroup_oom_notify_cb(iter);
3965 }
3966
3967 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3968         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3969 {
3970         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3971         struct mem_cgroup_thresholds *thresholds;
3972         struct mem_cgroup_threshold_ary *new;
3973         int type = MEMFILE_TYPE(cft->private);
3974         u64 threshold, usage;
3975         int i, size, ret;
3976
3977         ret = res_counter_memparse_write_strategy(args, &threshold);
3978         if (ret)
3979                 return ret;
3980
3981         mutex_lock(&memcg->thresholds_lock);
3982
3983         if (type == _MEM)
3984                 thresholds = &memcg->thresholds;
3985         else if (type == _MEMSWAP)
3986                 thresholds = &memcg->memsw_thresholds;
3987         else
3988                 BUG();
3989
3990         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3991
3992         /* Check if a threshold crossed before adding a new one */
3993         if (thresholds->primary)
3994                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3995
3996         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3997
3998         /* Allocate memory for new array of thresholds */
3999         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4000                         GFP_KERNEL);
4001         if (!new) {
4002                 ret = -ENOMEM;
4003                 goto unlock;
4004         }
4005         new->size = size;
4006
4007         /* Copy thresholds (if any) to new array */
4008         if (thresholds->primary) {
4009                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4010                                 sizeof(struct mem_cgroup_threshold));
4011         }
4012
4013         /* Add new threshold */
4014         new->entries[size - 1].eventfd = eventfd;
4015         new->entries[size - 1].threshold = threshold;
4016
4017         /* Sort thresholds. Registering of new threshold isn't time-critical */
4018         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4019                         compare_thresholds, NULL);
4020
4021         /* Find current threshold */
4022         new->current_threshold = -1;
4023         for (i = 0; i < size; i++) {
4024                 if (new->entries[i].threshold < usage) {
4025                         /*
4026                          * new->current_threshold will not be used until
4027                          * rcu_assign_pointer(), so it's safe to increment
4028                          * it here.
4029                          */
4030                         ++new->current_threshold;
4031                 }
4032         }
4033
4034         /* Free old spare buffer and save old primary buffer as spare */
4035         kfree(thresholds->spare);
4036         thresholds->spare = thresholds->primary;
4037
4038         rcu_assign_pointer(thresholds->primary, new);
4039
4040         /* To be sure that nobody uses thresholds */
4041         synchronize_rcu();
4042
4043 unlock:
4044         mutex_unlock(&memcg->thresholds_lock);
4045
4046         return ret;
4047 }
4048
4049 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4050         struct cftype *cft, struct eventfd_ctx *eventfd)
4051 {
4052         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4053         struct mem_cgroup_thresholds *thresholds;
4054         struct mem_cgroup_threshold_ary *new;
4055         int type = MEMFILE_TYPE(cft->private);
4056         u64 usage;
4057         int i, j, size;
4058
4059         mutex_lock(&memcg->thresholds_lock);
4060         if (type == _MEM)
4061                 thresholds = &memcg->thresholds;
4062         else if (type == _MEMSWAP)
4063                 thresholds = &memcg->memsw_thresholds;
4064         else
4065                 BUG();
4066
4067         /*
4068          * Something went wrong if we trying to unregister a threshold
4069          * if we don't have thresholds
4070          */
4071         BUG_ON(!thresholds);
4072
4073         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4074
4075         /* Check if a threshold crossed before removing */
4076         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4077
4078         /* Calculate new number of threshold */
4079         size = 0;
4080         for (i = 0; i < thresholds->primary->size; i++) {
4081                 if (thresholds->primary->entries[i].eventfd != eventfd)
4082                         size++;
4083         }
4084
4085         new = thresholds->spare;
4086
4087         /* Set thresholds array to NULL if we don't have thresholds */
4088         if (!size) {
4089                 kfree(new);
4090                 new = NULL;
4091                 goto swap_buffers;
4092         }
4093
4094         new->size = size;
4095
4096         /* Copy thresholds and find current threshold */
4097         new->current_threshold = -1;
4098         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4099                 if (thresholds->primary->entries[i].eventfd == eventfd)
4100                         continue;
4101
4102                 new->entries[j] = thresholds->primary->entries[i];
4103                 if (new->entries[j].threshold < usage) {
4104                         /*
4105                          * new->current_threshold will not be used
4106                          * until rcu_assign_pointer(), so it's safe to increment
4107                          * it here.
