0b2d4036f1cde795457b8e47794dac24174bad6c
[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/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
45 #include <linux/fs.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
52 #include "internal.h"
53 #include <net/sock.h>
54 #include <net/tcp_memcontrol.h>
55
56 #include <asm/uaccess.h>
57
58 #include <trace/events/vmscan.h>
59
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES      5
62 struct mem_cgroup *root_mem_cgroup __read_mostly;
63
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
67
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
71 #else
72 static int really_do_swap_account __initdata = 0;
73 #endif
74
75 #else
76 #define do_swap_account         (0)
77 #endif
78
79
80 /*
81  * Statistics for memory cgroup.
82  */
83 enum mem_cgroup_stat_index {
84         /*
85          * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
86          */
87         MEM_CGROUP_STAT_CACHE,     /* # of pages charged as cache */
88         MEM_CGROUP_STAT_RSS,       /* # of pages charged as anon rss */
89         MEM_CGROUP_STAT_FILE_MAPPED,  /* # of pages charged as file rss */
90         MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91         MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92         MEM_CGROUP_ON_MOVE,     /* someone is moving account between groups */
93         MEM_CGROUP_STAT_NSTATS,
94 };
95
96 enum mem_cgroup_events_index {
97         MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
98         MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
99         MEM_CGROUP_EVENTS_COUNT,        /* # of pages paged in/out */
100         MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
101         MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
102         MEM_CGROUP_EVENTS_NSTATS,
103 };
104 /*
105  * Per memcg event counter is incremented at every pagein/pageout. With THP,
106  * it will be incremated by the number of pages. This counter is used for
107  * for trigger some periodic events. This is straightforward and better
108  * than using jiffies etc. to handle periodic memcg event.
109  */
110 enum mem_cgroup_events_target {
111         MEM_CGROUP_TARGET_THRESH,
112         MEM_CGROUP_TARGET_SOFTLIMIT,
113         MEM_CGROUP_TARGET_NUMAINFO,
114         MEM_CGROUP_NTARGETS,
115 };
116 #define THRESHOLDS_EVENTS_TARGET (128)
117 #define SOFTLIMIT_EVENTS_TARGET (1024)
118 #define NUMAINFO_EVENTS_TARGET  (1024)
119
120 struct mem_cgroup_stat_cpu {
121         long count[MEM_CGROUP_STAT_NSTATS];
122         unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
123         unsigned long targets[MEM_CGROUP_NTARGETS];
124 };
125
126 /*
127  * per-zone information in memory controller.
128  */
129 struct mem_cgroup_per_zone {
130         /*
131          * spin_lock to protect the per cgroup LRU
132          */
133         struct list_head        lists[NR_LRU_LISTS];
134         unsigned long           count[NR_LRU_LISTS];
135
136         struct zone_reclaim_stat reclaim_stat;
137         struct rb_node          tree_node;      /* RB tree node */
138         unsigned long long      usage_in_excess;/* Set to the value by which */
139                                                 /* the soft limit is exceeded*/
140         bool                    on_tree;
141         struct mem_cgroup       *mem;           /* Back pointer, we cannot */
142                                                 /* use container_of        */
143 };
144 /* Macro for accessing counter */
145 #define MEM_CGROUP_ZSTAT(mz, idx)       ((mz)->count[(idx)])
146
147 struct mem_cgroup_per_node {
148         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 };
150
151 struct mem_cgroup_lru_info {
152         struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
153 };
154
155 /*
156  * Cgroups above their limits are maintained in a RB-Tree, independent of
157  * their hierarchy representation
158  */
159
160 struct mem_cgroup_tree_per_zone {
161         struct rb_root rb_root;
162         spinlock_t lock;
163 };
164
165 struct mem_cgroup_tree_per_node {
166         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 };
168
169 struct mem_cgroup_tree {
170         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 };
172
173 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
174
175 struct mem_cgroup_threshold {
176         struct eventfd_ctx *eventfd;
177         u64 threshold;
178 };
179
180 /* For threshold */
181 struct mem_cgroup_threshold_ary {
182         /* An array index points to threshold just below usage. */
183         int current_threshold;
184         /* Size of entries[] */
185         unsigned int size;
186         /* Array of thresholds */
187         struct mem_cgroup_threshold entries[0];
188 };
189
190 struct mem_cgroup_thresholds {
191         /* Primary thresholds array */
192         struct mem_cgroup_threshold_ary *primary;
193         /*
194          * Spare threshold array.
