e286e1603e4f388365506818b016cb2e3049da04
[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         if (!mz)
1047                 return NULL;
1048
1049         return &mz->reclaim_stat;
1050 }
1051
1052 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1053                                         struct list_head *dst,
1054                                         unsigned long *scanned, int order,
1055                                         int mode, struct zone *z,
1056                                         struct mem_cgroup *mem_cont,
1057                                         int active, int file)
1058 {
1059         unsigned long nr_taken = 0;
1060         struct page *page;
1061         unsigned long scan;
1062         LIST_HEAD(pc_list);
1063         struct list_head *src;
1064         struct page_cgroup *pc, *tmp;
1065         int nid = zone_to_nid(z);
1066         int zid = zone_idx(z);
1067         struct mem_cgroup_per_zone *mz;
1068         int lru = LRU_FILE * file + active;
1069         int ret;
1070
1071         BUG_ON(!mem_cont);
1072         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1073         src = &mz->lists[lru];
1074
1075         scan = 0;
1076         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1077                 if (scan >= nr_to_scan)
1078                         break;
1079
1080                 if (unlikely(!PageCgroupUsed(pc)))
1081                         continue;
1082
1083                 page = pc->page;
1084
1085                 if (unlikely(!PageLRU(page)))
1086                         continue;
1087
1088                 scan++;
1089                 ret = __isolate_lru_page(page, mode, file);
1090                 switch (ret) {
1091                 case 0:
1092                         list_move(&page->lru, dst);
1093                         mem_cgroup_del_lru(page);
1094                         nr_taken += hpage_nr_pages(page);
1095                         break;
1096                 case -EBUSY:
1097                         /* we don't affect global LRU but rotate in our LRU */
1098                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1099                         break;
1100                 default:
1101                         break;
1102                 }
1103         }
1104
1105         *scanned = scan;
1106
1107         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1108                                       0, 0, 0, mode);
1109
1110         return nr_taken;
1111 }
1112
1113 #define mem_cgroup_from_res_counter(counter, member)    \
1114         container_of(counter, struct mem_cgroup, member)
1115
1116 /**
1117  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1118  * @mem: the memory cgroup
1119  *
1120  * Returns the maximum amount of memory @mem can be charged with, in
1121  * bytes.
1122  */
1123 static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem)
1124 {
1125         unsigned long long margin;
1126
1127         margin = res_counter_margin(&mem->res);
1128         if (do_swap_account)
1129                 margin = min(margin, res_counter_margin(&mem->memsw));
1130         return margin;
1131 }
1132
1133 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1134 {
1135         struct cgroup *cgrp = memcg->css.cgroup;
1136         unsigned int swappiness;
1137
1138         /* root ? */
1139         if (cgrp->parent == NULL)
1140                 return vm_swappiness;
1141
1142         spin_lock(&memcg->reclaim_param_lock);
1143         swappiness = memcg->swappiness;
1144         spin_unlock(&memcg->reclaim_param_lock);
1145
1146         return swappiness;
1147 }
1148
1149 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1150 {
1151         int cpu;
1152
1153         get_online_cpus();
1154         spin_lock(&mem->pcp_counter_lock);
1155         for_each_online_cpu(cpu)
1156                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1157         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1158         spin_unlock(&mem->pcp_counter_lock);
1159         put_online_cpus();
1160
1161         synchronize_rcu();
1162 }
1163
1164 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1165 {
1166         int cpu;
1167
1168         if (!mem)
1169                 return;
1170         get_online_cpus();
1171         spin_lock(&mem->pcp_counter_lock);
1172         for_each_online_cpu(cpu)
1173                 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1174         mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1175         spin_unlock(&mem->pcp_counter_lock);
1176         put_online_cpus();
1177 }
1178 /*
1179  * 2 routines for checking "mem" is under move_account() or not.
1180  *
1181  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1182  *                        for avoiding race in accounting. If true,
1183  *                        pc->mem_cgroup may be overwritten.
1184  *
1185  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1186  *                        under hierarchy of moving cgroups. This is for
1187  *                        waiting at hith-memory prressure caused by "move".
1188  */
1189
1190 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1191 {
1192         VM_BUG_ON(!rcu_read_lock_held());
1193         return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1194 }
1195
1196 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1197 {
1198         struct mem_cgroup *from;
1199         struct mem_cgroup *to;
1200         bool ret = false;
1201         /*
1202          * Unlike task_move routines, we access mc.to, mc.from not under
1203          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1204          */
1205         spin_lock(&mc.lock);
1206         from = mc.from;
1207         to = mc.to;
1208         if (!from)
1209                 goto unlock;
1210         if (from == mem || to == mem
1211             || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1212             || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1213                 ret = true;
1214 unlock:
1215         spin_unlock(&mc.lock);
1216         return ret;
1217 }
1218
1219 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1220 {
1221         if (mc.moving_task && current != mc.moving_task) {
1222                 if (mem_cgroup_under_move(mem)) {
1223                         DEFINE_WAIT(wait);
1224                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1225                         /* moving charge context might have finished. */
1226                         if (mc.moving_task)
1227                                 schedule();
1228                         finish_wait(&mc.waitq, &wait);
1229                         return true;
1230                 }
1231         }
1232         return false;
1233 }
1234
1235 /**
1236  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1237  * @memcg: The memory cgroup that went over limit
1238  * @p: Task that is going to be killed
1239  *
1240  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1241  * enabled
1242  */
1243 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1244 {
1245         struct cgroup *task_cgrp;
1246         struct cgroup *mem_cgrp;
1247         /*
1248          * Need a buffer in BSS, can't rely on allocations. The code relies
1249          * on the assumption that OOM is serialized for memory controller.
1250          * If this assumption is broken, revisit this code.
1251          */
1252         static char memcg_name[PATH_MAX];
1253         int ret;
1254
1255         if (!memcg || !p)
1256                 return;
1257
1258
1259         rcu_read_lock();
1260
1261         mem_cgrp = memcg->css.cgroup;
1262         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1263
1264         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1265         if (ret < 0) {
1266                 /*
1267                  * Unfortunately, we are unable to convert to a useful name
1268                  * But we'll still print out the usage information
1269                  */
1270                 rcu_read_unlock();
1271                 goto done;
1272         }
1273         rcu_read_unlock();
1274
1275         printk(KERN_INFO "Task in %s killed", memcg_name);
1276
1277         rcu_read_lock();
1278         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1279         if (ret < 0) {
1280                 rcu_read_unlock();
1281                 goto done;
1282         }
1283         rcu_read_unlock();
1284
1285         /*
1286          * Continues from above, so we don't need an KERN_ level
1287          */
1288         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1289 done:
1290
1291         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1292                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1293                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1294                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1295         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1296                 "failcnt %llu\n",
1297                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1298                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1299                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1300 }
1301
1302 /*
1303  * This function returns the number of memcg under hierarchy tree. Returns
1304  * 1(self count) if no children.
1305  */
1306 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1307 {
1308         int num = 0;
1309         struct mem_cgroup *iter;
1310
1311         for_each_mem_cgroup_tree(iter, mem)
1312                 num++;
1313         return num;
1314 }
1315
1316 /*
1317  * Return the memory (and swap, if configured) limit for a memcg.
1318  */
1319 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1320 {
1321         u64 limit;
1322         u64 memsw;
1323
1324         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1325         limit += total_swap_pages << PAGE_SHIFT;
1326
1327         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1328         /*
1329          * If memsw is finite and limits the amount of swap space available
1330          * to this memcg, return that limit.
1331          */
1332         return min(limit, memsw);
1333 }
1334
1335 /*
1336  * Visit the first child (need not be the first child as per the ordering
1337  * of the cgroup list, since we track last_scanned_child) of @mem and use
1338  * that to reclaim free pages from.
1339  */
1340 static struct mem_cgroup *
1341 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1342 {
1343         struct mem_cgroup *ret = NULL;
1344         struct cgroup_subsys_state *css;
1345         int nextid, found;
1346
1347         if (!root_mem->use_hierarchy) {
1348                 css_get(&root_mem->css);
1349                 ret = root_mem;
1350         }
1351
1352         while (!ret) {
1353                 rcu_read_lock();
1354                 nextid = root_mem->last_scanned_child + 1;
1355                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1356                                    &found);
1357                 if (css && css_tryget(css))
1358                         ret = container_of(css, struct mem_cgroup, css);
1359
1360                 rcu_read_unlock();
1361                 /* Updates scanning parameter */
1362                 spin_lock(&root_mem->reclaim_param_lock);
1363                 if (!css) {
1364                         /* this means start scan from ID:1 */
1365                         root_mem->last_scanned_child = 0;
1366                 } else
1367                         root_mem->last_scanned_child = found;
1368                 spin_unlock(&root_mem->reclaim_param_lock);
1369         }
1370
1371         return ret;
1372 }
1373
1374 /*
1375  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1376  * we reclaimed from, so that we don't end up penalizing one child extensively
1377  * based on its position in the children list.
1378  *
1379  * root_mem is the original ancestor that we've been reclaim from.
1380  *
1381  * We give up and return to the caller when we visit root_mem twice.
1382  * (other groups can be removed while we're walking....)
1383  *
1384  * If shrink==true, for avoiding to free too much, this returns immedieately.
1385  */
1386 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1387                                                 struct zone *zone,
1388                                                 gfp_t gfp_mask,
1389                                                 unsigned long reclaim_options)
1390 {
1391         struct mem_cgroup *victim;
1392         int ret, total = 0;
1393         int loop = 0;
1394         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1395         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1396         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1397         unsigned long excess;
1398
1399         excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1400
1401         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1402         if (root_mem->memsw_is_minimum)
1403                 noswap = true;
1404
1405         while (1) {
1406                 victim = mem_cgroup_select_victim(root_mem);
1407                 if (victim == root_mem) {
1408                         loop++;
1409                         if (loop >= 1)
1410                                 drain_all_stock_async();
1411                         if (loop >= 2) {
1412                                 /*
1413                                  * If we have not been able to reclaim
1414                                  * anything, it might because there are
1415                                  * no reclaimable pages under this hierarchy
1416                                  */
1417                                 if (!check_soft || !total) {
1418                                         css_put(&victim->css);
1419                                         break;
1420                                 }
1421                                 /*
1422                                  * We want to do more targetted reclaim.
1423                                  * excess >> 2 is not to excessive so as to
1424                                  * reclaim too much, nor too less that we keep
1425                                  * coming back to reclaim from this cgroup
1426                                  */
1427                                 if (total >= (excess >> 2) ||
1428                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1429                                         css_put(&victim->css);
1430                                         break;
1431                                 }
1432                         }
1433                 }
1434                 if (!mem_cgroup_local_usage(victim)) {
1435                         /* this cgroup's local usage == 0 */
1436                         css_put(&victim->css);
1437                         continue;
1438                 }
1439                 /* we use swappiness of local cgroup */
1440                 if (check_soft)
1441                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1442                                 noswap, get_swappiness(victim), zone);
1443                 else
1444                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1445                                                 noswap, get_swappiness(victim));
1446                 css_put(&victim->css);
1447                 /*
1448                  * At shrinking usage, we can't check we should stop here or
1449                  * reclaim more. It's depends on callers. last_scanned_child
1450                  * will work enough for keeping fairness under tree.
1451                  */
1452                 if (shrink)
1453                         return ret;
1454                 total += ret;
1455                 if (check_soft) {
1456                         if (!res_counter_soft_limit_excess(&root_mem->res))
1457                                 return total;
1458                 } else if (mem_cgroup_margin(root_mem))
1459                         return 1 + total;
1460         }
1461         return total;
1462 }
1463
1464 /*
1465  * Check OOM-Killer is already running under our hierarchy.
1466  * If someone is running, return false.
1467  */
1468 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1469 {
1470         int x, lock_count = 0;
1471         struct mem_cgroup *iter;
1472
1473         for_each_mem_cgroup_tree(iter, mem) {
1474                 x = atomic_inc_return(&iter->oom_lock);
1475                 lock_count = max(x, lock_count);
1476         }
1477
1478         if (lock_count == 1)
1479                 return true;
1480         return false;
1481 }
1482
1483 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1484 {
1485         struct mem_cgroup *iter;
1486
1487         /*
1488          * When a new child is created while the hierarchy is under oom,
1489          * mem_cgroup_oom_lock() may not be called. We have to use
1490          * atomic_add_unless() here.
1491          */
1492         for_each_mem_cgroup_tree(iter, mem)
1493                 atomic_add_unless(&iter->oom_lock, -1, 0);
1494         return 0;
1495 }
1496
1497
1498 static DEFINE_MUTEX(memcg_oom_mutex);
1499 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1500
1501 struct oom_wait_info {
1502         struct mem_cgroup *mem;
1503         wait_queue_t    wait;
1504 };
1505
1506 static int memcg_oom_wake_function(wait_queue_t *wait,
1507         unsigned mode, int sync, void *arg)
1508 {
1509         struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1510         struct oom_wait_info *oom_wait_info;
1511
1512         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1513
1514         if (oom_wait_info->mem == wake_mem)
1515                 goto wakeup;
1516         /* if no hierarchy, no match */
1517         if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1518                 return 0;
1519         /*
1520          * Both of oom_wait_info->mem and wake_mem are stable under us.
1521          * Then we can use css_is_ancestor without taking care of RCU.
1522          */
1523         if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1524             !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1525                 return 0;
1526
1527 wakeup:
1528         return autoremove_wake_function(wait, mode, sync, arg);
1529 }
1530
1531 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1532 {
1533         /* for filtering, pass "mem" as argument. */
1534         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1535 }
1536
1537 static void memcg_oom_recover(struct mem_cgroup *mem)
1538 {
1539         if (mem && atomic_read(&mem->oom_lock))
1540                 memcg_wakeup_oom(mem);
1541 }
1542
1543 /*
1544  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1545  */
1546 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1547 {
1548         struct oom_wait_info owait;
1549         bool locked, need_to_kill;
1550
1551         owait.mem = mem;
1552         owait.wait.flags = 0;
1553         owait.wait.func = memcg_oom_wake_function;
1554         owait.wait.private = current;
1555         INIT_LIST_HEAD(&owait.wait.task_list);
1556         need_to_kill = true;
1557         /* At first, try to OOM lock hierarchy under mem.*/
1558         mutex_lock(&memcg_oom_mutex);
1559         locked = mem_cgroup_oom_lock(mem);
1560         /*
1561          * Even if signal_pending(), we can't quit charge() loop without
1562          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1563          * under OOM is always welcomed, use TASK_KILLABLE here.
1564          */
1565         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1566         if (!locked || mem->oom_kill_disable)
1567                 need_to_kill = false;
1568         if (locked)
1569                 mem_cgroup_oom_notify(mem);
1570         mutex_unlock(&memcg_oom_mutex);
1571
1572         if (need_to_kill) {
1573                 finish_wait(&memcg_oom_waitq, &owait.wait);
1574                 mem_cgroup_out_of_memory(mem, mask);
1575         } else {
1576                 schedule();
1577                 finish_wait(&memcg_oom_waitq, &owait.wait);
1578         }
1579         mutex_lock(&memcg_oom_mutex);
1580         mem_cgroup_oom_unlock(mem);
1581         memcg_wakeup_oom(mem);
1582         mutex_unlock(&memcg_oom_mutex);
1583
1584         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1585                 return false;
1586         /* Give chance to dying process */
1587         schedule_timeout(1);
1588         return true;
1589 }
1590
1591 /*
1592  * Currently used to update mapped file statistics, but the routine can be
1593  * generalized to update other statistics as well.