4108                          */
4109                         ++new->current_threshold;
4110                 }
4111                 j++;
4112         }
4113
4114 swap_buffers:
4115         /* Swap primary and spare array */
4116         thresholds->spare = thresholds->primary;
4117         rcu_assign_pointer(thresholds->primary, new);
4118
4119         /* To be sure that nobody uses thresholds */
4120         synchronize_rcu();
4121
4122         mutex_unlock(&memcg->thresholds_lock);
4123 }
4124
4125 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4126         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4127 {
4128         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4129         struct mem_cgroup_eventfd_list *event;
4130         int type = MEMFILE_TYPE(cft->private);
4131
4132         BUG_ON(type != _OOM_TYPE);
4133         event = kmalloc(sizeof(*event), GFP_KERNEL);
4134         if (!event)
4135                 return -ENOMEM;
4136
4137         mutex_lock(&memcg_oom_mutex);
4138
4139         event->eventfd = eventfd;
4140         list_add(&event->list, &memcg->oom_notify);
4141
4142         /* already in OOM ? */
4143         if (atomic_read(&memcg->oom_lock))
4144                 eventfd_signal(eventfd, 1);
4145         mutex_unlock(&memcg_oom_mutex);
4146
4147         return 0;
4148 }
4149
4150 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4151         struct cftype *cft, struct eventfd_ctx *eventfd)
4152 {
4153         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4154         struct mem_cgroup_eventfd_list *ev, *tmp;
4155         int type = MEMFILE_TYPE(cft->private);
4156
4157         BUG_ON(type != _OOM_TYPE);
4158
4159         mutex_lock(&memcg_oom_mutex);
4160
4161         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4162                 if (ev->eventfd == eventfd) {
4163                         list_del(&ev->list);
4164                         kfree(ev);
4165                 }
4166         }
4167
4168         mutex_unlock(&memcg_oom_mutex);
4169 }
4170
4171 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4172         struct cftype *cft,  struct cgroup_map_cb *cb)
4173 {
4174         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4175
4176         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4177
4178         if (atomic_read(&mem->oom_lock))
4179                 cb->fill(cb, "under_oom", 1);
4180         else
4181                 cb->fill(cb, "under_oom", 0);
4182         return 0;
4183 }
4184
4185 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4186         struct cftype *cft, u64 val)
4187 {
4188         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4189         struct mem_cgroup *parent;
4190
4191         /* cannot set to root cgroup and only 0 and 1 are allowed */
4192         if (!cgrp->parent || !((val == 0) || (val == 1)))
4193                 return -EINVAL;
4194
4195         parent = mem_cgroup_from_cont(cgrp->parent);
4196
4197         cgroup_lock();
4198         /* oom-kill-disable is a flag for subhierarchy. */
4199         if ((parent->use_hierarchy) ||
4200             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4201                 cgroup_unlock();
4202                 return -EINVAL;
4203         }
4204         mem->oom_kill_disable = val;
4205         if (!val)
4206                 memcg_oom_recover(mem);
4207         cgroup_unlock();
4208         return 0;
4209 }
4210
4211 static struct cftype mem_cgroup_files[] = {
4212         {
4213                 .name = "usage_in_bytes",
4214                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4215                 .read_u64 = mem_cgroup_read,
4216                 .register_event = mem_cgroup_usage_register_event,
4217                 .unregister_event = mem_cgroup_usage_unregister_event,
4218         },
4219         {
4220                 .name = "max_usage_in_bytes",
4221                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4222                 .trigger = mem_cgroup_reset,
4223                 .read_u64 = mem_cgroup_read,
4224         },
4225         {
4226                 .name = "limit_in_bytes",
4227                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4228                 .write_string = mem_cgroup_write,
4229                 .read_u64 = mem_cgroup_read,
4230         },
4231         {
4232                 .name = "soft_limit_in_bytes",
4233                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4234                 .write_string = mem_cgroup_write,
4235                 .read_u64 = mem_cgroup_read,
4236         },
4237         {
4238                 .name = "failcnt",
4239                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4240                 .trigger = mem_cgroup_reset,
4241                 .read_u64 = mem_cgroup_read,
4242         },
4243         {
4244                 .name = "stat",
4245                 .read_map = mem_control_stat_show,
4246         },
4247         {
4248                 .name = "force_empty",
4249                 .trigger = mem_cgroup_force_empty_write,
4250         },
4251         {
4252                 .name = "use_hierarchy",