195          * This is needed to make mem_cgroup_unregister_event() "never fail".
196          * It must be able to store at least primary->size - 1 entries.
197          */
198         struct mem_cgroup_threshold_ary *spare;
199 };
200
201 /* for OOM */
202 struct mem_cgroup_eventfd_list {
203         struct list_head list;
204         struct eventfd_ctx *eventfd;
205 };
206
207 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
208 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
209
210 /*
211  * The memory controller data structure. The memory controller controls both
212  * page cache and RSS per cgroup. We would eventually like to provide
213  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
214  * to help the administrator determine what knobs to tune.
215  *
216  * TODO: Add a water mark for the memory controller. Reclaim will begin when
217  * we hit the water mark. May be even add a low water mark, such that
218  * no reclaim occurs from a cgroup at it's low water mark, this is
219  * a feature that will be implemented much later in the future.
220  */
221 struct mem_cgroup {
222         struct cgroup_subsys_state css;
223         /*
224          * the counter to account for memory usage
225          */
226         struct res_counter res;
227         /*
228          * the counter to account for mem+swap usage.
229          */
230         struct res_counter memsw;
231         /*
232          * Per cgroup active and inactive list, similar to the
233          * per zone LRU lists.
234          */
235         struct mem_cgroup_lru_info info;
236         /*
237          * While reclaiming in a hierarchy, we cache the last child we
238          * reclaimed from.
239          */
240         int last_scanned_child;
241         int last_scanned_node;
242 #if MAX_NUMNODES > 1
243         nodemask_t      scan_nodes;
244         atomic_t        numainfo_events;
245         atomic_t        numainfo_updating;
246 #endif
247         /*
248          * Should the accounting and control be hierarchical, per subtree?
249          */
250         bool use_hierarchy;
251
252         bool            oom_lock;
253         atomic_t        under_oom;
254
255         atomic_t        refcnt;
256
257         int     swappiness;
258         /* OOM-Killer disable */
259         int             oom_kill_disable;
260
261         /* set when res.limit == memsw.limit */
262         bool            memsw_is_minimum;
263
264         /* protect arrays of thresholds */
265         struct mutex thresholds_lock;
266
267         /* thresholds for memory usage. RCU-protected */
268         struct mem_cgroup_thresholds thresholds;
269
270         /* thresholds for mem+swap usage. RCU-protected */
271         struct mem_cgroup_thresholds memsw_thresholds;
272
273         /* For oom notifier event fd */
274         struct list_head oom_notify;
275
276         /*
277          * Should we move charges of a task when a task is moved into this
278          * mem_cgroup ? And what type of charges should we move ?
279          */
280         unsigned long   move_charge_at_immigrate;
281         /*
282          * percpu counter.
283          */
284         struct mem_cgroup_stat_cpu *stat;
285         /*
286          * used when a cpu is offlined or other synchronizations
287          * See mem_cgroup_read_stat().
288          */
289         struct mem_cgroup_stat_cpu nocpu_base;
290         spinlock_t pcp_counter_lock;
291
292 #ifdef CONFIG_INET
293         struct tcp_memcontrol tcp_mem;
294 #endif
295 };
296
297 /* Stuffs for move charges at task migration. */
298 /*
299  * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
300  * left-shifted bitmap of these types.