1594  *
1595  * Notes: Race condition
1596  *
1597  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1598  * it tends to be costly. But considering some conditions, we doesn't need
1599  * to do so _always_.
1600  *
1601  * Considering "charge", lock_page_cgroup() is not required because all
1602  * file-stat operations happen after a page is attached to radix-tree. There
1603  * are no race with "charge".
1604  *
1605  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1606  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1607  * if there are race with "uncharge". Statistics itself is properly handled
1608  * by flags.
1609  *
1610  * Considering "move", this is an only case we see a race. To make the race
1611  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1612  * possibility of race condition. If there is, we take a lock.
1613  */
1614
1615 void mem_cgroup_update_page_stat(struct page *page,
1616                                  enum mem_cgroup_page_stat_item idx, int val)
1617 {
1618         struct mem_cgroup *mem;
1619         struct page_cgroup *pc = lookup_page_cgroup(page);
1620         bool need_unlock = false;
1621         unsigned long uninitialized_var(flags);
1622
1623         if (unlikely(!pc))
1624                 return;
1625
1626         rcu_read_lock();
1627         mem = pc->mem_cgroup;
1628         if (unlikely(!mem || !PageCgroupUsed(pc)))
1629                 goto out;
1630         /* pc->mem_cgroup is unstable ? */
1631         if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1632                 /* take a lock against to access pc->mem_cgroup */
1633                 move_lock_page_cgroup(pc, &flags);
1634                 need_unlock = true;
1635                 mem = pc->mem_cgroup;
1636                 if (!mem || !PageCgroupUsed(pc))
1637                         goto out;
1638         }
1639
1640         switch (idx) {
1641         case MEMCG_NR_FILE_MAPPED:
1642                 if (val > 0)
1643                         SetPageCgroupFileMapped(pc);
1644                 else if (!page_mapped(page))
1645                         ClearPageCgroupFileMapped(pc);
1646                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1647                 break;
1648         default:
1649                 BUG();
1650         }
1651
1652         this_cpu_add(mem->stat->count[idx], val);
1653
1654 out:
1655         if (unlikely(need_unlock))
1656                 move_unlock_page_cgroup(pc, &flags);
1657         rcu_read_unlock();
1658         return;
1659 }
1660 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1661
1662 /*
1663  * size of first charge trial. "32" comes from vmscan.c's magic value.
1664  * TODO: maybe necessary to use big numbers in big irons.
1665  */
1666 #define CHARGE_SIZE     (32 * PAGE_SIZE)
1667 struct memcg_stock_pcp {
1668         struct mem_cgroup *cached; /* this never be root cgroup */
1669         int charge;
1670         struct work_struct work;
1671 };
1672 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1673 static atomic_t memcg_drain_count;
1674
1675 /*
1676  * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1677  * from local stock and true is returned. If the stock is 0 or charges from a
1678  * cgroup which is not current target, returns false. This stock will be
1679  * refilled.
1680  */
1681 static bool consume_stock(struct mem_cgroup *mem)
1682 {
1683         struct memcg_stock_pcp *stock;
1684         bool ret = true;
1685
1686         stock = &get_cpu_var(memcg_stock);
1687         if (mem == stock->cached && stock->charge)
1688                 stock->charge -= PAGE_SIZE;
1689         else /* need to call res_counter_charge */
1690                 ret = false;
1691         put_cpu_var(memcg_stock);
1692         return ret;
1693 }
1694
1695 /*
1696  * Returns stocks cached in percpu to res_counter and reset cached information.
1697  */
1698 static void drain_stock(struct memcg_stock_pcp *stock)
1699 {
1700         struct mem_cgroup *old = stock->cached;
1701
1702         if (stock->charge) {
1703                 res_counter_uncharge(&old->res, stock->charge);
1704                 if (do_swap_account)
1705                         res_counter_uncharge(&old->memsw, stock->charge);
1706         }
1707         stock->cached = NULL;
1708         stock->charge = 0;
1709 }
1710
1711 /*
1712  * This must be called under preempt disabled or must be called by
1713  * a thread which is pinned to local cpu.
1714  */
1715 static void drain_local_stock(struct work_struct *dummy)
1716 {
1717         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1718         drain_stock(stock);
1719 }
1720
1721 /*
1722  * Cache charges(val) which is from res_counter, to local per_cpu area.
1723  * This will be consumed by consume_stock() function, later.
1724  */
1725 static void refill_stock(struct mem_cgroup *mem, int val)
1726 {
1727         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1728
1729         if (stock->cached != mem) { /* reset if necessary */
1730                 drain_stock(stock);
1731                 stock->cached = mem;
1732         }
1733         stock->charge += val;
1734         put_cpu_var(memcg_stock);
1735 }
1736
1737 /*
1738  * Tries to drain stocked charges in other cpus. This function is asynchronous
1739  * and just put a work per cpu for draining localy on each cpu. Caller can
1740  * expects some charges will be back to res_counter later but cannot wait for
1741  * it.
1742  */
1743 static void drain_all_stock_async(void)
1744 {
1745         int cpu;
1746         /* This function is for scheduling "drain" in asynchronous way.
1747          * The result of "drain" is not directly handled by callers. Then,
1748          * if someone is calling drain, we don't have to call drain more.
1749          * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1750          * there is a race. We just do loose check here.
1751          */
1752         if (atomic_read(&memcg_drain_count))
1753                 return;
1754         /* Notify other cpus that system-wide "drain" is running */
1755         atomic_inc(&memcg_drain_count);
1756         get_online_cpus();
1757         for_each_online_cpu(cpu) {
1758                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1759                 schedule_work_on(cpu, &stock->work);
1760         }
1761         put_online_cpus();
1762         atomic_dec(&memcg_drain_count);
1763         /* We don't wait for flush_work */
1764 }
1765
1766 /* This is a synchronous drain interface. */
1767 static void drain_all_stock_sync(void)
1768 {
1769         /* called when force_empty is called */
1770         atomic_inc(&memcg_drain_count);
1771         schedule_on_each_cpu(drain_local_stock);
1772         atomic_dec(&memcg_drain_count);
1773 }
1774
1775 /*
1776  * This function drains percpu counter value from DEAD cpu and
1777  * move it to local cpu. Note that this function can be preempted.
1778  */
1779 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1780 {
1781         int i;
1782
1783         spin_lock(&mem->pcp_counter_lock);
1784         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1785                 s64 x = per_cpu(mem->stat->count[i], cpu);
1786
1787                 per_cpu(mem->stat->count[i], cpu) = 0;
1788                 mem->nocpu_base.count[i] += x;
1789         }
1790         /* need to clear ON_MOVE value, works as a kind of lock. */
1791         per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1792         spin_unlock(&mem->pcp_counter_lock);
1793 }
1794
1795 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1796 {
1797         int idx = MEM_CGROUP_ON_MOVE;
1798
1799         spin_lock(&mem->pcp_counter_lock);
1800         per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1801         spin_unlock(&mem->pcp_counter_lock);
1802 }
1803
1804 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1805                                         unsigned long action,
1806                                         void *hcpu)
1807 {
1808         int cpu = (unsigned long)hcpu;
1809         struct memcg_stock_pcp *stock;
1810         struct mem_cgroup *iter;
1811
1812         if ((action == CPU_ONLINE)) {
1813                 for_each_mem_cgroup_all(iter)
1814                         synchronize_mem_cgroup_on_move(iter, cpu);
1815                 return NOTIFY_OK;
1816         }
1817
1818         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1819                 return NOTIFY_OK;
1820
1821         for_each_mem_cgroup_all(iter)
1822                 mem_cgroup_drain_pcp_counter(iter, cpu);
1823
1824         stock = &per_cpu(memcg_stock, cpu);
1825         drain_stock(stock);
1826         return NOTIFY_OK;
1827 }
1828
1829
1830 /* See __mem_cgroup_try_charge() for details */
1831 enum {
1832         CHARGE_OK,              /* success */
1833         CHARGE_RETRY,           /* need to retry but retry is not bad */
1834         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
1835         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
1836         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
1837 };
1838
1839 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1840                                 int csize, bool oom_check)
1841 {
1842         struct mem_cgroup *mem_over_limit;
1843         struct res_counter *fail_res;
1844         unsigned long flags = 0;
1845         int ret;
1846
1847         ret = res_counter_charge(&mem->res, csize, &fail_res);
1848
1849         if (likely(!ret)) {
1850                 if (!do_swap_account)
1851                         return CHARGE_OK;
1852                 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1853                 if (likely(!ret))
1854                         return CHARGE_OK;
1855
1856                 res_counter_uncharge(&mem->res, csize);
1857                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1858                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1859         } else
1860                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1861         /*
1862          * csize can be either a huge page (HPAGE_SIZE), a batch of
1863          * regular pages (CHARGE_SIZE), or a single regular page
1864          * (PAGE_SIZE).
1865          *
1866          * Never reclaim on behalf of optional batching, retry with a
1867          * single page instead.
1868          */
1869         if (csize == CHARGE_SIZE)
1870                 return CHARGE_RETRY;
1871
1872         if (!(gfp_mask & __GFP_WAIT))
1873                 return CHARGE_WOULDBLOCK;
1874
1875         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1876                                               gfp_mask, flags);
1877         if (mem_cgroup_margin(mem_over_limit) >= csize)
1878                 return CHARGE_RETRY;
1879         /*
1880          * Even though the limit is exceeded at this point, reclaim
1881          * may have been able to free some pages.  Retry the charge
1882          * before killing the task.
1883          *
1884          * Only for regular pages, though: huge pages are rather
1885          * unlikely to succeed so close to the limit, and we fall back
1886          * to regular pages anyway in case of failure.
1887          */
1888         if (csize == PAGE_SIZE && ret)
1889                 return CHARGE_RETRY;
1890
1891         /*
1892          * At task move, charge accounts can be doubly counted. So, it's
1893          * better to wait until the end of task_move if something is going on.
1894          */
1895         if (mem_cgroup_wait_acct_move(mem_over_limit))
1896                 return CHARGE_RETRY;
1897
1898         /* If we don't need to call oom-killer at el, return immediately */
1899         if (!oom_check)
1900                 return CHARGE_NOMEM;
1901         /* check OOM */
1902         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1903                 return CHARGE_OOM_DIE;
1904
1905         return CHARGE_RETRY;
1906 }
1907
1908 /*
1909  * Unlike exported interface, "oom" parameter is added. if oom==true,
1910  * oom-killer can be invoked.
1911  */
1912 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1913                                    gfp_t gfp_mask,
1914                                    struct mem_cgroup **memcg, bool oom,
1915                                    int page_size)
1916 {
1917         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1918         struct mem_cgroup *mem = NULL;
1919         int ret;
1920         int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1921
1922         /*
1923          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1924          * in system level. So, allow to go ahead dying process in addition to
1925          * MEMDIE process.
1926          */
1927         if (unlikely(test_thread_flag(TIF_MEMDIE)
1928                      || fatal_signal_pending(current)))
1929                 goto bypass;
1930
1931         /*
1932          * We always charge the cgroup the mm_struct belongs to.
1933          * The mm_struct's mem_cgroup changes on task migration if the
1934          * thread group leader migrates. It's possible that mm is not
1935          * set, if so charge the init_mm (happens for pagecache usage).
1936          */
1937         if (!*memcg && !mm)
1938                 goto bypass;
1939 again:
1940         if (*memcg) { /* css should be a valid one */
1941                 mem = *memcg;
1942                 VM_BUG_ON(css_is_removed(&mem->css));
1943                 if (mem_cgroup_is_root(mem))
1944                         goto done;
1945                 if (page_size == PAGE_SIZE && consume_stock(mem))
1946                         goto done;
1947                 css_get(&mem->css);
1948         } else {
1949                 struct task_struct *p;
1950
1951                 rcu_read_lock();
1952                 p = rcu_dereference(mm->owner);
1953                 /*
1954                  * Because we don't have task_lock(), "p" can exit.
1955                  * In that case, "mem" can point to root or p can be NULL with
1956                  * race with swapoff. Then, we have small risk of mis-accouning.
1957                  * But such kind of mis-account by race always happens because
1958                  * we don't have cgroup_mutex(). It's overkill and we allo that
1959                  * small race, here.
1960                  * (*) swapoff at el will charge against mm-struct not against
1961                  * task-struct. So, mm->owner can be NULL.
1962                  */
1963                 mem = mem_cgroup_from_task(p);
1964                 if (!mem || mem_cgroup_is_root(mem)) {
1965                         rcu_read_unlock();
1966                         goto done;
1967                 }
1968                 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1969                         /*
1970                          * It seems dagerous to access memcg without css_get().
1971                          * But considering how consume_stok works, it's not
1972                          * necessary. If consume_stock success, some charges
1973                          * from this memcg are cached on this cpu. So, we
1974                          * don't need to call css_get()/css_tryget() before
1975                          * calling consume_stock().
1976                          */
1977                         rcu_read_unlock();
1978                         goto done;
1979                 }
1980                 /* after here, we may be blocked. we need to get refcnt */
1981                 if (!css_tryget(&mem->css)) {
1982                         rcu_read_unlock();
1983                         goto again;
1984                 }
1985                 rcu_read_unlock();
1986         }
1987
1988         do {
1989                 bool oom_check;
1990
1991                 /* If killed, bypass charge */
1992                 if (fatal_signal_pending(current)) {
1993                         css_put(&mem->css);
1994                         goto bypass;
1995                 }
1996
1997                 oom_check = false;
1998                 if (oom && !nr_oom_retries) {
1999                         oom_check = true;
2000                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2001                 }
2002
2003                 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
2004
2005                 switch (ret) {
2006                 case CHARGE_OK:
2007                         break;
2008                 case CHARGE_RETRY: /* not in OOM situation but retry */
2009                         csize = page_size;
2010                         css_put(&mem->css);
2011                         mem = NULL;
2012                         goto again;
2013                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2014                         css_put(&mem->css);
2015                         goto nomem;
2016                 case CHARGE_NOMEM: /* OOM routine works */
2017                         if (!oom) {
2018                                 css_put(&mem->css);
2019                                 goto nomem;
2020                         }
2021                         /* If oom, we never return -ENOMEM */
2022                         nr_oom_retries--;
2023                         break;
2024                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2025                         css_put(&mem->css);
2026                         goto bypass;
2027                 }
2028         } while (ret != CHARGE_OK);
2029
2030         if (csize > page_size)
2031                 refill_stock(mem, csize - page_size);
2032         css_put(&mem->css);
2033 done:
2034         *memcg = mem;
2035         return 0;
2036 nomem:
2037         *memcg = NULL;
2038         return -ENOMEM;
2039 bypass:
2040         *memcg = NULL;
2041         return 0;
2042 }
2043
2044 /*
2045  * Somemtimes we have to undo a charge we got by try_charge().
2046  * This function is for that and do uncharge, put css's refcnt.
2047  * gotten by try_charge().
2048  */
2049 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2050                                                         unsigned long count)
2051 {
2052         if (!mem_cgroup_is_root(mem)) {
2053                 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2054                 if (do_swap_account)
2055                         res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2056         }
2057 }
2058
2059 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2060                                      int page_size)
2061 {
2062         __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2063 }
2064
2065 /*
2066  * A helper function to get mem_cgroup from ID. must be called under
2067  * rcu_read_lock(). The caller must check css_is_removed() or some if
2068  * it's concern. (dropping refcnt from swap can be called against removed
2069  * memcg.)