4253                 .write_u64 = mem_cgroup_hierarchy_write,
4254                 .read_u64 = mem_cgroup_hierarchy_read,
4255         },
4256         {
4257                 .name = "swappiness",
4258                 .read_u64 = mem_cgroup_swappiness_read,
4259                 .write_u64 = mem_cgroup_swappiness_write,
4260         },
4261         {
4262                 .name = "move_charge_at_immigrate",
4263                 .read_u64 = mem_cgroup_move_charge_read,
4264                 .write_u64 = mem_cgroup_move_charge_write,
4265         },
4266         {
4267                 .name = "oom_control",
4268                 .read_map = mem_cgroup_oom_control_read,
4269                 .write_u64 = mem_cgroup_oom_control_write,
4270                 .register_event = mem_cgroup_oom_register_event,
4271                 .unregister_event = mem_cgroup_oom_unregister_event,
4272                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4273         },
4274 };
4275
4276 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4277 static struct cftype memsw_cgroup_files[] = {
4278         {
4279                 .name = "memsw.usage_in_bytes",
4280                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4281                 .read_u64 = mem_cgroup_read,
4282                 .register_event = mem_cgroup_usage_register_event,
4283                 .unregister_event = mem_cgroup_usage_unregister_event,
4284         },
4285         {
4286                 .name = "memsw.max_usage_in_bytes",
4287                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4288                 .trigger = mem_cgroup_reset,
4289                 .read_u64 = mem_cgroup_read,
4290         },
4291         {
4292                 .name = "memsw.limit_in_bytes",
4293                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4294                 .write_string = mem_cgroup_write,
4295                 .read_u64 = mem_cgroup_read,
4296         },
4297         {
4298                 .name = "memsw.failcnt",
4299                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4300                 .trigger = mem_cgroup_reset,
4301                 .read_u64 = mem_cgroup_read,
4302         },
4303 };
4304
4305 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4306 {
4307         if (!do_swap_account)
4308                 return 0;
4309         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4310                                 ARRAY_SIZE(memsw_cgroup_files));
4311 };
4312 #else
4313 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4314 {
4315         return 0;
4316 }
4317 #endif
4318
4319 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4320 {
4321         struct mem_cgroup_per_node *pn;
4322         struct mem_cgroup_per_zone *mz;
4323         enum lru_list l;
4324         int zone, tmp = node;
4325         /*
4326          * This routine is called against possible nodes.
4327          * But it's BUG to call kmalloc() against offline node.
4328          *
4329          * TODO: this routine can waste much memory for nodes which will
4330          *       never be onlined. It's better to use memory hotplug callback
4331          *       function.
4332          */
4333         if (!node_state(node, N_NORMAL_MEMORY))
4334                 tmp = -1;
4335         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4336         if (!pn)
4337                 return 1;
4338
4339         mem->info.nodeinfo[node] = pn;
4340         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4341                 mz = &pn->zoneinfo[zone];
4342                 for_each_lru(l)
4343                         INIT_LIST_HEAD(&mz->lists[l]);
4344                 mz->usage_in_excess = 0;
4345                 mz->on_tree = false;
4346                 mz->mem = mem;
4347         }
4348         return 0;
4349 }
4350
4351 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4352 {
4353         kfree(mem->info.nodeinfo[node]);
4354 }
4355
4356 static struct mem_cgroup *mem_cgroup_alloc(void)
4357 {
4358         struct mem_cgroup *mem;
4359         int size = sizeof(struct mem_cgroup);
4360
4361         /* Can be very big if MAX_NUMNODES is very big */
4362         if (size < PAGE_SIZE)
4363                 mem = kzalloc(size, GFP_KERNEL);
4364         else
4365                 mem = vzalloc(size);
4366
4367         if (!mem)
4368                 return NULL;
4369
4370         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4371         if (!mem->stat)
4372                 goto out_free;
4373         spin_lock_init(&mem->pcp_counter_lock);
4374         return mem;
4375
4376 out_free:
4377         if (size < PAGE_SIZE)
4378                 kfree(mem);
4379         else
4380                 vfree(mem);
4381         return NULL;
4382 }
4383
4384 /*
4385  * At destroying mem_cgroup, references from swap_cgroup can remain.
4386  * (scanning all at force_empty is too costly...)
4387  *
4388  * Instead of clearing all references at force_empty, we remember
4389  * the number of reference from swap_cgroup and free mem_cgroup when
4390  * it goes down to 0.
4391  *
4392  * Removal of cgroup itself succeeds regardless of refs from swap.