301  */
302 enum move_type {
303         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
304         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
305         NR_MOVE_TYPE,
306 };
307
308 /* "mc" and its members are protected by cgroup_mutex */
309 static struct move_charge_struct {
310         spinlock_t        lock; /* for from, to */
311         struct mem_cgroup *from;
312         struct mem_cgroup *to;
313         unsigned long precharge;
314         unsigned long moved_charge;
315         unsigned long moved_swap;
316         struct task_struct *moving_task;        /* a task moving charges */
317         wait_queue_head_t waitq;                /* a waitq for other context */
318 } mc = {
319         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
320         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
321 };
322
323 static bool move_anon(void)
324 {
325         return test_bit(MOVE_CHARGE_TYPE_ANON,
326                                         &mc.to->move_charge_at_immigrate);
327 }
328
329 static bool move_file(void)
330 {
331         return test_bit(MOVE_CHARGE_TYPE_FILE,
332                                         &mc.to->move_charge_at_immigrate);
333 }
334
335 /*
336  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
337  * limit reclaim to prevent infinite loops, if they ever occur.
338  */
339 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            (100)
340 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
341
342 enum charge_type {
343         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
344         MEM_CGROUP_CHARGE_TYPE_MAPPED,
345         MEM_CGROUP_CHARGE_TYPE_SHMEM,   /* used by page migration of shmem */
346         MEM_CGROUP_CHARGE_TYPE_FORCE,   /* used by force_empty */
347         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
348         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
349         NR_CHARGE_TYPE,
350 };
351
352 /* for encoding cft->private value on file */
353 #define _MEM                    (0)
354 #define _MEMSWAP                (1)
355 #define _OOM_TYPE               (2)
356 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
357 #define MEMFILE_TYPE(val)       (((val) >> 16) & 0xffff)
358 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
359 /* Used for OOM nofiier */
360 #define OOM_CONTROL             (0)
361
362 /*
363  * Reclaim flags for mem_cgroup_hierarchical_reclaim
364  */
365 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
366 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
367 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
368 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
369 #define MEM_CGROUP_RECLAIM_SOFT_BIT     0x2
370 #define MEM_CGROUP_RECLAIM_SOFT         (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
371
372 static void mem_cgroup_get(struct mem_cgroup *memcg);
373 static void mem_cgroup_put(struct mem_cgroup *memcg);
374
375 /* Writing them here to avoid exposing memcg's inner layout */
376 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
377 #ifdef CONFIG_INET
378 #include <net/sock.h>
379 #include <net/ip.h>
380
381 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
382 void sock_update_memcg(struct sock *sk)
383 {
384         if (static_branch(&memcg_socket_limit_enabled)) {
385                 struct mem_cgroup *memcg;
386
387                 BUG_ON(!sk->sk_prot->proto_cgroup);
388
389                 /* Socket cloning can throw us here with sk_cgrp already
390                  * filled. It won't however, necessarily happen from
391                  * process context. So the test for root memcg given
392                  * the current task's memcg won't help us in this case.
393                  *
394                  * Respecting the original socket's memcg is a better
395                  * decision in this case.