2070  */
2071 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2072 {
2073         struct cgroup_subsys_state *css;
2074
2075         /* ID 0 is unused ID */
2076         if (!id)
2077                 return NULL;
2078         css = css_lookup(&mem_cgroup_subsys, id);
2079         if (!css)
2080                 return NULL;
2081         return container_of(css, struct mem_cgroup, css);
2082 }
2083
2084 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2085 {
2086         struct mem_cgroup *mem = NULL;
2087         struct page_cgroup *pc;
2088         unsigned short id;
2089         swp_entry_t ent;
2090
2091         VM_BUG_ON(!PageLocked(page));
2092
2093         pc = lookup_page_cgroup(page);
2094         lock_page_cgroup(pc);
2095         if (PageCgroupUsed(pc)) {
2096                 mem = pc->mem_cgroup;
2097                 if (mem && !css_tryget(&mem->css))
2098                         mem = NULL;
2099         } else if (PageSwapCache(page)) {
2100                 ent.val = page_private(page);
2101                 id = lookup_swap_cgroup(ent);
2102                 rcu_read_lock();
2103                 mem = mem_cgroup_lookup(id);
2104                 if (mem && !css_tryget(&mem->css))
2105                         mem = NULL;
2106                 rcu_read_unlock();
2107         }
2108         unlock_page_cgroup(pc);
2109         return mem;
2110 }
2111
2112 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2113                                        struct page *page,
2114                                        struct page_cgroup *pc,
2115                                        enum charge_type ctype,
2116                                        int page_size)
2117 {
2118         int nr_pages = page_size >> PAGE_SHIFT;
2119
2120         lock_page_cgroup(pc);
2121         if (unlikely(PageCgroupUsed(pc))) {
2122                 unlock_page_cgroup(pc);
2123                 mem_cgroup_cancel_charge(mem, page_size);
2124                 return;
2125         }
2126         /*
2127          * we don't need page_cgroup_lock about tail pages, becase they are not
2128          * accessed by any other context at this point.
2129          */
2130         pc->mem_cgroup = mem;
2131         /*
2132          * We access a page_cgroup asynchronously without lock_page_cgroup().
2133          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2134          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2135          * before USED bit, we need memory barrier here.
2136          * See mem_cgroup_add_lru_list(), etc.
2137          */
2138         smp_wmb();
2139         switch (ctype) {
2140         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2141         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2142                 SetPageCgroupCache(pc);
2143                 SetPageCgroupUsed(pc);
2144                 break;
2145         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2146                 ClearPageCgroupCache(pc);
2147                 SetPageCgroupUsed(pc);
2148                 break;
2149         default:
2150                 break;
2151         }
2152
2153         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2154         unlock_page_cgroup(pc);
2155         /*
2156          * "charge_statistics" updated event counter. Then, check it.
2157          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2158          * if they exceeds softlimit.
2159          */
2160         memcg_check_events(mem, page);
2161 }
2162
2163 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2164
2165 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2166                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2167 /*
2168  * Because tail pages are not marked as "used", set it. We're under
2169  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2170  */
2171 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2172 {
2173         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2174         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2175         unsigned long flags;
2176
2177         if (mem_cgroup_disabled())
2178                 return;
2179         /*
2180          * We have no races with charge/uncharge but will have races with
2181          * page state accounting.
2182          */
2183         move_lock_page_cgroup(head_pc, &flags);
2184
2185         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2186         smp_wmb(); /* see __commit_charge() */
2187         if (PageCgroupAcctLRU(head_pc)) {
2188                 enum lru_list lru;
2189                 struct mem_cgroup_per_zone *mz;
2190
2191                 /*
2192                  * LRU flags cannot be copied because we need to add tail
2193                  *.page to LRU by generic call and our hook will be called.
2194                  * We hold lru_lock, then, reduce counter directly.
2195                  */
2196                 lru = page_lru(head);
2197                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2198                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2199         }
2200         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2201         move_unlock_page_cgroup(head_pc, &flags);
2202 }
2203 #endif
2204
2205 /**
2206  * mem_cgroup_move_account - move account of the page
2207  * @page: the page
2208  * @pc: page_cgroup of the page.
2209  * @from: mem_cgroup which the page is moved from.
2210  * @to: mem_cgroup which the page is moved to. @from != @to.
2211  * @uncharge: whether we should call uncharge and css_put against @from.
2212  * @charge_size: number of bytes to charge (regular or huge page)
2213  *
2214  * The caller must confirm following.
2215  * - page is not on LRU (isolate_page() is useful.)
2216  * - compound_lock is held when charge_size > PAGE_SIZE
2217  *
2218  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2219  * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2220  * true, this function does "uncharge" from old cgroup, but it doesn't if
2221  * @uncharge is false, so a caller should do "uncharge".
2222  */
2223 static int mem_cgroup_move_account(struct page *page, struct page_cgroup *pc,
2224                                    struct mem_cgroup *from, struct mem_cgroup *to,
2225                                    bool uncharge, int charge_size)
2226 {
2227         int nr_pages = charge_size >> PAGE_SHIFT;
2228         unsigned long flags;
2229         int ret;
2230
2231         VM_BUG_ON(from == to);
2232         VM_BUG_ON(PageLRU(page));
2233         /*
2234          * The page is isolated from LRU. So, collapse function
2235          * will not handle this page. But page splitting can happen.
2236          * Do this check under compound_page_lock(). The caller should
2237          * hold it.
2238          */
2239         ret = -EBUSY;
2240         if (charge_size > PAGE_SIZE && !PageTransHuge(page))
2241                 goto out;
2242
2243         lock_page_cgroup(pc);
2244
2245         ret = -EINVAL;
2246         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2247                 goto unlock;
2248
2249         move_lock_page_cgroup(pc, &flags);
2250
2251         if (PageCgroupFileMapped(pc)) {
2252                 /* Update mapped_file data for mem_cgroup */
2253                 preempt_disable();
2254                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2255                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2256                 preempt_enable();
2257         }
2258         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2259         if (uncharge)
2260                 /* This is not "cancel", but cancel_charge does all we need. */
2261                 mem_cgroup_cancel_charge(from, charge_size);
2262
2263         /* caller should have done css_get */
2264         pc->mem_cgroup = to;
2265         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2266         /*
2267          * We charges against "to" which may not have any tasks. Then, "to"
2268          * can be under rmdir(). But in current implementation, caller of
2269          * this function is just force_empty() and move charge, so it's
2270          * garanteed that "to" is never removed. So, we don't check rmdir
2271          * status here.
2272          */
2273         move_unlock_page_cgroup(pc, &flags);
2274         ret = 0;
2275 unlock:
2276         unlock_page_cgroup(pc);
2277         /*
2278          * check events
2279          */
2280         memcg_check_events(to, page);
2281         memcg_check_events(from, page);
2282 out:
2283         return ret;
2284 }
2285
2286 /*
2287  * move charges to its parent.
2288  */
2289
2290 static int mem_cgroup_move_parent(struct page *page,
2291                                   struct page_cgroup *pc,
2292                                   struct mem_cgroup *child,
2293                                   gfp_t gfp_mask)
2294 {
2295         struct cgroup *cg = child->css.cgroup;
2296         struct cgroup *pcg = cg->parent;
2297         struct mem_cgroup *parent;
2298         int page_size = PAGE_SIZE;
2299         unsigned long flags;
2300         int ret;
2301
2302         /* Is ROOT ? */
2303         if (!pcg)
2304                 return -EINVAL;
2305
2306         ret = -EBUSY;
2307         if (!get_page_unless_zero(page))
2308                 goto out;
2309         if (isolate_lru_page(page))
2310                 goto put;
2311
2312         if (PageTransHuge(page))
2313                 page_size = HPAGE_SIZE;
2314
2315         parent = mem_cgroup_from_cont(pcg);
2316         ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2317                                 &parent, false, page_size);
2318         if (ret || !parent)
2319                 goto put_back;
2320
2321         if (page_size > PAGE_SIZE)
2322                 flags = compound_lock_irqsave(page);
2323
2324         ret = mem_cgroup_move_account(page, pc, child, parent, true, page_size);
2325         if (ret)
2326                 mem_cgroup_cancel_charge(parent, page_size);
2327
2328         if (page_size > PAGE_SIZE)
2329                 compound_unlock_irqrestore(page, flags);
2330 put_back:
2331         putback_lru_page(page);
2332 put:
2333         put_page(page);
2334 out:
2335         return ret;
2336 }
2337
2338 /*
2339  * Charge the memory controller for page usage.
2340  * Return
2341  * 0 if the charge was successful
2342  * < 0 if the cgroup is over its limit
2343  */
2344 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2345                                 gfp_t gfp_mask, enum charge_type ctype)
2346 {
2347         struct mem_cgroup *mem = NULL;
2348         int page_size = PAGE_SIZE;
2349         struct page_cgroup *pc;
2350         bool oom = true;
2351         int ret;
2352
2353         if (PageTransHuge(page)) {
2354                 page_size <<= compound_order(page);
2355                 VM_BUG_ON(!PageTransHuge(page));
2356                 /*
2357                  * Never OOM-kill a process for a huge page.  The
2358                  * fault handler will fall back to regular pages.
2359                  */
2360                 oom = false;
2361         }
2362
2363         pc = lookup_page_cgroup(page);
2364         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2365
2366         ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2367         if (ret || !mem)
2368                 return ret;
2369
2370         __mem_cgroup_commit_charge(mem, page, pc, ctype, page_size);
2371         return 0;
2372 }
2373
2374 int mem_cgroup_newpage_charge(struct page *page,
2375                               struct mm_struct *mm, gfp_t gfp_mask)
2376 {
2377         if (mem_cgroup_disabled())
2378                 return 0;
2379         /*
2380          * If already mapped, we don't have to account.
2381          * If page cache, page->mapping has address_space.
2382          * But page->mapping may have out-of-use anon_vma pointer,
2383          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2384          * is NULL.
2385          */
2386         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2387                 return 0;
2388         if (unlikely(!mm))
2389                 mm = &init_mm;
2390         return mem_cgroup_charge_common(page, mm, gfp_mask,
2391                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2392 }
2393
2394 static void
2395 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2396                                         enum charge_type ctype);
2397
2398 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2399                                 gfp_t gfp_mask)
2400 {
2401         int ret;
2402
2403         if (mem_cgroup_disabled())
2404                 return 0;
2405         if (PageCompound(page))
2406                 return 0;
2407         /*
2408          * Corner case handling. This is called from add_to_page_cache()
2409          * in usual. But some FS (shmem) precharges this page before calling it
2410          * and call add_to_page_cache() with GFP_NOWAIT.
2411          *
2412          * For GFP_NOWAIT case, the page may be pre-charged before calling
2413          * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2414          * charge twice. (It works but has to pay a bit larger cost.)
2415          * And when the page is SwapCache, it should take swap information
2416          * into account. This is under lock_page() now.
2417          */
2418         if (!(gfp_mask & __GFP_WAIT)) {
2419                 struct page_cgroup *pc;
2420
2421                 pc = lookup_page_cgroup(page);
2422                 if (!pc)
2423                         return 0;
2424                 lock_page_cgroup(pc);
2425                 if (PageCgroupUsed(pc)) {
2426                         unlock_page_cgroup(pc);
2427                         return 0;
2428                 }
2429                 unlock_page_cgroup(pc);
2430         }
2431
2432         if (unlikely(!mm))
2433                 mm = &init_mm;
2434
2435         if (page_is_file_cache(page))
2436                 return mem_cgroup_charge_common(page, mm, gfp_mask,
2437                                 MEM_CGROUP_CHARGE_TYPE_CACHE);
2438
2439         /* shmem */
2440         if (PageSwapCache(page)) {
2441                 struct mem_cgroup *mem;
2442
2443                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2444                 if (!ret)
2445                         __mem_cgroup_commit_charge_swapin(page, mem,
2446                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2447         } else
2448                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2449                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2450
2451         return ret;
2452 }
2453
2454 /*
2455  * While swap-in, try_charge -> commit or cancel, the page is locked.
2456  * And when try_charge() successfully returns, one refcnt to memcg without
2457  * struct page_cgroup is acquired. This refcnt will be consumed by
2458  * "commit()" or removed by "cancel()"
2459  */
2460 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2461                                  struct page *page,
2462                                  gfp_t mask, struct mem_cgroup **ptr)
2463 {
2464         struct mem_cgroup *mem;
2465         int ret;
2466
2467         *ptr = NULL;
2468
2469         if (mem_cgroup_disabled())
2470                 return 0;
2471
2472         if (!do_swap_account)
2473                 goto charge_cur_mm;
2474         /*
2475          * A racing thread's fault, or swapoff, may have already updated
2476          * the pte, and even removed page from swap cache: in those cases
2477          * do_swap_page()'s pte_same() test will fail; but there's also a
2478          * KSM case which does need to charge the page.
2479          */
2480         if (!PageSwapCache(page))
2481                 goto charge_cur_mm;
2482         mem = try_get_mem_cgroup_from_page(page);
2483         if (!mem)
2484                 goto charge_cur_mm;
2485         *ptr = mem;
2486         ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2487         css_put(&mem->css);
2488         return ret;
2489 charge_cur_mm:
2490         if (unlikely(!mm))
2491                 mm = &init_mm;
2492         return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2493 }
2494
2495 static void
2496 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2497                                         enum charge_type ctype)
2498 {
2499         struct page_cgroup *pc;
2500
2501         if (mem_cgroup_disabled())
2502                 return;
2503         if (!ptr)
2504                 return;
2505         cgroup_exclude_rmdir(&ptr->css);
2506         pc = lookup_page_cgroup(page);
2507         mem_cgroup_lru_del_before_commit_swapcache(page);
2508         __mem_cgroup_commit_charge(ptr, page, pc, ctype, PAGE_SIZE);
2509         mem_cgroup_lru_add_after_commit_swapcache(page);
2510         /*
2511          * Now swap is on-memory. This means this page may be
2512          * counted both as mem and swap....double count.
2513          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2514          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2515          * may call delete_from_swap_cache() before reach here.
2516          */
2517         if (do_swap_account && PageSwapCache(page)) {
2518                 swp_entry_t ent = {.val = page_private(page)};
2519                 unsigned short id;
2520                 struct mem_cgroup *memcg;
2521
2522                 id = swap_cgroup_record(ent, 0);
2523                 rcu_read_lock();
2524                 memcg = mem_cgroup_lookup(id);
2525                 if (memcg) {
2526                         /*
2527                          * This recorded memcg can be obsolete one. So, avoid
2528                          * calling css_tryget
2529                          */
2530                         if (!mem_cgroup_is_root(memcg))
2531                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2532                         mem_cgroup_swap_statistics(memcg, false);
2533                         mem_cgroup_put(memcg);
2534                 }
2535                 rcu_read_unlock();
2536         }
2537         /*
2538          * At swapin, we may charge account against cgroup which has no tasks.
2539          * So, rmdir()->pre_destroy() can be called while we do this charge.
2540          * In that case, we need to call pre_destroy() again. check it here.
2541          */
2542         cgroup_release_and_wakeup_rmdir(&ptr->css);
2543 }
2544
2545 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2546 {
2547         __mem_cgroup_commit_charge_swapin(page, ptr,
2548                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2549 }
2550
2551 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2552 {
2553         if (mem_cgroup_disabled())
2554                 return;
2555         if (!mem)
2556                 return;
2557         mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2558 }
2559
2560 static void
2561 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2562               int page_size)
2563 {
2564         struct memcg_batch_info *batch = NULL;
2565         bool uncharge_memsw = true;
2566         /* If swapout, usage of swap doesn't decrease */
2567         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2568                 uncharge_memsw = false;
2569
2570         batch = &current->memcg_batch;
2571         /*
2572          * In usual, we do css_get() when we remember memcg pointer.