4393  */
4394
4395 static void __mem_cgroup_free(struct mem_cgroup *mem)
4396 {
4397         int node;
4398
4399         mem_cgroup_remove_from_trees(mem);
4400         free_css_id(&mem_cgroup_subsys, &mem->css);
4401
4402         for_each_node_state(node, N_POSSIBLE)
4403                 free_mem_cgroup_per_zone_info(mem, node);
4404
4405         free_percpu(mem->stat);
4406         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4407                 kfree(mem);
4408         else
4409                 vfree(mem);
4410 }
4411
4412 static void mem_cgroup_get(struct mem_cgroup *mem)
4413 {
4414         atomic_inc(&mem->refcnt);
4415 }
4416
4417 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4418 {
4419         if (atomic_sub_and_test(count, &mem->refcnt)) {
4420                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4421                 __mem_cgroup_free(mem);
4422                 if (parent)
4423                         mem_cgroup_put(parent);
4424         }
4425 }
4426
4427 static void mem_cgroup_put(struct mem_cgroup *mem)
4428 {
4429         __mem_cgroup_put(mem, 1);
4430 }
4431
4432 /*
4433  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4434  */
4435 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4436 {
4437         if (!mem->res.parent)
4438                 return NULL;
4439         return mem_cgroup_from_res_counter(mem->res.parent, res);
4440 }
4441
4442 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4443 static void __init enable_swap_cgroup(void)
4444 {
4445         if (!mem_cgroup_disabled() && really_do_swap_account)
4446                 do_swap_account = 1;
4447 }
4448 #else
4449 static void __init enable_swap_cgroup(void)
4450 {
4451 }
4452 #endif
4453
4454 static int mem_cgroup_soft_limit_tree_init(void)
4455 {
4456         struct mem_cgroup_tree_per_node *rtpn;
4457         struct mem_cgroup_tree_per_zone *rtpz;
4458         int tmp, node, zone;
4459
4460         for_each_node_state(node, N_POSSIBLE) {
4461                 tmp = node;
4462                 if (!node_state(node, N_NORMAL_MEMORY))
4463                         tmp = -1;
4464                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4465                 if (!rtpn)
4466                         return 1;
4467
4468                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4469
4470                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4471                         rtpz = &rtpn->rb_tree_per_zone[zone];
4472                         rtpz->rb_root = RB_ROOT;
4473                         spin_lock_init(&rtpz->lock);
4474                 }
4475         }
4476         return 0;
4477 }
4478
4479 static struct cgroup_subsys_state * __ref
4480 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4481 {
4482         struct mem_cgroup *mem, *parent;
4483         long error = -ENOMEM;
4484         int node;
4485
4486         mem = mem_cgroup_alloc();
4487         if (!mem)
4488                 return ERR_PTR(error);
4489
4490         for_each_node_state(node, N_POSSIBLE)
4491                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4492                         goto free_out;
4493
4494         /* root ? */
4495         if (cont->parent == NULL) {
4496                 int cpu;
4497                 enable_swap_cgroup();
4498                 parent = NULL;
4499                 root_mem_cgroup = mem;
4500                 if (mem_cgroup_soft_limit_tree_init())
4501                         goto free_out;
4502                 for_each_possible_cpu(cpu) {
4503                         struct memcg_stock_pcp *stock =
4504                                                 &per_cpu(memcg_stock, cpu);
4505                         INIT_WORK(&stock->work, drain_local_stock);
4506                 }
4507                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4508         } else {
4509                 parent = mem_cgroup_from_cont(cont->parent);
4510                 mem->use_hierarchy = parent->use_hierarchy;
4511                 mem->oom_kill_disable = parent->oom_kill_disable;
4512         }
4513
4514         if (parent && parent->use_hierarchy) {
4515                 res_counter_init(&mem->res, &parent->res);
4516                 res_counter_init(&mem->memsw, &parent->memsw);
4517                 /*
4518                  * We increment refcnt of the parent to ensure that we can
4519                  * safely access it on res_counter_charge/uncharge.
4520                  * This refcnt will be decremented when freeing this
4521                  * mem_cgroup(see mem_cgroup_put).