396                  */
397                 if (sk->sk_cgrp) {
398                         BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
399                         mem_cgroup_get(sk->sk_cgrp->memcg);
400                         return;
401                 }
402
403                 rcu_read_lock();
404                 memcg = mem_cgroup_from_task(current);
405                 if (!mem_cgroup_is_root(memcg)) {
406                         mem_cgroup_get(memcg);
407                         sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
408                 }
409                 rcu_read_unlock();
410         }
411 }
412 EXPORT_SYMBOL(sock_update_memcg);
413
414 void sock_release_memcg(struct sock *sk)
415 {
416         if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
417                 struct mem_cgroup *memcg;
418                 WARN_ON(!sk->sk_cgrp->memcg);
419                 memcg = sk->sk_cgrp->memcg;
420                 mem_cgroup_put(memcg);
421         }
422 }
423
424 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
425 {
426         if (!memcg || mem_cgroup_is_root(memcg))
427                 return NULL;
428
429         return &memcg->tcp_mem.cg_proto;
430 }
431 EXPORT_SYMBOL(tcp_proto_cgroup);
432 #endif /* CONFIG_INET */
433 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
434
435 static void drain_all_stock_async(struct mem_cgroup *memcg);
436
437 static struct mem_cgroup_per_zone *
438 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
439 {
440         return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
441 }
442
443 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
444 {
445         return &memcg->css;
446 }
447
448 static struct mem_cgroup_per_zone *
449 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
450 {
451         int nid = page_to_nid(page);
452         int zid = page_zonenum(page);
453
454         return mem_cgroup_zoneinfo(memcg, nid, zid);
455 }
456
457 static struct mem_cgroup_tree_per_zone *
458 soft_limit_tree_node_zone(int nid, int zid)
459 {
460         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
461 }
462
463 static struct mem_cgroup_tree_per_zone *
464 soft_limit_tree_from_page(struct page *page)
465 {
466         int nid = page_to_nid(page);
467         int zid = page_zonenum(page);
468
469         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
470 }
471
472 static void
473 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
474                                 struct mem_cgroup_per_zone *mz,
475                                 struct mem_cgroup_tree_per_zone *mctz,
476                                 unsigned long long new_usage_in_excess)
477 {
478         struct rb_node **p = &mctz->rb_root.rb_node;
479         struct rb_node *parent = NULL;
480         struct mem_cgroup_per_zone *mz_node;
481
482         if (mz->on_tree)
483                 return;
484
485         mz->usage_in_excess = new_usage_in_excess;
486         if (!mz->usage_in_excess)
487                 return;
488         while (*p) {
489                 parent = *p;
490                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
491                                         tree_node);
492                 if (mz->usage_in_excess < mz_node->usage_in_excess)
493                         p = &(*p)->rb_left;
494                 /*
495                  * We can't avoid mem cgroups that are over their soft
496                  * limit by the same amount
497                  */
498                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
499                         p = &(*p)->rb_right;
500         }
501         rb_link_node(&mz->tree_node, parent, p);
502         rb_insert_color(&mz->tree_node, &mctz->rb_root);
503         mz->on_tree = true;
504 }
505
506 static void
507 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
508                                 struct mem_cgroup_per_zone *mz,
509                                 struct mem_cgroup_tree_per_zone *mctz)
510 {
511         if (!mz->on_tree)
512                 return;
513         rb_erase(&mz->tree_node, &mctz->rb_root);
514         mz->on_tree = false;
515 }
516
517 static void
518 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
519                                 struct mem_cgroup_per_zone *mz,
520                                 struct mem_cgroup_tree_per_zone *mctz)
521 {
522         spin_lock(&mctz->lock);
523         __mem_cgroup_remove_exceeded(memcg, mz, mctz);
524         spin_unlock(&mctz->lock);
525 }
526
527
528 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
529 {
530         unsigned long long excess;
531         struct mem_cgroup_per_zone *mz;
532         struct mem_cgroup_tree_per_zone *mctz;
533         int nid = page_to_nid(page);
534         int zid = page_zonenum(page);
535         mctz = soft_limit_tree_from_page(page);
536
537         /*
538          * Necessary to update all ancestors when hierarchy is used.
539          * because their event counter is not touched.
540          */
541         for (; memcg; memcg = parent_mem_cgroup(memcg)) {
542                 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
543                 excess = res_counter_soft_limit_excess(&memcg->res);
544                 /*
545                  * We have to update the tree if mz is on RB-tree or
546                  * mem is over its softlimit.
547                  */
548                 if (excess || mz->on_tree) {
549                         spin_lock(&mctz->lock);
550                         /* if on-tree, remove it */
551                         if (mz->on_tree)
552                                 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
553                         /*
554                          * Insert again. mz->usage_in_excess will be updated.
555                          * If excess is 0, no tree ops.