2573          * But in this case, we keep res->usage until end of a series of
2574          * uncharges. Then, it's ok to ignore memcg's refcnt.
2575          */
2576         if (!batch->memcg)
2577                 batch->memcg = mem;
2578         /*
2579          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2580          * In those cases, all pages freed continously can be expected to be in
2581          * the same cgroup and we have chance to coalesce uncharges.
2582          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2583          * because we want to do uncharge as soon as possible.
2584          */
2585
2586         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2587                 goto direct_uncharge;
2588
2589         if (page_size != PAGE_SIZE)
2590                 goto direct_uncharge;
2591
2592         /*
2593          * In typical case, batch->memcg == mem. This means we can
2594          * merge a series of uncharges to an uncharge of res_counter.
2595          * If not, we uncharge res_counter ony by one.
2596          */
2597         if (batch->memcg != mem)
2598                 goto direct_uncharge;
2599         /* remember freed charge and uncharge it later */
2600         batch->bytes += PAGE_SIZE;
2601         if (uncharge_memsw)
2602                 batch->memsw_bytes += PAGE_SIZE;
2603         return;
2604 direct_uncharge:
2605         res_counter_uncharge(&mem->res, page_size);
2606         if (uncharge_memsw)
2607                 res_counter_uncharge(&mem->memsw, page_size);
2608         if (unlikely(batch->memcg != mem))
2609                 memcg_oom_recover(mem);
2610         return;
2611 }
2612
2613 /*
2614  * uncharge if !page_mapped(page)
2615  */
2616 static struct mem_cgroup *
2617 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2618 {
2619         int count;
2620         struct page_cgroup *pc;
2621         struct mem_cgroup *mem = NULL;
2622         int page_size = PAGE_SIZE;
2623
2624         if (mem_cgroup_disabled())
2625                 return NULL;
2626
2627         if (PageSwapCache(page))
2628                 return NULL;
2629
2630         if (PageTransHuge(page)) {
2631                 page_size <<= compound_order(page);
2632                 VM_BUG_ON(!PageTransHuge(page));
2633         }
2634
2635         count = page_size >> PAGE_SHIFT;
2636         /*
2637          * Check if our page_cgroup is valid
2638          */
2639         pc = lookup_page_cgroup(page);
2640         if (unlikely(!pc || !PageCgroupUsed(pc)))
2641                 return NULL;
2642
2643         lock_page_cgroup(pc);
2644
2645         mem = pc->mem_cgroup;
2646
2647         if (!PageCgroupUsed(pc))
2648                 goto unlock_out;
2649
2650         switch (ctype) {
2651         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2652         case MEM_CGROUP_CHARGE_TYPE_DROP:
2653                 /* See mem_cgroup_prepare_migration() */
2654                 if (page_mapped(page) || PageCgroupMigration(pc))
2655                         goto unlock_out;
2656                 break;
2657         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2658                 if (!PageAnon(page)) {  /* Shared memory */
2659                         if (page->mapping && !page_is_file_cache(page))
2660                                 goto unlock_out;
2661                 } else if (page_mapped(page)) /* Anon */
2662                                 goto unlock_out;
2663                 break;
2664         default:
2665                 break;
2666         }
2667
2668         mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2669
2670         ClearPageCgroupUsed(pc);
2671         /*
2672          * pc->mem_cgroup is not cleared here. It will be accessed when it's
2673          * freed from LRU. This is safe because uncharged page is expected not
2674          * to be reused (freed soon). Exception is SwapCache, it's handled by
2675          * special functions.
2676          */
2677
2678         unlock_page_cgroup(pc);
2679         /*
2680          * even after unlock, we have mem->res.usage here and this memcg
2681          * will never be freed.
2682          */
2683         memcg_check_events(mem, page);
2684         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2685                 mem_cgroup_swap_statistics(mem, true);
2686                 mem_cgroup_get(mem);
2687         }
2688         if (!mem_cgroup_is_root(mem))
2689                 __do_uncharge(mem, ctype, page_size);
2690
2691         return mem;
2692
2693 unlock_out:
2694         unlock_page_cgroup(pc);
2695         return NULL;
2696 }
2697
2698 void mem_cgroup_uncharge_page(struct page *page)
2699 {
2700         /* early check. */
2701         if (page_mapped(page))
2702                 return;
2703         if (page->mapping && !PageAnon(page))
2704                 return;
2705         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2706 }
2707
2708 void mem_cgroup_uncharge_cache_page(struct page *page)
2709 {
2710         VM_BUG_ON(page_mapped(page));
2711         VM_BUG_ON(page->mapping);
2712         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2713 }
2714
2715 /*
2716  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2717  * In that cases, pages are freed continuously and we can expect pages
2718  * are in the same memcg. All these calls itself limits the number of
2719  * pages freed at once, then uncharge_start/end() is called properly.
2720  * This may be called prural(2) times in a context,
2721  */
2722
2723 void mem_cgroup_uncharge_start(void)
2724 {
2725         current->memcg_batch.do_batch++;
2726         /* We can do nest. */
2727         if (current->memcg_batch.do_batch == 1) {
2728                 current->memcg_batch.memcg = NULL;
2729                 current->memcg_batch.bytes = 0;
2730                 current->memcg_batch.memsw_bytes = 0;
2731         }
2732 }
2733
2734 void mem_cgroup_uncharge_end(void)
2735 {
2736         struct memcg_batch_info *batch = &current->memcg_batch;
2737
2738         if (!batch->do_batch)
2739                 return;
2740
2741         batch->do_batch--;
2742         if (batch->do_batch) /* If stacked, do nothing. */
2743                 return;
2744
2745         if (!batch->memcg)
2746                 return;
2747         /*
2748          * This "batch->memcg" is valid without any css_get/put etc...
2749          * bacause we hide charges behind us.
2750          */
2751         if (batch->bytes)
2752                 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2753         if (batch->memsw_bytes)
2754                 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2755         memcg_oom_recover(batch->memcg);
2756         /* forget this pointer (for sanity check) */
2757         batch->memcg = NULL;
2758 }
2759
2760 #ifdef CONFIG_SWAP
2761 /*
2762  * called after __delete_from_swap_cache() and drop "page" account.
2763  * memcg information is recorded to swap_cgroup of "ent"
2764  */
2765 void
2766 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2767 {
2768         struct mem_cgroup *memcg;
2769         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2770
2771         if (!swapout) /* this was a swap cache but the swap is unused ! */
2772                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2773
2774         memcg = __mem_cgroup_uncharge_common(page, ctype);
2775
2776         /*
2777          * record memcg information,  if swapout && memcg != NULL,
2778          * mem_cgroup_get() was called in uncharge().
2779          */
2780         if (do_swap_account && swapout && memcg)
2781                 swap_cgroup_record(ent, css_id(&memcg->css));
2782 }
2783 #endif
2784
2785 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2786 /*
2787  * called from swap_entry_free(). remove record in swap_cgroup and
2788  * uncharge "memsw" account.
2789  */
2790 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2791 {
2792         struct mem_cgroup *memcg;
2793         unsigned short id;
2794
2795         if (!do_swap_account)
2796                 return;
2797
2798         id = swap_cgroup_record(ent, 0);
2799         rcu_read_lock();
2800         memcg = mem_cgroup_lookup(id);
2801         if (memcg) {
2802                 /*
2803                  * We uncharge this because swap is freed.
2804                  * This memcg can be obsolete one. We avoid calling css_tryget
2805                  */
2806                 if (!mem_cgroup_is_root(memcg))
2807                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2808                 mem_cgroup_swap_statistics(memcg, false);
2809                 mem_cgroup_put(memcg);
2810         }
2811         rcu_read_unlock();
2812 }
2813
2814 /**
2815  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2816  * @entry: swap entry to be moved
2817  * @from:  mem_cgroup which the entry is moved from
2818  * @to:  mem_cgroup which the entry is moved to
2819  * @need_fixup: whether we should fixup res_counters and refcounts.
2820  *
2821  * It succeeds only when the swap_cgroup's record for this entry is the same
2822  * as the mem_cgroup's id of @from.
2823  *
2824  * Returns 0 on success, -EINVAL on failure.
2825  *
2826  * The caller must have charged to @to, IOW, called res_counter_charge() about
2827  * both res and memsw, and called css_get().
2828  */
2829 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2830                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2831 {
2832         unsigned short old_id, new_id;
2833
2834         old_id = css_id(&from->css);
2835         new_id = css_id(&to->css);
2836
2837         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2838                 mem_cgroup_swap_statistics(from, false);
2839                 mem_cgroup_swap_statistics(to, true);
2840                 /*
2841                  * This function is only called from task migration context now.
2842                  * It postpones res_counter and refcount handling till the end
2843                  * of task migration(mem_cgroup_clear_mc()) for performance
2844                  * improvement. But we cannot postpone mem_cgroup_get(to)
2845                  * because if the process that has been moved to @to does
2846                  * swap-in, the refcount of @to might be decreased to 0.
2847                  */
2848                 mem_cgroup_get(to);
2849                 if (need_fixup) {
2850                         if (!mem_cgroup_is_root(from))
2851                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2852                         mem_cgroup_put(from);
2853                         /*
2854                          * we charged both to->res and to->memsw, so we should
2855                          * uncharge to->res.
2856                          */
2857                         if (!mem_cgroup_is_root(to))
2858                                 res_counter_uncharge(&to->res, PAGE_SIZE);
2859                 }
2860                 return 0;
2861         }
2862         return -EINVAL;
2863 }
2864 #else
2865 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2866                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2867 {
2868         return -EINVAL;
2869 }
2870 #endif
2871
2872 /*
2873  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2874  * page belongs to.
2875  */
2876 int mem_cgroup_prepare_migration(struct page *page,
2877         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2878 {
2879         struct page_cgroup *pc;
2880         struct mem_cgroup *mem = NULL;
2881         enum charge_type ctype;
2882         int ret = 0;
2883
2884         *ptr = NULL;
2885
2886         VM_BUG_ON(PageTransHuge(page));
2887         if (mem_cgroup_disabled())
2888                 return 0;
2889
2890         pc = lookup_page_cgroup(page);
2891         lock_page_cgroup(pc);
2892         if (PageCgroupUsed(pc)) {
2893                 mem = pc->mem_cgroup;
2894                 css_get(&mem->css);
2895                 /*
2896                  * At migrating an anonymous page, its mapcount goes down
2897                  * to 0 and uncharge() will be called. But, even if it's fully
2898                  * unmapped, migration may fail and this page has to be
2899                  * charged again. We set MIGRATION flag here and delay uncharge
2900                  * until end_migration() is called
2901                  *
2902                  * Corner Case Thinking
2903                  * A)
2904                  * When the old page was mapped as Anon and it's unmap-and-freed
2905                  * while migration was ongoing.
2906                  * If unmap finds the old page, uncharge() of it will be delayed
2907                  * until end_migration(). If unmap finds a new page, it's
2908                  * uncharged when it make mapcount to be 1->0. If unmap code
2909                  * finds swap_migration_entry, the new page will not be mapped
2910                  * and end_migration() will find it(mapcount==0).
2911                  *
2912                  * B)
2913                  * When the old page was mapped but migraion fails, the kernel
2914                  * remaps it. A charge for it is kept by MIGRATION flag even
2915                  * if mapcount goes down to 0. We can do remap successfully
2916                  * without charging it again.
2917                  *
2918                  * C)
2919                  * The "old" page is under lock_page() until the end of
2920                  * migration, so, the old page itself will not be swapped-out.
2921                  * If the new page is swapped out before end_migraton, our
2922                  * hook to usual swap-out path will catch the event.
2923                  */
2924                 if (PageAnon(page))
2925                         SetPageCgroupMigration(pc);
2926         }
2927         unlock_page_cgroup(pc);
2928         /*
2929          * If the page is not charged at this point,
2930          * we return here.
2931          */
2932         if (!mem)
2933                 return 0;
2934
2935         *ptr = mem;
2936         ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2937         css_put(&mem->css);/* drop extra refcnt */
2938         if (ret || *ptr == NULL) {
2939                 if (PageAnon(page)) {
2940                         lock_page_cgroup(pc);
2941                         ClearPageCgroupMigration(pc);
2942                         unlock_page_cgroup(pc);
2943                         /*
2944                          * The old page may be fully unmapped while we kept it.
2945                          */
2946                         mem_cgroup_uncharge_page(page);
2947                 }
2948                 return -ENOMEM;
2949         }
2950         /*
2951          * We charge new page before it's used/mapped. So, even if unlock_page()
2952          * is called before end_migration, we can catch all events on this new
2953          * page. In the case new page is migrated but not remapped, new page's
2954          * mapcount will be finally 0 and we call uncharge in end_migration().
2955          */
2956         pc = lookup_page_cgroup(newpage);
2957         if (PageAnon(page))
2958                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2959         else if (page_is_file_cache(page))
2960                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2961         else
2962                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2963         __mem_cgroup_commit_charge(mem, page, pc, ctype, PAGE_SIZE);
2964         return ret;
2965 }
2966
2967 /* remove redundant charge if migration failed*/
2968 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2969         struct page *oldpage, struct page *newpage, bool migration_ok)
2970 {
2971         struct page *used, *unused;
2972         struct page_cgroup *pc;
2973
2974         if (!mem)
2975                 return;
2976         /* blocks rmdir() */
2977         cgroup_exclude_rmdir(&mem->css);
2978         if (!migration_ok) {
2979                 used = oldpage;
2980                 unused = newpage;
2981         } else {
2982                 used = newpage;
2983                 unused = oldpage;
2984         }
2985         /*
2986          * We disallowed uncharge of pages under migration because mapcount
2987          * of the page goes down to zero, temporarly.
2988          * Clear the flag and check the page should be charged.
2989          */
2990         pc = lookup_page_cgroup(oldpage);
2991         lock_page_cgroup(pc);
2992         ClearPageCgroupMigration(pc);
2993         unlock_page_cgroup(pc);
2994
2995         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2996
2997         /*
2998          * If a page is a file cache, radix-tree replacement is very atomic
2999          * and we can skip this check. When it was an Anon page, its mapcount
3000          * goes down to 0. But because we added MIGRATION flage, it's not
3001          * uncharged yet. There are several case but page->mapcount check
3002          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3003          * check. (see prepare_charge() also)
3004          */
3005         if (PageAnon(used))
3006                 mem_cgroup_uncharge_page(used);
3007         /*
3008          * At migration, we may charge account against cgroup which has no
3009          * tasks.
3010          * So, rmdir()->pre_destroy() can be called while we do this charge.
3011          * In that case, we need to call pre_destroy() again. check it here.
3012          */
3013         cgroup_release_and_wakeup_rmdir(&mem->css);
3014 }
3015
3016 /*
3017  * A call to try to shrink memory usage on charge failure at shmem's swapin.
3018  * Calling hierarchical_reclaim is not enough because we should update
3019  * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3020  * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3021  * not from the memcg which this page would be charged to.
3022  * try_charge_swapin does all of these works properly.