4522                  */
4523                 mem_cgroup_get(parent);
4524         } else {
4525                 res_counter_init(&mem->res, NULL);
4526                 res_counter_init(&mem->memsw, NULL);
4527         }
4528         mem->last_scanned_child = 0;
4529         spin_lock_init(&mem->reclaim_param_lock);
4530         INIT_LIST_HEAD(&mem->oom_notify);
4531
4532         if (parent)
4533                 mem->swappiness = get_swappiness(parent);
4534         atomic_set(&mem->refcnt, 1);
4535         mem->move_charge_at_immigrate = 0;
4536         mutex_init(&mem->thresholds_lock);
4537         return &mem->css;
4538 free_out:
4539         __mem_cgroup_free(mem);
4540         root_mem_cgroup = NULL;
4541         return ERR_PTR(error);
4542 }
4543
4544 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4545                                         struct cgroup *cont)
4546 {
4547         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4548
4549         return mem_cgroup_force_empty(mem, false);
4550 }
4551
4552 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4553                                 struct cgroup *cont)
4554 {
4555         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4556
4557         mem_cgroup_put(mem);
4558 }
4559
4560 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4561                                 struct cgroup *cont)
4562 {
4563         int ret;
4564
4565         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4566                                 ARRAY_SIZE(mem_cgroup_files));
4567
4568         if (!ret)
4569                 ret = register_memsw_files(cont, ss);
4570         return ret;
4571 }
4572
4573 #ifdef CONFIG_MMU
4574 /* Handlers for move charge at task migration. */
4575 #define PRECHARGE_COUNT_AT_ONCE 256
4576 static int mem_cgroup_do_precharge(unsigned long count)
4577 {
4578         int ret = 0;
4579         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4580         struct mem_cgroup *mem = mc.to;
4581
4582         if (mem_cgroup_is_root(mem)) {
4583                 mc.precharge += count;
4584                 /* we don't need css_get for root */
4585                 return ret;
4586         }
4587         /* try to charge at once */
4588         if (count > 1) {
4589                 struct res_counter *dummy;
4590                 /*
4591                  * "mem" cannot be under rmdir() because we've already checked
4592                  * by cgroup_lock_live_cgroup() that it is not removed and we
4593                  * are still under the same cgroup_mutex. So we can postpone
4594                  * css_get().
4595                  */
4596                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4597                         goto one_by_one;
4598                 if (do_swap_account && res_counter_charge(&mem->memsw,
4599                                                 PAGE_SIZE * count, &dummy)) {
4600                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4601                         goto one_by_one;
4602                 }
4603                 mc.precharge += count;
4604                 return ret;
4605         }
4606 one_by_one:
4607         /* fall back to one by one charge */
4608         while (count--) {
4609                 if (signal_pending(current)) {
4610                         ret = -EINTR;
4611                         break;
4612                 }
4613                 if (!batch_count--) {
4614                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4615                         cond_resched();
4616                 }
4617                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4618                                               PAGE_SIZE);
4619                 if (ret || !mem)
4620                         /* mem_cgroup_clear_mc() will do uncharge later */
4621                         return -ENOMEM;
4622                 mc.precharge++;
4623         }
4624         return ret;
4625 }
4626
4627 /**
4628  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4629  * @vma: the vma the pte to be checked belongs
4630  * @addr: the address corresponding to the pte to be checked
4631  * @ptent: the pte to be checked
4632  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4633  *
4634  * Returns
4635  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4636  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4637  *     move charge. if @target is not NULL, the page is stored in target->page
4638  *     with extra refcnt got(Callers should handle it).
4639  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4640  *     target for charge migration. if @target is not NULL, the entry is stored
4641  *     in target->ent.
4642  *
4643  * Called with pte lock held.
4644  */
4645 union mc_target {
4646         struct page     *page;
4647         swp_entry_t     ent;
4648 };
4649
4650 enum mc_target_type {
4651         MC_TARGET_NONE, /* not used */
4652         MC_TARGET_PAGE,
4653         MC_TARGET_SWAP,
4654 };
4655
4656 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4657                                                 unsigned long addr, pte_t ptent)
4658 {
4659         struct page *page = vm_normal_page(vma, addr, ptent);
4660
4661         if (!page || !page_mapped(page))
4662                 return NULL;
4663         if (PageAnon(page)) {
4664                 /* we don't move shared anon */
4665                 if (!move_anon() || page_mapcount(page) > 2)
4666                         return NULL;
4667         } else if (!move_file())
4668                 /* we ignore mapcount for file pages */
4669                 return NULL;
4670         if (!get_page_unless_zero(page))
4671                 return NULL;
4672
4673         return page;
4674 }
4675
4676 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4677                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4678 {
4679         int usage_count;
4680         struct page *page = NULL;
4681         swp_entry_t ent = pte_to_swp_entry(ptent);
4682
