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