3023  */
3024 int mem_cgroup_shmem_charge_fallback(struct page *page,
3025                             struct mm_struct *mm,
3026                             gfp_t gfp_mask)
3027 {
3028         struct mem_cgroup *mem;
3029         int ret;
3030
3031         if (mem_cgroup_disabled())
3032                 return 0;
3033
3034         ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3035         if (!ret)
3036                 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3037
3038         return ret;
3039 }
3040
3041 #ifdef CONFIG_DEBUG_VM
3042 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3043 {
3044         struct page_cgroup *pc;
3045
3046         pc = lookup_page_cgroup(page);
3047         if (likely(pc) && PageCgroupUsed(pc))
3048                 return pc;
3049         return NULL;
3050 }
3051
3052 bool mem_cgroup_bad_page_check(struct page *page)
3053 {
3054         if (mem_cgroup_disabled())
3055                 return false;
3056
3057         return lookup_page_cgroup_used(page) != NULL;
3058 }
3059
3060 void mem_cgroup_print_bad_page(struct page *page)
3061 {
3062         struct page_cgroup *pc;
3063
3064         pc = lookup_page_cgroup_used(page);
3065         if (pc) {
3066                 int ret = -1;
3067                 char *path;
3068
3069                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3070                        pc, pc->flags, pc->mem_cgroup);
3071
3072                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3073                 if (path) {
3074                         rcu_read_lock();
3075                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3076                                                         path, PATH_MAX);
3077                         rcu_read_unlock();
3078                 }
3079
3080                 printk(KERN_CONT "(%s)\n",
3081                                 (ret < 0) ? "cannot get the path" : path);
3082                 kfree(path);
3083         }
3084 }
3085 #endif
3086
3087 static DEFINE_MUTEX(set_limit_mutex);
3088
3089 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3090                                 unsigned long long val)
3091 {
3092         int retry_count;
3093         u64 memswlimit, memlimit;
3094         int ret = 0;
3095         int children = mem_cgroup_count_children(memcg);
3096         u64 curusage, oldusage;
3097         int enlarge;
3098
3099         /*
3100          * For keeping hierarchical_reclaim simple, how long we should retry
3101          * is depends on callers. We set our retry-count to be function
3102          * of # of children which we should visit in this loop.
3103          */
3104         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3105
3106         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3107
3108         enlarge = 0;
3109         while (retry_count) {
3110                 if (signal_pending(current)) {
3111                         ret = -EINTR;
3112                         break;
3113                 }
3114                 /*
3115                  * Rather than hide all in some function, I do this in
3116                  * open coded manner. You see what this really does.
3117                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3118                  */
3119                 mutex_lock(&set_limit_mutex);
3120                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3121                 if (memswlimit < val) {
3122                         ret = -EINVAL;
3123                         mutex_unlock(&set_limit_mutex);
3124                         break;
3125                 }
3126
3127                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3128                 if (memlimit < val)
3129                         enlarge = 1;
3130
3131                 ret = res_counter_set_limit(&memcg->res, val);
3132                 if (!ret) {
3133                         if (memswlimit == val)
3134                                 memcg->memsw_is_minimum = true;
3135                         else
3136                                 memcg->memsw_is_minimum = false;
3137                 }
3138                 mutex_unlock(&set_limit_mutex);
3139
3140                 if (!ret)
3141                         break;
3142
3143                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3144                                                 MEM_CGROUP_RECLAIM_SHRINK);
3145                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3146                 /* Usage is reduced ? */
3147                 if (curusage >= oldusage)
3148                         retry_count--;
3149                 else
3150                         oldusage = curusage;
3151         }
3152         if (!ret && enlarge)
3153                 memcg_oom_recover(memcg);
3154
3155         return ret;
3156 }
3157
3158 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3159                                         unsigned long long val)
3160 {
3161         int retry_count;
3162         u64 memlimit, memswlimit, oldusage, curusage;
3163         int children = mem_cgroup_count_children(memcg);
3164         int ret = -EBUSY;
3165         int enlarge = 0;
3166
3167         /* see mem_cgroup_resize_res_limit */
3168         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3169         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3170         while (retry_count) {
3171                 if (signal_pending(current)) {
3172                         ret = -EINTR;
3173                         break;
3174                 }
3175                 /*
3176                  * Rather than hide all in some function, I do this in
3177                  * open coded manner. You see what this really does.
3178                  * We have to guarantee mem->res.limit < mem->memsw.limit.
3179                  */
3180                 mutex_lock(&set_limit_mutex);
3181                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3182                 if (memlimit > val) {
3183                         ret = -EINVAL;
3184                         mutex_unlock(&set_limit_mutex);
3185                         break;
3186                 }
3187                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3188                 if (memswlimit < val)
3189                         enlarge = 1;
3190                 ret = res_counter_set_limit(&memcg->memsw, val);
3191                 if (!ret) {
3192                         if (memlimit == val)
3193                                 memcg->memsw_is_minimum = true;
3194                         else
3195                                 memcg->memsw_is_minimum = false;
3196                 }
3197                 mutex_unlock(&set_limit_mutex);
3198
3199                 if (!ret)
3200                         break;
3201
3202                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3203                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3204                                                 MEM_CGROUP_RECLAIM_SHRINK);
3205                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3206                 /* Usage is reduced ? */
3207                 if (curusage >= oldusage)
3208                         retry_count--;
3209                 else
3210                         oldusage = curusage;
3211         }
3212         if (!ret && enlarge)
3213                 memcg_oom_recover(memcg);
3214         return ret;
3215 }
3216
3217 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3218                                             gfp_t gfp_mask)
3219 {
3220         unsigned long nr_reclaimed = 0;
3221         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3222         unsigned long reclaimed;
3223         int loop = 0;
3224         struct mem_cgroup_tree_per_zone *mctz;
3225         unsigned long long excess;
3226
3227         if (order > 0)
3228                 return 0;
3229
3230         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3231         /*
3232          * This loop can run a while, specially if mem_cgroup's continuously
3233          * keep exceeding their soft limit and putting the system under
3234          * pressure
3235          */
3236         do {
3237                 if (next_mz)
3238                         mz = next_mz;
3239                 else
3240                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3241                 if (!mz)
3242                         break;
3243
3244                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3245                                                 gfp_mask,
3246                                                 MEM_CGROUP_RECLAIM_SOFT);
3247                 nr_reclaimed += reclaimed;
3248                 spin_lock(&mctz->lock);
3249
3250                 /*
3251                  * If we failed to reclaim anything from this memory cgroup
3252                  * it is time to move on to the next cgroup
3253                  */
3254                 next_mz = NULL;
3255                 if (!reclaimed) {
3256                         do {
3257                                 /*
3258                                  * Loop until we find yet another one.
3259                                  *
3260                                  * By the time we get the soft_limit lock
3261                                  * again, someone might have aded the
3262                                  * group back on the RB tree. Iterate to
3263                                  * make sure we get a different mem.
3264                                  * mem_cgroup_largest_soft_limit_node returns
3265                                  * NULL if no other cgroup is present on
3266                                  * the tree
3267                                  */
3268                                 next_mz =
3269                                 __mem_cgroup_largest_soft_limit_node(mctz);
3270                                 if (next_mz == mz) {
3271                                         css_put(&next_mz->mem->css);
3272                                         next_mz = NULL;
3273                                 } else /* next_mz == NULL or other memcg */
3274                                         break;
3275                         } while (1);
3276                 }
3277                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3278                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3279                 /*
3280                  * One school of thought says that we should not add
3281                  * back the node to the tree if reclaim returns 0.
3282                  * But our reclaim could return 0, simply because due
3283                  * to priority we are exposing a smaller subset of
3284                  * memory to reclaim from. Consider this as a longer
3285                  * term TODO.
3286                  */
3287                 /* If excess == 0, no tree ops */
3288                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3289                 spin_unlock(&mctz->lock);
3290                 css_put(&mz->mem->css);
3291                 loop++;
3292                 /*
3293                  * Could not reclaim anything and there are no more
3294                  * mem cgroups to try or we seem to be looping without
3295                  * reclaiming anything.
3296                  */
3297                 if (!nr_reclaimed &&
3298                         (next_mz == NULL ||
3299                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3300                         break;
3301         } while (!nr_reclaimed);
3302         if (next_mz)
3303                 css_put(&next_mz->mem->css);
3304         return nr_reclaimed;
3305 }
3306
3307 /*
3308  * This routine traverse page_cgroup in given list and drop them all.
3309  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3310  */
3311 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3312                                 int node, int zid, enum lru_list lru)
3313 {
3314         struct zone *zone;
3315         struct mem_cgroup_per_zone *mz;
3316         struct page_cgroup *pc, *busy;
3317         unsigned long flags, loop;
3318         struct list_head *list;
3319         int ret = 0;
3320
3321         zone = &NODE_DATA(node)->node_zones[zid];
3322         mz = mem_cgroup_zoneinfo(mem, node, zid);
3323         list = &mz->lists[lru];
3324
3325         loop = MEM_CGROUP_ZSTAT(mz, lru);
3326         /* give some margin against EBUSY etc...*/
3327         loop += 256;
3328         busy = NULL;
3329         while (loop--) {
3330                 struct page *page;
3331
3332                 ret = 0;
3333                 spin_lock_irqsave(&zone->lru_lock, flags);
3334                 if (list_empty(list)) {
3335                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3336                         break;
3337                 }
3338                 pc = list_entry(list->prev, struct page_cgroup, lru);
3339                 if (busy == pc) {
3340                         list_move(&pc->lru, list);
3341                         busy = NULL;
3342                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3343                         continue;
3344                 }
3345                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3346
3347                 page = pc->page;
3348
3349                 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3350                 if (ret == -ENOMEM)
3351                         break;
3352
3353                 if (ret == -EBUSY || ret == -EINVAL) {
3354                         /* found lock contention or "pc" is obsolete. */
3355                         busy = pc;
3356                         cond_resched();
3357                 } else
3358                         busy = NULL;
3359         }
3360
3361         if (!ret && !list_empty(list))
3362                 return -EBUSY;
3363         return ret;
3364 }
3365
3366 /*
3367  * make mem_cgroup's charge to be 0 if there is no task.
3368  * This enables deleting this mem_cgroup.
3369  */
3370 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3371 {
3372         int ret;
3373         int node, zid, shrink;
3374         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3375         struct cgroup *cgrp = mem->css.cgroup;
3376
3377         css_get(&mem->css);
3378
3379         shrink = 0;
3380         /* should free all ? */
3381         if (free_all)
3382                 goto try_to_free;
3383 move_account:
3384         do {
3385                 ret = -EBUSY;
3386                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3387                         goto out;
3388                 ret = -EINTR;
3389                 if (signal_pending(current))
3390                         goto out;
3391                 /* This is for making all *used* pages to be on LRU. */
3392                 lru_add_drain_all();
3393                 drain_all_stock_sync();
3394                 ret = 0;
3395                 mem_cgroup_start_move(mem);
3396                 for_each_node_state(node, N_HIGH_MEMORY) {
3397                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3398                                 enum lru_list l;
3399                                 for_each_lru(l) {
3400                                         ret = mem_cgroup_force_empty_list(mem,
3401                                                         node, zid, l);
3402                                         if (ret)
3403                                                 break;
3404                                 }
3405                         }
3406                         if (ret)
3407                                 break;
3408                 }
3409                 mem_cgroup_end_move(mem);
3410                 memcg_oom_recover(mem);
3411                 /* it seems parent cgroup doesn't have enough mem */
3412                 if (ret == -ENOMEM)
3413                         goto try_to_free;
3414                 cond_resched();
3415         /* "ret" should also be checked to ensure all lists are empty. */
3416         } while (mem->res.usage > 0 || ret);
3417 out:
3418         css_put(&mem->css);
3419         return ret;
3420
3421 try_to_free:
3422         /* returns EBUSY if there is a task or if we come here twice. */
3423         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3424                 ret = -EBUSY;
3425                 goto out;
3426         }
3427         /* we call try-to-free pages for make this cgroup empty */
3428         lru_add_drain_all();
3429         /* try to free all pages in this cgroup */
3430         shrink = 1;
3431         while (nr_retries && mem->res.usage > 0) {
3432                 int progress;
3433
3434                 if (signal_pending(current)) {
3435                         ret = -EINTR;
3436                         goto out;
3437                 }
3438                 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3439                                                 false, get_swappiness(mem));
3440                 if (!progress) {
3441                         nr_retries--;
3442                         /* maybe some writeback is necessary */
3443                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3444                 }
3445
3446         }
3447         lru_add_drain();
3448         /* try move_account...there may be some *locked* pages. */
3449         goto move_account;
3450 }
3451
3452 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3453 {
3454         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3455 }
3456
3457
3458 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3459 {
3460         return mem_cgroup_from_cont(cont)->use_hierarchy;
3461 }
3462
3463 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3464                                         u64 val)
3465 {
3466         int retval = 0;
3467         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3468         struct cgroup *parent = cont->parent;
3469         struct mem_cgroup *parent_mem = NULL;
3470
3471         if (parent)
3472                 parent_mem = mem_cgroup_from_cont(parent);
3473
3474         cgroup_lock();
3475         /*
3476          * If parent's use_hierarchy is set, we can't make any modifications
3477          * in the child subtrees. If it is unset, then the change can
3478          * occur, provided the current cgroup has no children.
3479          *
3480          * For the root cgroup, parent_mem is NULL, we allow value to be
3481          * set if there are no children.
3482          */
3483         if ((!parent_mem || !parent_mem->use_hierarchy) &&
3484                                 (val == 1 || val == 0)) {
3485                 if (list_empty(&cont->children))
3486                         mem->use_hierarchy = val;
3487                 else
3488                         retval = -EBUSY;
3489         } else
3490                 retval = -EINVAL;
3491         cgroup_unlock();
3492
3493         return retval;
3494 }
3495
3496
3497 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3498                                 enum mem_cgroup_stat_index idx)
3499 {
3500         struct mem_cgroup *iter;
3501         s64 val = 0;
3502
3503         /* each per cpu's value can be minus.Then, use s64 */
3504         for_each_mem_cgroup_tree(iter, mem)
3505                 val += mem_cgroup_read_stat(iter, idx);
3506
3507         if (val < 0) /* race ? */
3508                 val = 0;
3509         return val;
3510 }
3511
3512 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3513 {
3514         u64 val;
3515
3516         if (!mem_cgroup_is_root(mem)) {
3517                 if (!swap)
3518                         return res_counter_read_u64(&mem->res, RES_USAGE);
3519                 else
3520                         return res_counter_read_u64(&mem->memsw, RES_USAGE);
3521         }
3522
3523         val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3524         val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3525
3526         if (swap)
3527                 val += mem_cgroup_get_recursive_idx_stat(mem,
3528                                 MEM_CGROUP_STAT_SWAPOUT);
3529
3530         return val << PAGE_SHIFT;
3531 }
3532
3533 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3534 {
3535         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3536         u64 val;
3537         int type, name;
3538
3539         type = MEMFILE_TYPE(cft->private);
3540         name = MEMFILE_ATTR(cft->private);
3541         switch (type) {
3542         case _MEM:
3543                 if (name == RES_USAGE)
3544                         val = mem_cgroup_usage(mem, false);
3545                 else
3546                         val = res_counter_read_u64(&mem->res, name);
3547                 break;
3548         case _MEMSWAP:
3549                 if (name == RES_USAGE)
3550                         val = mem_cgroup_usage(mem, true);
3551                 else
3552                         val = res_counter_read_u64(&mem->memsw, name);
3553                 break;
3554         default:
3555                 BUG();
3556                 break;
3557         }
3558         return val;
3559 }
3560 /*
3561  * The user of this function is...
3562  * RES_LIMIT.