4683         if (!move_anon() || non_swap_entry(ent))
4684                 return NULL;
4685         usage_count = mem_cgroup_count_swap_user(ent, &page);
4686         if (usage_count > 1) { /* we don't move shared anon */
4687                 if (page)
4688                         put_page(page);
4689                 return NULL;
4690         }
4691         if (do_swap_account)
4692                 entry->val = ent.val;
4693
4694         return page;
4695 }
4696
4697 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4698                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4699 {
4700         struct page *page = NULL;
4701         struct inode *inode;
4702         struct address_space *mapping;
4703         pgoff_t pgoff;
4704
4705         if (!vma->vm_file) /* anonymous vma */
4706                 return NULL;
4707         if (!move_file())
4708                 return NULL;
4709
4710         inode = vma->vm_file->f_path.dentry->d_inode;
4711         mapping = vma->vm_file->f_mapping;
4712         if (pte_none(ptent))
4713                 pgoff = linear_page_index(vma, addr);
4714         else /* pte_file(ptent) is true */
4715                 pgoff = pte_to_pgoff(ptent);
4716
4717         /* page is moved even if it's not RSS of this task(page-faulted). */
4718         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4719                 page = find_get_page(mapping, pgoff);
4720         } else { /* shmem/tmpfs file. we should take account of swap too. */
4721                 swp_entry_t ent;
4722                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4723                 if (do_swap_account)
4724                         entry->val = ent.val;
4725         }
4726
4727         return page;
4728 }
4729
4730 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4731                 unsigned long addr, pte_t ptent, union mc_target *target)
4732 {
4733         struct page *page = NULL;
4734         struct page_cgroup *pc;
4735         int ret = 0;
4736         swp_entry_t ent = { .val = 0 };
4737
4738         if (pte_present(ptent))
4739                 page = mc_handle_present_pte(vma, addr, ptent);
4740         else if (is_swap_pte(ptent))
4741                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4742         else if (pte_none(ptent) || pte_file(ptent))
4743                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4744
4745         if (!page && !ent.val)
4746                 return 0;
4747         if (page) {
4748                 pc = lookup_page_cgroup(page);
4749                 /*
4750                  * Do only loose check w/o page_cgroup lock.
4751                  * mem_cgroup_move_account() checks the pc is valid or not under
4752                  * the lock.
4753                  */
4754                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4755                         ret = MC_TARGET_PAGE;
4756                         if (target)
4757                                 target->page = page;
4758                 }
4759                 if (!ret || !target)
4760                         put_page(page);
4761         }
4762         /* There is a swap entry and a page doesn't exist or isn't charged */
4763         if (ent.val && !ret &&
4764                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4765                 ret = MC_TARGET_SWAP;
4766                 if (target)
4767                         target->ent = ent;
4768         }
4769         return ret;
4770 }
4771
4772 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4773                                         unsigned long addr, unsigned long end,
4774                                         struct mm_walk *walk)
4775 {
4776         struct vm_area_struct *vma = walk->private;
4777         pte_t *pte;
4778         spinlock_t *ptl;
4779
4780         split_huge_page_pmd(walk->mm, pmd);
4781
4782         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4783         for (; addr != end; pte++, addr += PAGE_SIZE)
4784                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4785                         mc.precharge++; /* increment precharge temporarily */
4786         pte_unmap_unlock(pte - 1, ptl);
4787         cond_resched();
4788
4789         return 0;
4790 }
4791
4792 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4793 {
4794         unsigned long precharge;
4795         struct vm_area_struct *vma;
4796
4797         down_read(&mm->mmap_sem);
4798         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4799                 struct mm_walk mem_cgroup_count_precharge_walk = {
4800                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4801                         .mm = mm,
4802                         .private = vma,
4803                 };
4804                 if (is_vm_hugetlb_page(vma))
4805                         continue;
4806                 walk_page_range(vma->vm_start, vma->vm_end,
4807                                         &mem_cgroup_count_precharge_walk);
4808         }
4809         up_read(&mm->mmap_sem);
4810
4811         precharge = mc.precharge;
4812         mc.precharge = 0;
4813
4814         return precharge;
4815 }
4816
4817 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4818 {
4819         unsigned long precharge = mem_cgroup_count_precharge(mm);
4820
4821         VM_BUG_ON(mc.moving_task);
4822         mc.moving_task = current;
4823         return mem_cgroup_do_precharge(precharge);
4824 }
4825
4826 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4827 static void __mem_cgroup_clear_mc(void)
4828 {
4829         struct mem_cgroup *from = mc.from;
4830         struct mem_cgroup *to = mc.to;
4831
4832         /* we must uncharge all the leftover precharges from mc.to */
4833         if (mc.precharge) {
4834                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4835                 mc.precharge = 0;
4836         }
4837         /*
4838          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4839          * we must uncharge here.