556                          */
557                         __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
558                         spin_unlock(&mctz->lock);
559                 }
560         }
561 }
562
563 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
564 {
565         int node, zone;
566         struct mem_cgroup_per_zone *mz;
567         struct mem_cgroup_tree_per_zone *mctz;
568
569         for_each_node_state(node, N_POSSIBLE) {
570                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
571                         mz = mem_cgroup_zoneinfo(memcg, node, zone);
572                         mctz = soft_limit_tree_node_zone(node, zone);
573                         mem_cgroup_remove_exceeded(memcg, mz, mctz);
574                 }
575         }
576 }
577
578 static struct mem_cgroup_per_zone *
579 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
580 {
581         struct rb_node *rightmost = NULL;
582         struct mem_cgroup_per_zone *mz;
583
584 retry:
585         mz = NULL;
586         rightmost = rb_last(&mctz->rb_root);
587         if (!rightmost)
588                 goto done;              /* Nothing to reclaim from */
589
590         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
591         /*
592          * Remove the node now but someone else can add it back,
593          * we will to add it back at the end of reclaim to its correct
594          * position in the tree.
595          */
596         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
597         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
598                 !css_tryget(&mz->mem->css))
599                 goto retry;
600 done:
601         return mz;
602 }
603
604 static struct mem_cgroup_per_zone *
605 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
606 {
607         struct mem_cgroup_per_zone *mz;
608
609         spin_lock(&mctz->lock);
610         mz = __mem_cgroup_largest_soft_limit_node(mctz);
611         spin_unlock(&mctz->lock);
612         return mz;
613 }
614
615 /*
616  * Implementation Note: reading percpu statistics for memcg.
617  *
618  * Both of vmstat[] and percpu_counter has threshold and do periodic
619  * synchronization to implement "quick" read. There are trade-off between
620  * reading cost and precision of value. Then, we may have a chance to implement
621  * a periodic synchronizion of counter in memcg's counter.
622  *
623  * But this _read() function is used for user interface now. The user accounts
624  * memory usage by memory cgroup and he _always_ requires exact value because
625  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
626  * have to visit all online cpus and make sum. So, for now, unnecessary
627  * synchronization is not implemented. (just implemented for cpu hotplug)
628  *
629  * If there are kernel internal actions which can make use of some not-exact
630  * value, and reading all cpu value can be performance bottleneck in some
631  * common workload, threashold and synchonization as vmstat[] should be
632  * implemented.
633  */
634 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
635                                  enum mem_cgroup_stat_index idx)
636 {
637         long val = 0;
638         int cpu;
639
640         get_online_cpus();
641         for_each_online_cpu(cpu)
642                 val += per_cpu(memcg->stat->count[idx], cpu);
643 #ifdef CONFIG_HOTPLUG_CPU
644         spin_lock(&memcg->pcp_counter_lock);
645         val += memcg->nocpu_base.count[idx];
646         spin_unlock(&memcg->pcp_counter_lock);
647 #endif
648         put_online_cpus();
649         return val;
650 }
651
652 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
653                                          bool charge)
654 {
655         int val = (charge) ? 1 : -1;
656         this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
657 }
658
659 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
660 {
661         this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
662 }
663
664 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
665 {
666         this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
667 }
668
669 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
670                                             enum mem_cgroup_events_index idx)
671 {
672         unsigned long val = 0;
673         int cpu;
674
675         for_each_online_cpu(cpu)
676                 val += per_cpu(memcg->stat->events[idx], cpu);
677 #ifdef CONFIG_HOTPLUG_CPU
678         spin_lock(&memcg->pcp_counter_lock);
679         val += memcg->nocpu_base.events[idx];
680         spin_unlock(&memcg->pcp_counter_lock);
681 #endif
682         return val;
683 }
684
685 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
686                                          bool file, int nr_pages)
687 {
688         preempt_disable();
689
690         if (file)
691                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
692                                 nr_pages);
693         else
694                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
695                                 nr_pages);