3563  */
3564 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3565                             const char *buffer)
3566 {
3567         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3568         int type, name;
3569         unsigned long long val;
3570         int ret;
3571
3572         type = MEMFILE_TYPE(cft->private);
3573         name = MEMFILE_ATTR(cft->private);
3574         switch (name) {
3575         case RES_LIMIT:
3576                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3577                         ret = -EINVAL;
3578                         break;
3579                 }
3580                 /* This function does all necessary parse...reuse it */
3581                 ret = res_counter_memparse_write_strategy(buffer, &val);
3582                 if (ret)
3583                         break;
3584                 if (type == _MEM)
3585                         ret = mem_cgroup_resize_limit(memcg, val);
3586                 else
3587                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3588                 break;
3589         case RES_SOFT_LIMIT:
3590                 ret = res_counter_memparse_write_strategy(buffer, &val);
3591                 if (ret)
3592                         break;
3593                 /*
3594                  * For memsw, soft limits are hard to implement in terms
3595                  * of semantics, for now, we support soft limits for
3596                  * control without swap
3597                  */
3598                 if (type == _MEM)
3599                         ret = res_counter_set_soft_limit(&memcg->res, val);
3600                 else
3601                         ret = -EINVAL;
3602                 break;
3603         default:
3604                 ret = -EINVAL; /* should be BUG() ? */
3605                 break;
3606         }
3607         return ret;
3608 }
3609
3610 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3611                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3612 {
3613         struct cgroup *cgroup;
3614         unsigned long long min_limit, min_memsw_limit, tmp;
3615
3616         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3617         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3618         cgroup = memcg->css.cgroup;
3619         if (!memcg->use_hierarchy)
3620                 goto out;
3621
3622         while (cgroup->parent) {
3623                 cgroup = cgroup->parent;
3624                 memcg = mem_cgroup_from_cont(cgroup);
3625                 if (!memcg->use_hierarchy)
3626                         break;
3627                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3628                 min_limit = min(min_limit, tmp);
3629                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3630                 min_memsw_limit = min(min_memsw_limit, tmp);
3631         }
3632 out:
3633         *mem_limit = min_limit;
3634         *memsw_limit = min_memsw_limit;
3635         return;
3636 }
3637
3638 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3639 {
3640         struct mem_cgroup *mem;
3641         int type, name;
3642
3643         mem = mem_cgroup_from_cont(cont);
3644         type = MEMFILE_TYPE(event);
3645         name = MEMFILE_ATTR(event);
3646         switch (name) {
3647         case RES_MAX_USAGE:
3648                 if (type == _MEM)
3649                         res_counter_reset_max(&mem->res);
3650                 else
3651                         res_counter_reset_max(&mem->memsw);
3652                 break;
3653         case RES_FAILCNT:
3654                 if (type == _MEM)
3655                         res_counter_reset_failcnt(&mem->res);
3656                 else
3657                         res_counter_reset_failcnt(&mem->memsw);
3658                 break;
3659         }
3660
3661         return 0;
3662 }
3663
3664 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3665                                         struct cftype *cft)
3666 {
3667         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3668 }
3669
3670 #ifdef CONFIG_MMU
3671 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3672                                         struct cftype *cft, u64 val)
3673 {
3674         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3675
3676         if (val >= (1 << NR_MOVE_TYPE))
3677                 return -EINVAL;
3678         /*
3679          * We check this value several times in both in can_attach() and
3680          * attach(), so we need cgroup lock to prevent this value from being
3681          * inconsistent.
3682          */
3683         cgroup_lock();
3684         mem->move_charge_at_immigrate = val;
3685         cgroup_unlock();
3686
3687         return 0;
3688 }
3689 #else
3690 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3691                                         struct cftype *cft, u64 val)
3692 {
3693         return -ENOSYS;
3694 }
3695 #endif
3696
3697
3698 /* For read statistics */
3699 enum {
3700         MCS_CACHE,
3701         MCS_RSS,
3702         MCS_FILE_MAPPED,
3703         MCS_PGPGIN,
3704         MCS_PGPGOUT,
3705         MCS_SWAP,
3706         MCS_INACTIVE_ANON,
3707         MCS_ACTIVE_ANON,
3708         MCS_INACTIVE_FILE,
3709         MCS_ACTIVE_FILE,
3710         MCS_UNEVICTABLE,
3711         NR_MCS_STAT,
3712 };
3713
3714 struct mcs_total_stat {
3715         s64 stat[NR_MCS_STAT];
3716 };
3717
3718 struct {
3719         char *local_name;
3720         char *total_name;
3721 } memcg_stat_strings[NR_MCS_STAT] = {
3722         {"cache", "total_cache"},
3723         {"rss", "total_rss"},
3724         {"mapped_file", "total_mapped_file"},
3725         {"pgpgin", "total_pgpgin"},
3726         {"pgpgout", "total_pgpgout"},
3727         {"swap", "total_swap"},
3728         {"inactive_anon", "total_inactive_anon"},
3729         {"active_anon", "total_active_anon"},
3730         {"inactive_file", "total_inactive_file"},
3731         {"active_file", "total_active_file"},
3732         {"unevictable", "total_unevictable"}
3733 };
3734
3735
3736 static void
3737 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3738 {
3739         s64 val;
3740
3741         /* per cpu stat */
3742         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3743         s->stat[MCS_CACHE] += val * PAGE_SIZE;
3744         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3745         s->stat[MCS_RSS] += val * PAGE_SIZE;
3746         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3747         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3748         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3749         s->stat[MCS_PGPGIN] += val;
3750         val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3751         s->stat[MCS_PGPGOUT] += val;
3752         if (do_swap_account) {
3753                 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3754                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3755         }
3756
3757         /* per zone stat */
3758         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3759         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3760         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3761         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3762         val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3763         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3764         val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3765         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3766         val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3767         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3768 }
3769
3770 static void
3771 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3772 {
3773         struct mem_cgroup *iter;
3774
3775         for_each_mem_cgroup_tree(iter, mem)
3776                 mem_cgroup_get_local_stat(iter, s);
3777 }
3778
3779 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3780                                  struct cgroup_map_cb *cb)
3781 {
3782         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3783         struct mcs_total_stat mystat;
3784         int i;
3785
3786         memset(&mystat, 0, sizeof(mystat));
3787         mem_cgroup_get_local_stat(mem_cont, &mystat);
3788
3789         for (i = 0; i < NR_MCS_STAT; i++) {
3790                 if (i == MCS_SWAP && !do_swap_account)
3791                         continue;
3792                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3793         }
3794
3795         /* Hierarchical information */
3796         {
3797                 unsigned long long limit, memsw_limit;
3798                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3799                 cb->fill(cb, "hierarchical_memory_limit", limit);
3800                 if (do_swap_account)
3801                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3802         }
3803
3804         memset(&mystat, 0, sizeof(mystat));
3805         mem_cgroup_get_total_stat(mem_cont, &mystat);
3806         for (i = 0; i < NR_MCS_STAT; i++) {
3807                 if (i == MCS_SWAP && !do_swap_account)
3808                         continue;
3809                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3810         }
3811
3812 #ifdef CONFIG_DEBUG_VM
3813         cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3814
3815         {
3816                 int nid, zid;
3817                 struct mem_cgroup_per_zone *mz;
3818                 unsigned long recent_rotated[2] = {0, 0};
3819                 unsigned long recent_scanned[2] = {0, 0};
3820
3821                 for_each_online_node(nid)
3822                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3823                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3824
3825                                 recent_rotated[0] +=
3826                                         mz->reclaim_stat.recent_rotated[0];
3827                                 recent_rotated[1] +=
3828                                         mz->reclaim_stat.recent_rotated[1];
3829                                 recent_scanned[0] +=
3830                                         mz->reclaim_stat.recent_scanned[0];
3831                                 recent_scanned[1] +=
3832                                         mz->reclaim_stat.recent_scanned[1];
3833                         }
3834                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3835                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3836                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3837                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3838         }
3839 #endif
3840
3841         return 0;
3842 }
3843
3844 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3845 {
3846         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3847
3848         return get_swappiness(memcg);
3849 }
3850
3851 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3852                                        u64 val)
3853 {
3854         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3855         struct mem_cgroup *parent;
3856
3857         if (val > 100)
3858                 return -EINVAL;
3859
3860         if (cgrp->parent == NULL)
3861                 return -EINVAL;
3862
3863         parent = mem_cgroup_from_cont(cgrp->parent);
3864
3865         cgroup_lock();
3866
3867         /* If under hierarchy, only empty-root can set this value */
3868         if ((parent->use_hierarchy) ||
3869             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3870                 cgroup_unlock();
3871                 return -EINVAL;
3872         }
3873
3874         spin_lock(&memcg->reclaim_param_lock);
3875         memcg->swappiness = val;
3876         spin_unlock(&memcg->reclaim_param_lock);
3877
3878         cgroup_unlock();
3879
3880         return 0;
3881 }
3882
3883 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3884 {
3885         struct mem_cgroup_threshold_ary *t;
3886         u64 usage;
3887         int i;
3888
3889         rcu_read_lock();
3890         if (!swap)
3891                 t = rcu_dereference(memcg->thresholds.primary);
3892         else
3893                 t = rcu_dereference(memcg->memsw_thresholds.primary);
3894
3895         if (!t)
3896                 goto unlock;
3897
3898         usage = mem_cgroup_usage(memcg, swap);
3899
3900         /*
3901          * current_threshold points to threshold just below usage.
3902          * If it's not true, a threshold was crossed after last
3903          * call of __mem_cgroup_threshold().
3904          */
3905         i = t->current_threshold;
3906
3907         /*
3908          * Iterate backward over array of thresholds starting from
3909          * current_threshold and check if a threshold is crossed.
3910          * If none of thresholds below usage is crossed, we read
3911          * only one element of the array here.
3912          */
3913         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3914                 eventfd_signal(t->entries[i].eventfd, 1);
3915
3916         /* i = current_threshold + 1 */
3917         i++;
3918
3919         /*
3920          * Iterate forward over array of thresholds starting from
3921          * current_threshold+1 and check if a threshold is crossed.
3922          * If none of thresholds above usage is crossed, we read
3923          * only one element of the array here.
3924          */
3925         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3926                 eventfd_signal(t->entries[i].eventfd, 1);
3927
3928         /* Update current_threshold */
3929         t->current_threshold = i - 1;
3930 unlock:
3931         rcu_read_unlock();
3932 }
3933
3934 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3935 {
3936         while (memcg) {
3937                 __mem_cgroup_threshold(memcg, false);
3938                 if (do_swap_account)
3939                         __mem_cgroup_threshold(memcg, true);
3940
3941                 memcg = parent_mem_cgroup(memcg);
3942         }
3943 }
3944
3945 static int compare_thresholds(const void *a, const void *b)
3946 {
3947         const struct mem_cgroup_threshold *_a = a;
3948         const struct mem_cgroup_threshold *_b = b;
3949
3950         return _a->threshold - _b->threshold;
3951 }
3952
3953 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3954 {
3955         struct mem_cgroup_eventfd_list *ev;
3956
3957         list_for_each_entry(ev, &mem->oom_notify, list)
3958                 eventfd_signal(ev->eventfd, 1);
3959         return 0;
3960 }
3961
3962 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3963 {
3964         struct mem_cgroup *iter;
3965
3966         for_each_mem_cgroup_tree(iter, mem)
3967                 mem_cgroup_oom_notify_cb(iter);
3968 }
3969
3970 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3971         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3972 {
3973         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3974         struct mem_cgroup_thresholds *thresholds;
3975         struct mem_cgroup_threshold_ary *new;
3976         int type = MEMFILE_TYPE(cft->private);
3977         u64 threshold, usage;
3978         int i, size, ret;
3979
3980         ret = res_counter_memparse_write_strategy(args, &threshold);
3981         if (ret)
3982                 return ret;
3983
3984         mutex_lock(&memcg->thresholds_lock);
3985
3986         if (type == _MEM)
3987                 thresholds = &memcg->thresholds;
3988         else if (type == _MEMSWAP)
3989                 thresholds = &memcg->memsw_thresholds;
3990         else
3991                 BUG();
3992
3993         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3994
3995         /* Check if a threshold crossed before adding a new one */
3996         if (thresholds->primary)
3997                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3998
3999         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4000
4001         /* Allocate memory for new array of thresholds */
4002         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4003                         GFP_KERNEL);
4004         if (!new) {
4005                 ret = -ENOMEM;
4006                 goto unlock;
4007         }
4008         new->size = size;
4009
4010         /* Copy thresholds (if any) to new array */
4011         if (thresholds->primary) {
4012                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4013                                 sizeof(struct mem_cgroup_threshold));
4014         }
4015
4016         /* Add new threshold */
4017         new->entries[size - 1].eventfd = eventfd;
4018         new->entries[size - 1].threshold = threshold;
4019
4020         /* Sort thresholds. Registering of new threshold isn't time-critical */
4021         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4022                         compare_thresholds, NULL);
4023
4024         /* Find current threshold */
4025         new->current_threshold = -1;
4026         for (i = 0; i < size; i++) {
4027                 if (new->entries[i].threshold < usage) {
4028                         /*
4029                          * new->current_threshold will not be used until
4030                          * rcu_assign_pointer(), so it's safe to increment
4031                          * it here.
4032                          */
4033                         ++new->current_threshold;
4034                 }
4035         }
4036
4037         /* Free old spare buffer and save old primary buffer as spare */
4038         kfree(thresholds->spare);
4039         thresholds->spare = thresholds->primary;
4040
4041         rcu_assign_pointer(thresholds->primary, new);
4042
4043         /* To be sure that nobody uses thresholds */
4044         synchronize_rcu();
4045
4046 unlock:
4047         mutex_unlock(&memcg->thresholds_lock);
4048
4049         return ret;
4050 }
4051
4052 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4053         struct cftype *cft, struct eventfd_ctx *eventfd)
4054 {
4055         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4056         struct mem_cgroup_thresholds *thresholds;
4057         struct mem_cgroup_threshold_ary *new;
4058         int type = MEMFILE_TYPE(cft->private);
4059         u64 usage;
4060         int i, j, size;
4061
4062         mutex_lock(&memcg->thresholds_lock);
4063         if (type == _MEM)
4064                 thresholds = &memcg->thresholds;
4065         else if (type == _MEMSWAP)
4066                 thresholds = &memcg->memsw_thresholds;
4067         else
4068                 BUG();
4069
4070         /*
4071          * Something went wrong if we trying to unregister a threshold
4072          * if we don't have thresholds
4073          */
4074         BUG_ON(!thresholds);
4075
4076         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4077
4078         /* Check if a threshold crossed before removing */
4079         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4080
4081         /* Calculate new number of threshold */
4082         size = 0;
4083         for (i = 0; i < thresholds->primary->size; i++) {
4084                 if (thresholds->primary->entries[i].eventfd != eventfd)
4085                         size++;
4086         }
4087
4088         new = thresholds->spare;
4089
4090         /* Set thresholds array to NULL if we don't have thresholds */
4091         if (!size) {
4092                 kfree(new);
4093                 new = NULL;
4094                 goto swap_buffers;
4095         }
4096
4097         new->size = size;
4098
4099         /* Copy thresholds and find current threshold */
4100         new->current_threshold = -1;
4101         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4102                 if (thresholds->primary->entries[i].eventfd == eventfd)
4103                         continue;
4104
4105                 new->entries[j] = thresholds->primary->entries[i];
4106                 if (new->entries[j].threshold < usage) {
4107                         /*
4108                          * new->current_threshold will not be used
4109                          * until rcu_assign_pointer(), so it's safe to increment
4110                          * it here.