4840          */
4841         if (mc.moved_charge) {
4842                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4843                 mc.moved_charge = 0;
4844         }
4845         /* we must fixup refcnts and charges */
4846         if (mc.moved_swap) {
4847                 /* uncharge swap account from the old cgroup */
4848                 if (!mem_cgroup_is_root(mc.from))
4849                         res_counter_uncharge(&mc.from->memsw,
4850                                                 PAGE_SIZE * mc.moved_swap);
4851                 __mem_cgroup_put(mc.from, mc.moved_swap);
4852
4853                 if (!mem_cgroup_is_root(mc.to)) {
4854                         /*
4855                          * we charged both to->res and to->memsw, so we should
4856                          * uncharge to->res.
4857                          */
4858                         res_counter_uncharge(&mc.to->res,
4859                                                 PAGE_SIZE * mc.moved_swap);
4860                 }
4861                 /* we've already done mem_cgroup_get(mc.to) */
4862                 mc.moved_swap = 0;
4863         }
4864         memcg_oom_recover(from);
4865         memcg_oom_recover(to);
4866         wake_up_all(&mc.waitq);
4867 }
4868
4869 static void mem_cgroup_clear_mc(void)
4870 {
4871         struct mem_cgroup *from = mc.from;
4872
4873         /*
4874          * we must clear moving_task before waking up waiters at the end of
4875          * task migration.
4876          */
4877         mc.moving_task = NULL;
4878         __mem_cgroup_clear_mc();
4879         spin_lock(&mc.lock);
4880         mc.from = NULL;
4881         mc.to = NULL;
4882         spin_unlock(&mc.lock);
4883         mem_cgroup_end_move(from);
4884 }
4885
4886 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4887                                 struct cgroup *cgroup,
4888                                 struct task_struct *p,
4889                                 bool threadgroup)
4890 {
4891         int ret = 0;
4892         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4893
4894         if (mem->move_charge_at_immigrate) {
4895                 struct mm_struct *mm;
4896                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4897
4898                 VM_BUG_ON(from == mem);
4899
4900                 mm = get_task_mm(p);
4901                 if (!mm)
4902                         return 0;
4903                 /* We move charges only when we move a owner of the mm */
4904                 if (mm->owner == p) {
4905                         VM_BUG_ON(mc.from);
4906                         VM_BUG_ON(mc.to);
4907                         VM_BUG_ON(mc.precharge);
4908                         VM_BUG_ON(mc.moved_charge);
4909                         VM_BUG_ON(mc.moved_swap);
4910                         mem_cgroup_start_move(from);
4911                         spin_lock(&mc.lock);
4912                         mc.from = from;
4913                         mc.to = mem;
4914                         spin_unlock(&mc.lock);
4915                         /* We set mc.moving_task later */
4916
4917                         ret = mem_cgroup_precharge_mc(mm);
4918                         if (ret)
4919                                 mem_cgroup_clear_mc();
4920                 }
4921                 mmput(mm);
4922         }
4923         return ret;
4924 }
4925
4926 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4927                                 struct cgroup *cgroup,
4928                                 struct task_struct *p,
4929                                 bool threadgroup)
4930 {
4931         mem_cgroup_clear_mc();
4932 }
4933
4934 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4935                                 unsigned long addr, unsigned long end,
4936                                 struct mm_walk *walk)
4937 {
4938         int ret = 0;
4939         struct vm_area_struct *vma = walk->private;
4940         pte_t *pte;
4941         spinlock_t *ptl;
4942
4943         split_huge_page_pmd(walk->mm, pmd);
4944 retry:
4945         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4946         for (; addr != end; addr += PAGE_SIZE) {
4947                 pte_t ptent = *(pte++);
4948                 union mc_target target;
4949                 int type;
4950                 struct page *page;
4951                 struct page_cgroup *pc;
4952                 swp_entry_t ent;
4953
4954                 if (!mc.precharge)
4955                         break;
4956
4957                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4958                 switch (type) {
4959                 case MC_TARGET_PAGE:
4960                         page = target.page;
4961                         if (isolate_lru_page(page))
4962                                 goto put;
4963                         pc = lookup_page_cgroup(page);
4964                         if (!mem_cgroup_move_account(page, pc,
4965                                         mc.from, mc.to, false, PAGE_SIZE)) {
4966                                 mc.precharge--;
4967                                 /* we uncharge from mc.from later. */
4968                                 mc.moved_charge++;
4969                         }
4970                         putback_lru_page(page);
4971 put:                    /* is_target_pte_for_mc() gets the page */
4972                         put_page(page);
4973                         break;
4974                 case MC_TARGET_SWAP:
4975                         ent = target.ent;
4976                         if (!mem_cgroup_move_swap_account(ent,
4977                                                 mc.from, mc.to, false)) {
4978                                 mc.precharge--;
4979                                 /* we fixup refcnts and charges later. */
4980                                 mc.moved_swap++;
4981                         }
4982                         break;
4983                 default:
4984                         break;
4985                 }
4986         }
4987         pte_unmap_unlock(pte - 1, ptl);
4988         cond_resched();
4989
4990         if (addr != end) {
4991                 /*
4992                  * We have consumed all precharges we got in can_attach().