696
697         /* pagein of a big page is an event. So, ignore page size */
698         if (nr_pages > 0)
699                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
700         else {
701                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
702                 nr_pages = -nr_pages; /* for event */
703         }
704
705         __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
706
707         preempt_enable();
708 }
709
710 unsigned long
711 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
712                         unsigned int lru_mask)
713 {
714         struct mem_cgroup_per_zone *mz;
715         enum lru_list l;
716         unsigned long ret = 0;
717
718         mz = mem_cgroup_zoneinfo(memcg, nid, zid);
719
720         for_each_lru(l) {
721                 if (BIT(l) & lru_mask)
722                         ret += MEM_CGROUP_ZSTAT(mz, l);
723         }
724         return ret;
725 }
726
727 static unsigned long
728 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
729                         int nid, unsigned int lru_mask)
730 {
731         u64 total = 0;
732         int zid;
733
734         for (zid = 0; zid < MAX_NR_ZONES; zid++)
735                 total += mem_cgroup_zone_nr_lru_pages(memcg,
736                                                 nid, zid, lru_mask);
737
738         return total;
739 }
740
741 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
742                         unsigned int lru_mask)
743 {
744         int nid;
745         u64 total = 0;
746
747         for_each_node_state(nid, N_HIGH_MEMORY)
748                 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
749         return total;
750 }
751
752 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
753 {
754         unsigned long val, next;
755
756         val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
757         next = __this_cpu_read(memcg->stat->targets[target]);
758         /* from time_after() in jiffies.h */
759         return ((long)next - (long)val < 0);
760 }
761
762 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
763 {
764         unsigned long val, next;
765
766         val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
767
768         switch (target) {
769         case MEM_CGROUP_TARGET_THRESH:
770                 next = val + THRESHOLDS_EVENTS_TARGET;
771                 break;
772         case MEM_CGROUP_TARGET_SOFTLIMIT:
773                 next = val + SOFTLIMIT_EVENTS_TARGET;
774                 break;
775         case MEM_CGROUP_TARGET_NUMAINFO:
776                 next = val + NUMAINFO_EVENTS_TARGET;
777                 break;
778         default:
779                 return;
780         }
781
782         __this_cpu_write(memcg->stat->targets[target], next);
783 }
784
785 /*
786  * Check events in order.
787  *
788  */
789 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
790 {
791         preempt_disable();
792         /* threshold event is triggered in finer grain than soft limit */
793         if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
794                 mem_cgroup_threshold(memcg);
795                 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
796                 if (unlikely(__memcg_event_check(memcg,
797                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
798                         mem_cgroup_update_tree(memcg, page);
799                         __mem_cgroup_target_update(memcg,
800                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
801                 }
802 #if MAX_NUMNODES > 1
803                 if (unlikely(__memcg_event_check(memcg,
804                         MEM_CGROUP_TARGET_NUMAINFO))) {
805                         atomic_inc(&memcg->numainfo_events);
806                         __mem_cgroup_target_update(memcg,
807                                 MEM_CGROUP_TARGET_NUMAINFO);
808                 }
809 #endif
810         }
811         preempt_enable();
812 }
813
814 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
815 {
816         return container_of(cgroup_subsys_state(cont,
817                                 mem_cgroup_subsys_id), struct mem_cgroup,
818                                 css);
819 }
820
821 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
822 {
823         /*
824          * mm_update_next_owner() may clear mm->owner to NULL
825          * if it races with swapoff, page migration, etc.
826          * So this can be called with p == NULL.
827          */
828         if (unlikely(!p))
829                 return NULL;
830
831         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
832                                 struct mem_cgroup, css);
833 }
834
835 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
836 {
837         struct mem_cgroup *memcg = NULL;
838
839         if (!mm)
840                 return NULL;
841         /*
842          * Because we have no locks, mm->owner's may be being moved to other
843          * cgroup. We use css_tryget() here even if this looks
844          * pessimistic (rather than adding locks here).