4111                          */
4112                         ++new->current_threshold;
4113                 }
4114                 j++;
4115         }
4116
4117 swap_buffers:
4118         /* Swap primary and spare array */
4119         thresholds->spare = thresholds->primary;
4120         rcu_assign_pointer(thresholds->primary, new);
4121
4122         /* To be sure that nobody uses thresholds */
4123         synchronize_rcu();
4124
4125         mutex_unlock(&memcg->thresholds_lock);
4126 }
4127
4128 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4129         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4130 {
4131         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4132         struct mem_cgroup_eventfd_list *event;
4133         int type = MEMFILE_TYPE(cft->private);
4134
4135         BUG_ON(type != _OOM_TYPE);
4136         event = kmalloc(sizeof(*event), GFP_KERNEL);
4137         if (!event)
4138                 return -ENOMEM;
4139
4140         mutex_lock(&memcg_oom_mutex);
4141
4142         event->eventfd = eventfd;
4143         list_add(&event->list, &memcg->oom_notify);
4144
4145         /* already in OOM ? */
4146         if (atomic_read(&memcg->oom_lock))
4147                 eventfd_signal(eventfd, 1);
4148         mutex_unlock(&memcg_oom_mutex);
4149
4150         return 0;
4151 }
4152
4153 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4154         struct cftype *cft, struct eventfd_ctx *eventfd)
4155 {
4156         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4157         struct mem_cgroup_eventfd_list *ev, *tmp;
4158         int type = MEMFILE_TYPE(cft->private);
4159
4160         BUG_ON(type != _OOM_TYPE);
4161
4162         mutex_lock(&memcg_oom_mutex);
4163
4164         list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4165                 if (ev->eventfd == eventfd) {
4166                         list_del(&ev->list);
4167                         kfree(ev);
4168                 }
4169         }
4170
4171         mutex_unlock(&memcg_oom_mutex);
4172 }
4173
4174 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4175         struct cftype *cft,  struct cgroup_map_cb *cb)
4176 {
4177         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4178
4179         cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4180
4181         if (atomic_read(&mem->oom_lock))
4182                 cb->fill(cb, "under_oom", 1);
4183         else
4184                 cb->fill(cb, "under_oom", 0);
4185         return 0;
4186 }
4187
4188 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4189         struct cftype *cft, u64 val)
4190 {
4191         struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4192         struct mem_cgroup *parent;
4193
4194         /* cannot set to root cgroup and only 0 and 1 are allowed */
4195         if (!cgrp->parent || !((val == 0) || (val == 1)))
4196                 return -EINVAL;
4197
4198         parent = mem_cgroup_from_cont(cgrp->parent);
4199
4200         cgroup_lock();
4201         /* oom-kill-disable is a flag for subhierarchy. */
4202         if ((parent->use_hierarchy) ||
4203             (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4204                 cgroup_unlock();
4205                 return -EINVAL;
4206         }
4207         mem->oom_kill_disable = val;
4208         if (!val)
4209                 memcg_oom_recover(mem);
4210         cgroup_unlock();
4211         return 0;
4212 }
4213
4214 static struct cftype mem_cgroup_files[] = {
4215         {
4216                 .name = "usage_in_bytes",
4217                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4218                 .read_u64 = mem_cgroup_read,
4219                 .register_event = mem_cgroup_usage_register_event,
4220                 .unregister_event = mem_cgroup_usage_unregister_event,
4221         },
4222         {
4223                 .name = "max_usage_in_bytes",
4224                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4225                 .trigger = mem_cgroup_reset,
4226                 .read_u64 = mem_cgroup_read,
4227         },
4228         {
4229                 .name = "limit_in_bytes",
4230                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4231                 .write_string = mem_cgroup_write,
4232                 .read_u64 = mem_cgroup_read,
4233         },
4234         {
4235                 .name = "soft_limit_in_bytes",
4236                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4237                 .write_string = mem_cgroup_write,
4238                 .read_u64 = mem_cgroup_read,
4239         },
4240         {
4241                 .name = "failcnt",
4242                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4243                 .trigger = mem_cgroup_reset,
4244                 .read_u64 = mem_cgroup_read,
4245         },
4246         {
4247                 .name = "stat",
4248                 .read_map = mem_control_stat_show,
4249         },
4250         {
4251                 .name = "force_empty",
4252                 .trigger = mem_cgroup_force_empty_write,
4253         },
4254         {
4255                 .name = "use_hierarchy",
4256                 .write_u64 = mem_cgroup_hierarchy_write,
4257                 .read_u64 = mem_cgroup_hierarchy_read,
4258         },
4259         {
4260                 .name = "swappiness",
4261                 .read_u64 = mem_cgroup_swappiness_read,
4262                 .write_u64 = mem_cgroup_swappiness_write,
4263         },
4264         {
4265                 .name = "move_charge_at_immigrate",
4266                 .read_u64 = mem_cgroup_move_charge_read,
4267                 .write_u64 = mem_cgroup_move_charge_write,
4268         },
4269         {
4270                 .name = "oom_control",
4271                 .read_map = mem_cgroup_oom_control_read,
4272                 .write_u64 = mem_cgroup_oom_control_write,
4273                 .register_event = mem_cgroup_oom_register_event,
4274                 .unregister_event = mem_cgroup_oom_unregister_event,
4275                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4276         },
4277 };
4278
4279 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4280 static struct cftype memsw_cgroup_files[] = {
4281         {
4282                 .name = "memsw.usage_in_bytes",
4283                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4284                 .read_u64 = mem_cgroup_read,
4285                 .register_event = mem_cgroup_usage_register_event,
4286                 .unregister_event = mem_cgroup_usage_unregister_event,
4287         },
4288         {
4289                 .name = "memsw.max_usage_in_bytes",
4290                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4291                 .trigger = mem_cgroup_reset,
4292                 .read_u64 = mem_cgroup_read,
4293         },
4294         {
4295                 .name = "memsw.limit_in_bytes",
4296                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4297                 .write_string = mem_cgroup_write,
4298                 .read_u64 = mem_cgroup_read,
4299         },
4300         {
4301                 .name = "memsw.failcnt",
4302                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4303                 .trigger = mem_cgroup_reset,
4304                 .read_u64 = mem_cgroup_read,
4305         },
4306 };
4307
4308 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4309 {
4310         if (!do_swap_account)
4311                 return 0;
4312         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4313                                 ARRAY_SIZE(memsw_cgroup_files));
4314 };
4315 #else
4316 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4317 {
4318         return 0;
4319 }
4320 #endif
4321
4322 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4323 {
4324         struct mem_cgroup_per_node *pn;
4325         struct mem_cgroup_per_zone *mz;
4326         enum lru_list l;
4327         int zone, tmp = node;
4328         /*
4329          * This routine is called against possible nodes.
4330          * But it's BUG to call kmalloc() against offline node.
4331          *
4332          * TODO: this routine can waste much memory for nodes which will
4333          *       never be onlined. It's better to use memory hotplug callback
4334          *       function.
4335          */
4336         if (!node_state(node, N_NORMAL_MEMORY))
4337                 tmp = -1;
4338         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4339         if (!pn)
4340                 return 1;
4341
4342         mem->info.nodeinfo[node] = pn;
4343         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4344                 mz = &pn->zoneinfo[zone];
4345                 for_each_lru(l)
4346                         INIT_LIST_HEAD(&mz->lists[l]);
4347                 mz->usage_in_excess = 0;
4348                 mz->on_tree = false;
4349                 mz->mem = mem;
4350         }
4351         return 0;
4352 }
4353
4354 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4355 {
4356         kfree(mem->info.nodeinfo[node]);
4357 }
4358
4359 static struct mem_cgroup *mem_cgroup_alloc(void)
4360 {
4361         struct mem_cgroup *mem;
4362         int size = sizeof(struct mem_cgroup);
4363
4364         /* Can be very big if MAX_NUMNODES is very big */
4365         if (size < PAGE_SIZE)
4366                 mem = kzalloc(size, GFP_KERNEL);
4367         else
4368                 mem = vzalloc(size);
4369
4370         if (!mem)
4371                 return NULL;
4372
4373         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4374         if (!mem->stat)
4375                 goto out_free;
4376         spin_lock_init(&mem->pcp_counter_lock);
4377         return mem;
4378
4379 out_free:
4380         if (size < PAGE_SIZE)
4381                 kfree(mem);
4382         else
4383                 vfree(mem);
4384         return NULL;
4385 }
4386
4387 /*
4388  * At destroying mem_cgroup, references from swap_cgroup can remain.
4389  * (scanning all at force_empty is too costly...)
4390  *
4391  * Instead of clearing all references at force_empty, we remember
4392  * the number of reference from swap_cgroup and free mem_cgroup when
4393  * it goes down to 0.
4394  *
4395  * Removal of cgroup itself succeeds regardless of refs from swap.
4396  */
4397
4398 static void __mem_cgroup_free(struct mem_cgroup *mem)
4399 {
4400         int node;
4401
4402         mem_cgroup_remove_from_trees(mem);
4403         free_css_id(&mem_cgroup_subsys, &mem->css);
4404
4405         for_each_node_state(node, N_POSSIBLE)
4406                 free_mem_cgroup_per_zone_info(mem, node);
4407
4408         free_percpu(mem->stat);
4409         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4410                 kfree(mem);
4411         else
4412                 vfree(mem);
4413 }
4414
4415 static void mem_cgroup_get(struct mem_cgroup *mem)
4416 {
4417         atomic_inc(&mem->refcnt);
4418 }
4419
4420 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4421 {
4422         if (atomic_sub_and_test(count, &mem->refcnt)) {
4423                 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4424                 __mem_cgroup_free(mem);
4425                 if (parent)
4426                         mem_cgroup_put(parent);
4427         }
4428 }
4429
4430 static void mem_cgroup_put(struct mem_cgroup *mem)
4431 {
4432         __mem_cgroup_put(mem, 1);
4433 }
4434
4435 /*
4436  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4437  */
4438 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4439 {
4440         if (!mem->res.parent)
4441                 return NULL;
4442         return mem_cgroup_from_res_counter(mem->res.parent, res);
4443 }
4444
4445 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4446 static void __init enable_swap_cgroup(void)
4447 {
4448         if (!mem_cgroup_disabled() && really_do_swap_account)
4449                 do_swap_account = 1;
4450 }
4451 #else
4452 static void __init enable_swap_cgroup(void)
4453 {
4454 }
4455 #endif
4456
4457 static int mem_cgroup_soft_limit_tree_init(void)
4458 {
4459         struct mem_cgroup_tree_per_node *rtpn;
4460         struct mem_cgroup_tree_per_zone *rtpz;
4461         int tmp, node, zone;
4462
4463         for_each_node_state(node, N_POSSIBLE) {
4464                 tmp = node;
4465                 if (!node_state(node, N_NORMAL_MEMORY))
4466                         tmp = -1;
4467                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4468                 if (!rtpn)
4469                         return 1;
4470
4471                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4472
4473                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4474                         rtpz = &rtpn->rb_tree_per_zone[zone];
4475                         rtpz->rb_root = RB_ROOT;
4476                         spin_lock_init(&rtpz->lock);
4477                 }
4478         }
4479         return 0;
4480 }
4481
4482 static struct cgroup_subsys_state * __ref
4483 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4484 {
4485         struct mem_cgroup *mem, *parent;
4486         long error = -ENOMEM;
4487         int node;
4488
4489         mem = mem_cgroup_alloc();
4490         if (!mem)
4491                 return ERR_PTR(error);
4492
4493         for_each_node_state(node, N_POSSIBLE)
4494                 if (alloc_mem_cgroup_per_zone_info(mem, node))
4495                         goto free_out;
4496
4497         /* root ? */
4498         if (cont->parent == NULL) {
4499                 int cpu;
4500                 enable_swap_cgroup();
4501                 parent = NULL;
4502                 root_mem_cgroup = mem;
4503                 if (mem_cgroup_soft_limit_tree_init())
4504                         goto free_out;
4505                 for_each_possible_cpu(cpu) {
4506                         struct memcg_stock_pcp *stock =
4507                                                 &per_cpu(memcg_stock, cpu);
4508                         INIT_WORK(&stock->work, drain_local_stock);
4509                 }
4510                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4511         } else {
4512                 parent = mem_cgroup_from_cont(cont->parent);
4513                 mem->use_hierarchy = parent->use_hierarchy;
4514                 mem->oom_kill_disable = parent->oom_kill_disable;
4515         }
4516
4517         if (parent && parent->use_hierarchy) {
4518                 res_counter_init(&mem->res, &parent->res);
4519                 res_counter_init(&mem->memsw, &parent->memsw);
4520                 /*
4521                  * We increment refcnt of the parent to ensure that we can
4522                  * safely access it on res_counter_charge/uncharge.
4523                  * This refcnt will be decremented when freeing this
4524                  * mem_cgroup(see mem_cgroup_put).
4525                  */
4526                 mem_cgroup_get(parent);
4527         } else {
4528                 res_counter_init(&mem->res, NULL);
4529                 res_counter_init(&mem->memsw, NULL);
4530         }
4531         mem->last_scanned_child = 0;
4532         spin_lock_init(&mem->reclaim_param_lock);
4533         INIT_LIST_HEAD(&mem->oom_notify);
4534
4535         if (parent)
4536                 mem->swappiness = get_swappiness(parent);
4537         atomic_set(&mem->refcnt, 1);
4538         mem->move_charge_at_immigrate = 0;
4539         mutex_init(&mem->thresholds_lock);
4540         return &mem->css;
4541 free_out:
4542         __mem_cgroup_free(mem);
4543         root_mem_cgroup = NULL;
4544         return ERR_PTR(error);
4545 }
4546
4547 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4548                                         struct cgroup *cont)
4549 {
4550         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4551
4552         return mem_cgroup_force_empty(mem, false);
4553 }
4554
4555 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4556                                 struct cgroup *cont)
4557 {
4558         struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4559
4560         mem_cgroup_put(mem);
4561 }
4562
4563 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4564                                 struct cgroup *cont)
4565 {
4566         int ret;
4567
4568         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4569                                 ARRAY_SIZE(mem_cgroup_files));
4570
4571         if (!ret)
4572                 ret = register_memsw_files(cont, ss);
4573         return ret;
4574 }
4575
4576 #ifdef CONFIG_MMU
4577 /* Handlers for move charge at task migration. */
4578 #define PRECHARGE_COUNT_AT_ONCE 256
4579 static int mem_cgroup_do_precharge(unsigned long count)
4580 {
4581         int ret = 0;
4582         int batch_count = PRECHARGE_COUNT_AT_ONCE;
4583         struct mem_cgroup *mem = mc.to;
4584
4585         if (mem_cgroup_is_root(mem)) {
4586                 mc.precharge += count;
4587                 /* we don't need css_get for root */
4588                 return ret;
4589         }
4590         /* try to charge at once */
4591         if (count > 1) {
4592                 struct res_counter *dummy;
4593                 /*
4594                  * "mem" cannot be under rmdir() because we've already checked
4595                  * by cgroup_lock_live_cgroup() that it is not removed and we
4596                  * are still under the same cgroup_mutex. So we can postpone
4597                  * css_get().
4598                  */
4599                 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4600                         goto one_by_one;
4601                 if (do_swap_account && res_counter_charge(&mem->memsw,
4602                                                 PAGE_SIZE * count, &dummy)) {
4603                         res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4604                         goto one_by_one;
4605                 }
4606                 mc.precharge += count;
4607                 return ret;
4608         }
4609 one_by_one:
4610         /* fall back to one by one charge */
4611         while (count--) {
4612                 if (signal_pending(current)) {
4613                         ret = -EINTR;
4614                         break;
4615                 }
4616                 if (!batch_count--) {
4617                         batch_count = PRECHARGE_COUNT_AT_ONCE;
4618                         cond_resched();
4619                 }
4620                 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4621                                               PAGE_SIZE);
4622                 if (ret || !mem)
4623                         /* mem_cgroup_clear_mc() will do uncharge later */
4624                         return -ENOMEM;
4625                 mc.precharge++;
4626         }
4627         return ret;
4628 }
4629
4630 /**
4631  * is_target_pte_for_mc - check a pte whether it is valid for move charge
4632  * @vma: the vma the pte to be checked belongs
4633  * @addr: the address corresponding to the pte to be checked
4634  * @ptent: the pte to be checked
4635  * @target: the pointer the target page or swap ent will be stored(can be NULL)
4636  *
4637  * Returns
4638  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
4639  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4640  *     move charge. if @target is not NULL, the page is stored in target->page
4641  *     with extra refcnt got(Callers should handle it).