4993                  * We try charge one by one, but don't do any additional
4994                  * charges to mc.to if we have failed in charge once in attach()
4995                  * phase.
4996                  */
4997                 ret = mem_cgroup_do_precharge(1);
4998                 if (!ret)
4999                         goto retry;
5000         }
5001
5002         return ret;
5003 }
5004
5005 static void mem_cgroup_move_charge(struct mm_struct *mm)
5006 {
5007         struct vm_area_struct *vma;
5008
5009         lru_add_drain_all();
5010 retry:
5011         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5012                 /*
5013                  * Someone who are holding the mmap_sem might be waiting in
5014                  * waitq. So we cancel all extra charges, wake up all waiters,
5015                  * and retry. Because we cancel precharges, we might not be able
5016                  * to move enough charges, but moving charge is a best-effort
5017                  * feature anyway, so it wouldn't be a big problem.
5018                  */
5019                 __mem_cgroup_clear_mc();
5020                 cond_resched();
5021                 goto retry;
5022         }
5023         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5024                 int ret;
5025                 struct mm_walk mem_cgroup_move_charge_walk = {
5026                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5027                         .mm = mm,
5028                         .private = vma,
5029                 };
5030                 if (is_vm_hugetlb_page(vma))
5031                         continue;
5032                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5033                                                 &mem_cgroup_move_charge_walk);
5034                 if (ret)
5035                         /*
5036                          * means we have consumed all precharges and failed in
5037                          * doing additional charge. Just abandon here.
5038                          */
5039                         break;
5040         }
5041         up_read(&mm->mmap_sem);
5042 }
5043
5044 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5045                                 struct cgroup *cont,
5046                                 struct cgroup *old_cont,
5047                                 struct task_struct *p,
5048                                 bool threadgroup)
5049 {
5050         struct mm_struct *mm;
5051
5052         if (!mc.to)
5053                 /* no need to move charge */
5054                 return;
5055
5056         mm = get_task_mm(p);
5057         if (mm) {
5058                 mem_cgroup_move_charge(mm);
5059                 mmput(mm);
5060         }
5061         mem_cgroup_clear_mc();
5062 }
5063 #else   /* !CONFIG_MMU */
5064 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5065                                 struct cgroup *cgroup,
5066                                 struct task_struct *p,
5067                                 bool threadgroup)
5068 {
5069         return 0;
5070 }
5071 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5072                                 struct cgroup *cgroup,
5073                                 struct task_struct *p,
5074                                 bool threadgroup)
5075 {
5076 }
5077 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5078                                 struct cgroup *cont,
5079                                 struct cgroup *old_cont,
5080                                 struct task_struct *p,
5081                                 bool threadgroup)
5082 {
5083 }
5084 #endif
5085
5086 struct cgroup_subsys mem_cgroup_subsys = {
5087         .name = "memory",
5088         .subsys_id = mem_cgroup_subsys_id,
5089         .create = mem_cgroup_create,
5090         .pre_destroy = mem_cgroup_pre_destroy,
5091         .destroy = mem_cgroup_destroy,
5092         .populate = mem_cgroup_populate,
5093         .can_attach = mem_cgroup_can_attach,
5094         .cancel_attach = mem_cgroup_cancel_attach,
5095         .attach = mem_cgroup_move_task,
5096         .early_init = 0,
5097         .use_id = 1,
5098 };
5099
5100 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5101 static int __init enable_swap_account(char *s)
5102 {
5103         /* consider enabled if no parameter or 1 is given */
5104         if (!(*s) || !strcmp(s, "=1"))
5105                 really_do_swap_account = 1;
5106         else if (!strcmp(s, "=0"))
5107                 really_do_swap_account = 0;
5108         return 1;
5109 }
5110 __setup("swapaccount", enable_swap_account);
5111
5112 static int __init disable_swap_account(char *s)
5113 {
5114         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5115         enable_swap_account("=0");
5116         return 1;
5117 }
5118 __setup("noswapaccount", disable_swap_account);
5119 #endif