4642  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4643  *     target for charge migration. if @target is not NULL, the entry is stored
4644  *     in target->ent.
4645  *
4646  * Called with pte lock held.
4647  */
4648 union mc_target {
4649         struct page     *page;
4650         swp_entry_t     ent;
4651 };
4652
4653 enum mc_target_type {
4654         MC_TARGET_NONE, /* not used */
4655         MC_TARGET_PAGE,
4656         MC_TARGET_SWAP,
4657 };
4658
4659 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4660                                                 unsigned long addr, pte_t ptent)
4661 {
4662         struct page *page = vm_normal_page(vma, addr, ptent);
4663
4664         if (!page || !page_mapped(page))
4665                 return NULL;
4666         if (PageAnon(page)) {
4667                 /* we don't move shared anon */
4668                 if (!move_anon() || page_mapcount(page) > 2)
4669                         return NULL;
4670         } else if (!move_file())
4671                 /* we ignore mapcount for file pages */
4672                 return NULL;
4673         if (!get_page_unless_zero(page))
4674                 return NULL;
4675
4676         return page;
4677 }
4678
4679 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4680                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4681 {
4682         int usage_count;
4683         struct page *page = NULL;
4684         swp_entry_t ent = pte_to_swp_entry(ptent);
4685
4686         if (!move_anon() || non_swap_entry(ent))
4687                 return NULL;
4688         usage_count = mem_cgroup_count_swap_user(ent, &page);
4689         if (usage_count > 1) { /* we don't move shared anon */
4690                 if (page)
4691                         put_page(page);
4692                 return NULL;
4693         }
4694         if (do_swap_account)
4695                 entry->val = ent.val;
4696
4697         return page;
4698 }
4699
4700 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4701                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
4702 {
4703         struct page *page = NULL;
4704         struct inode *inode;
4705         struct address_space *mapping;
4706         pgoff_t pgoff;
4707
4708         if (!vma->vm_file) /* anonymous vma */
4709                 return NULL;
4710         if (!move_file())
4711                 return NULL;
4712
4713         inode = vma->vm_file->f_path.dentry->d_inode;
4714         mapping = vma->vm_file->f_mapping;
4715         if (pte_none(ptent))
4716                 pgoff = linear_page_index(vma, addr);
4717         else /* pte_file(ptent) is true */
4718                 pgoff = pte_to_pgoff(ptent);
4719
4720         /* page is moved even if it's not RSS of this task(page-faulted). */
4721         if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4722                 page = find_get_page(mapping, pgoff);
4723         } else { /* shmem/tmpfs file. we should take account of swap too. */
4724                 swp_entry_t ent;
4725                 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4726                 if (do_swap_account)
4727                         entry->val = ent.val;
4728         }
4729
4730         return page;
4731 }
4732
4733 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4734                 unsigned long addr, pte_t ptent, union mc_target *target)
4735 {
4736         struct page *page = NULL;
4737         struct page_cgroup *pc;
4738         int ret = 0;
4739         swp_entry_t ent = { .val = 0 };
4740
4741         if (pte_present(ptent))
4742                 page = mc_handle_present_pte(vma, addr, ptent);
4743         else if (is_swap_pte(ptent))
4744                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4745         else if (pte_none(ptent) || pte_file(ptent))
4746                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4747
4748         if (!page && !ent.val)
4749                 return 0;
4750         if (page) {
4751                 pc = lookup_page_cgroup(page);
4752                 /*
4753                  * Do only loose check w/o page_cgroup lock.
4754                  * mem_cgroup_move_account() checks the pc is valid or not under
4755                  * the lock.
4756                  */
4757                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4758                         ret = MC_TARGET_PAGE;
4759                         if (target)
4760                                 target->page = page;
4761                 }
4762                 if (!ret || !target)
4763                         put_page(page);
4764         }
4765         /* There is a swap entry and a page doesn't exist or isn't charged */
4766         if (ent.val && !ret &&
4767                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4768                 ret = MC_TARGET_SWAP;
4769                 if (target)
4770                         target->ent = ent;
4771         }
4772         return ret;
4773 }
4774
4775 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4776                                         unsigned long addr, unsigned long end,
4777                                         struct mm_walk *walk)
4778 {
4779         struct vm_area_struct *vma = walk->private;
4780         pte_t *pte;
4781         spinlock_t *ptl;
4782
4783         split_huge_page_pmd(walk->mm, pmd);
4784
4785         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4786         for (; addr != end; pte++, addr += PAGE_SIZE)
4787                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4788                         mc.precharge++; /* increment precharge temporarily */
4789         pte_unmap_unlock(pte - 1, ptl);
4790         cond_resched();
4791
4792         return 0;
4793 }
4794
4795 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4796 {
4797         unsigned long precharge;
4798         struct vm_area_struct *vma;
4799
4800         down_read(&mm->mmap_sem);
4801         for (vma = mm->mmap; vma; vma = vma->vm_next) {
4802                 struct mm_walk mem_cgroup_count_precharge_walk = {
4803                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
4804                         .mm = mm,
4805                         .private = vma,
4806                 };
4807                 if (is_vm_hugetlb_page(vma))
4808                         continue;
4809                 walk_page_range(vma->vm_start, vma->vm_end,
4810                                         &mem_cgroup_count_precharge_walk);
4811         }
4812         up_read(&mm->mmap_sem);
4813
4814         precharge = mc.precharge;
4815         mc.precharge = 0;
4816
4817         return precharge;
4818 }
4819
4820 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4821 {
4822         unsigned long precharge = mem_cgroup_count_precharge(mm);
4823
4824         VM_BUG_ON(mc.moving_task);
4825         mc.moving_task = current;
4826         return mem_cgroup_do_precharge(precharge);
4827 }
4828
4829 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4830 static void __mem_cgroup_clear_mc(void)
4831 {
4832         struct mem_cgroup *from = mc.from;
4833         struct mem_cgroup *to = mc.to;
4834
4835         /* we must uncharge all the leftover precharges from mc.to */
4836         if (mc.precharge) {
4837                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4838                 mc.precharge = 0;
4839         }
4840         /*
4841          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4842          * we must uncharge here.
4843          */
4844         if (mc.moved_charge) {
4845                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4846                 mc.moved_charge = 0;
4847         }
4848         /* we must fixup refcnts and charges */
4849         if (mc.moved_swap) {
4850                 /* uncharge swap account from the old cgroup */
4851                 if (!mem_cgroup_is_root(mc.from))
4852                         res_counter_uncharge(&mc.from->memsw,
4853                                                 PAGE_SIZE * mc.moved_swap);
4854                 __mem_cgroup_put(mc.from, mc.moved_swap);
4855
4856                 if (!mem_cgroup_is_root(mc.to)) {
4857                         /*
4858                          * we charged both to->res and to->memsw, so we should
4859                          * uncharge to->res.
4860                          */
4861                         res_counter_uncharge(&mc.to->res,
4862                                                 PAGE_SIZE * mc.moved_swap);
4863                 }
4864                 /* we've already done mem_cgroup_get(mc.to) */
4865                 mc.moved_swap = 0;
4866         }
4867         memcg_oom_recover(from);
4868         memcg_oom_recover(to);
4869         wake_up_all(&mc.waitq);
4870 }
4871
4872 static void mem_cgroup_clear_mc(void)
4873 {
4874         struct mem_cgroup *from = mc.from;
4875
4876         /*
4877          * we must clear moving_task before waking up waiters at the end of
4878          * task migration.
4879          */
4880         mc.moving_task = NULL;
4881         __mem_cgroup_clear_mc();
4882         spin_lock(&mc.lock);
4883         mc.from = NULL;
4884         mc.to = NULL;
4885         spin_unlock(&mc.lock);
4886         mem_cgroup_end_move(from);
4887 }
4888
4889 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4890                                 struct cgroup *cgroup,
4891                                 struct task_struct *p,
4892                                 bool threadgroup)
4893 {
4894         int ret = 0;
4895         struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4896
4897         if (mem->move_charge_at_immigrate) {
4898                 struct mm_struct *mm;
4899                 struct mem_cgroup *from = mem_cgroup_from_task(p);
4900
4901                 VM_BUG_ON(from == mem);
4902
4903                 mm = get_task_mm(p);
4904                 if (!mm)
4905                         return 0;
4906                 /* We move charges only when we move a owner of the mm */
4907                 if (mm->owner == p) {
4908                         VM_BUG_ON(mc.from);
4909                         VM_BUG_ON(mc.to);
4910                         VM_BUG_ON(mc.precharge);
4911                         VM_BUG_ON(mc.moved_charge);
4912                         VM_BUG_ON(mc.moved_swap);
4913                         mem_cgroup_start_move(from);
4914                         spin_lock(&mc.lock);
4915                         mc.from = from;
4916                         mc.to = mem;
4917                         spin_unlock(&mc.lock);
4918                         /* We set mc.moving_task later */
4919
4920                         ret = mem_cgroup_precharge_mc(mm);
4921                         if (ret)
4922                                 mem_cgroup_clear_mc();
4923                 }
4924                 mmput(mm);
4925         }
4926         return ret;
4927 }
4928
4929 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4930                                 struct cgroup *cgroup,
4931                                 struct task_struct *p,
4932                                 bool threadgroup)
4933 {
4934         mem_cgroup_clear_mc();
4935 }
4936
4937 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4938                                 unsigned long addr, unsigned long end,
4939                                 struct mm_walk *walk)
4940 {
4941         int ret = 0;
4942         struct vm_area_struct *vma = walk->private;
4943         pte_t *pte;
4944         spinlock_t *ptl;
4945
4946         split_huge_page_pmd(walk->mm, pmd);
4947 retry:
4948         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4949         for (; addr != end; addr += PAGE_SIZE) {
4950                 pte_t ptent = *(pte++);
4951                 union mc_target target;
4952                 int type;
4953                 struct page *page;
4954                 struct page_cgroup *pc;
4955                 swp_entry_t ent;
4956
4957                 if (!mc.precharge)
4958                         break;
4959
4960                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4961                 switch (type) {
4962                 case MC_TARGET_PAGE:
4963                         page = target.page;
4964                         if (isolate_lru_page(page))
4965                                 goto put;
4966                         pc = lookup_page_cgroup(page);
4967                         if (!mem_cgroup_move_account(page, pc,
4968                                         mc.from, mc.to, false, PAGE_SIZE)) {
4969                                 mc.precharge--;
4970                                 /* we uncharge from mc.from later. */
4971                                 mc.moved_charge++;
4972                         }
4973                         putback_lru_page(page);
4974 put:                    /* is_target_pte_for_mc() gets the page */
4975                         put_page(page);
4976                         break;
4977                 case MC_TARGET_SWAP:
4978                         ent = target.ent;
4979                         if (!mem_cgroup_move_swap_account(ent,
4980                                                 mc.from, mc.to, false)) {
4981                                 mc.precharge--;
4982                                 /* we fixup refcnts and charges later. */
4983                                 mc.moved_swap++;
4984                         }
4985                         break;
4986                 default:
4987                         break;
4988                 }
4989         }
4990         pte_unmap_unlock(pte - 1, ptl);
4991         cond_resched();
4992
4993         if (addr != end) {
4994                 /*
4995                  * We have consumed all precharges we got in can_attach().
4996                  * We try charge one by one, but don't do any additional
4997                  * charges to mc.to if we have failed in charge once in attach()
4998                  * phase.
4999                  */
5000                 ret = mem_cgroup_do_precharge(1);
5001                 if (!ret)
5002                         goto retry;
5003         }
5004
5005         return ret;
5006 }
5007
5008 static void mem_cgroup_move_charge(struct mm_struct *mm)
5009 {
5010         struct vm_area_struct *vma;
5011
5012         lru_add_drain_all();
5013 retry:
5014         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5015                 /*
5016                  * Someone who are holding the mmap_sem might be waiting in
5017                  * waitq. So we cancel all extra charges, wake up all waiters,
5018                  * and retry. Because we cancel precharges, we might not be able
5019                  * to move enough charges, but moving charge is a best-effort
5020                  * feature anyway, so it wouldn't be a big problem.
5021                  */
5022                 __mem_cgroup_clear_mc();
5023                 cond_resched();
5024                 goto retry;
5025         }
5026         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5027                 int ret;
5028                 struct mm_walk mem_cgroup_move_charge_walk = {
5029                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5030                         .mm = mm,
5031                         .private = vma,
5032                 };
5033                 if (is_vm_hugetlb_page(vma))
5034                         continue;
5035                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5036                                                 &mem_cgroup_move_charge_walk);
5037                 if (ret)
5038                         /*
5039                          * means we have consumed all precharges and failed in
5040                          * doing additional charge. Just abandon here.
5041                          */
5042                         break;
5043         }
5044         up_read(&mm->mmap_sem);
5045 }
5046
5047 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5048                                 struct cgroup *cont,
5049                                 struct cgroup *old_cont,
5050                                 struct task_struct *p,
5051                                 bool threadgroup)
5052 {
5053         struct mm_struct *mm;
5054
5055         if (!mc.to)
5056                 /* no need to move charge */
5057                 return;
5058
5059         mm = get_task_mm(p);
5060         if (mm) {
5061                 mem_cgroup_move_charge(mm);
5062                 mmput(mm);
5063         }
5064         mem_cgroup_clear_mc();
5065 }
5066 #else   /* !CONFIG_MMU */
5067 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5068                                 struct cgroup *cgroup,
5069                                 struct task_struct *p,
5070                                 bool threadgroup)
5071 {
5072         return 0;
5073 }
5074 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5075                                 struct cgroup *cgroup,
5076                                 struct task_struct *p,
5077                                 bool threadgroup)
5078 {
5079 }
5080 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5081                                 struct cgroup *cont,
5082                                 struct cgroup *old_cont,
5083                                 struct task_struct *p,
5084                                 bool threadgroup)
5085 {
5086 }
5087 #endif
5088
5089 struct cgroup_subsys mem_cgroup_subsys = {
5090         .name = "memory",
5091         .subsys_id = mem_cgroup_subsys_id,
5092         .create = mem_cgroup_create,
5093         .pre_destroy = mem_cgroup_pre_destroy,
5094         .destroy = mem_cgroup_destroy,
5095         .populate = mem_cgroup_populate,
5096         .can_attach = mem_cgroup_can_attach,
5097         .cancel_attach = mem_cgroup_cancel_attach,
5098         .attach = mem_cgroup_move_task,
5099         .early_init = 0,
5100         .use_id = 1,
5101 };
5102
5103 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5104 static int __init enable_swap_account(char *s)
5105 {
5106         /* consider enabled if no parameter or 1 is given */
5107         if (!(*s) || !strcmp(s, "=1"))
5108                 really_do_swap_account = 1;
5109         else if (!strcmp(s, "=0"))
5110                 really_do_swap_account = 0;
5111         return 1;
5112 }
5113 __setup("swapaccount", enable_swap_account);
5114
5115 static int __init disable_swap_account(char *s)
5116 {
5117         printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5118         enable_swap_account("=0");
5119         return 1;
5120 }
5121 __setup("noswapaccount", disable_swap_account);
5122 #endif