845          */
846         rcu_read_lock();
847         do {
848                 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
849                 if (unlikely(!memcg))
850                         break;
851         } while (!css_tryget(&memcg->css));
852         rcu_read_unlock();
853         return memcg;
854 }
855
856 /* The caller has to guarantee "mem" exists before calling this */
857 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
858 {
859         struct cgroup_subsys_state *css;
860         int found;
861
862         if (!memcg) /* ROOT cgroup has the smallest ID */
863                 return root_mem_cgroup; /*css_put/get against root is ignored*/
864         if (!memcg->use_hierarchy) {
865                 if (css_tryget(&memcg->css))
866                         return memcg;
867                 return NULL;
868         }
869         rcu_read_lock();
870         /*
871          * searching a memory cgroup which has the smallest ID under given
872          * ROOT cgroup. (ID >= 1)
873          */
874         css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
875         if (css && css_tryget(css))
876                 memcg = container_of(css, struct mem_cgroup, css);
877         else
878                 memcg = NULL;
879         rcu_read_unlock();
880         return memcg;
881 }
882
883 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
884                                         struct mem_cgroup *root,
885                                         bool cond)
886 {
887         int nextid = css_id(&iter->css) + 1;
888         int found;
889         int hierarchy_used;
890         struct cgroup_subsys_state *css;
891
892         hierarchy_used = iter->use_hierarchy;
893
894         css_put(&iter->css);
895         /* If no ROOT, walk all, ignore hierarchy */
896         if (!cond || (root && !hierarchy_used))
897                 return NULL;
898
899         if (!root)
900                 root = root_mem_cgroup;
901
902         do {
903                 iter = NULL;
904                 rcu_read_lock();
905
906                 css = css_get_next(&mem_cgroup_subsys, nextid,
907                                 &root->css, &found);
908                 if (css && css_tryget(css))
909                         iter = container_of(css, struct mem_cgroup, css);
910                 rcu_read_unlock();
911                 /* If css is NULL, no more cgroups will be found */
912                 nextid = found + 1;
913         } while (css && !iter);
914
915         return iter;
916 }
917 /*
918  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
919  * be careful that "break" loop is not allowed. We have reference count.
920  * Instead of that modify "cond" to be false and "continue" to exit the loop.
921  */
922 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
923         for (iter = mem_cgroup_start_loop(root);\
924              iter != NULL;\
925              iter = mem_cgroup_get_next(iter, root, cond))
926
927 #define for_each_mem_cgroup_tree(iter, root) \
928         for_each_mem_cgroup_tree_cond(iter, root, true)
929
930 #define for_each_mem_cgroup_all(iter) \
931         for_each_mem_cgroup_tree_cond(iter, NULL, true)
932
933
934 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
935 {
936         return (memcg == root_mem_cgroup);
937 }
938
939 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
940 {
941         struct mem_cgroup *memcg;
942
943         if (!mm)
944                 return;
945
946         rcu_read_lock();
947         memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
948         if (unlikely(!memcg))
949                 goto out;
950
951         switch (idx) {
952         case PGMAJFAULT:
953                 mem_cgroup_pgmajfault(memcg, 1);
954                 break;
955         case PGFAULT:
956                 mem_cgroup_pgfault(memcg, 1);
957                 break;
958         default:
959                 BUG();
960         }
961 out:
962         rcu_read_unlock();
963 }
964 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
965
966 /*
967  * Following LRU functions are allowed to be used without PCG_LOCK.
968  * Operations are called by routine of global LRU independently from memcg.
969  * What we have to take care of here is validness of pc->mem_cgroup.
970  *
971  * Changes to pc->mem_cgroup happens when
972  * 1. charge
973  * 2. moving account
974  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
975  * It is added to LRU before charge.
976  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
977  * When moving account, the page is not on LRU. It's isolated.
978  */
979
980 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
981 {
982         struct page_cgroup *pc;
983         struct mem_cgroup_per_zone *mz;
984
985         if (mem_cgroup_disabled())
986                 return;
987         pc = lookup_page_cgroup(page);
988         /* can happen while we handle swapcache. */
989         if (!TestClearPageCgroupAcctLRU(pc))
990                 return;
991         VM_BUG_ON(!pc->mem_cgroup);
992         /*
993          * We don't check PCG_USED bit. It's cleared when the "page" is finally
994          * removed from global LRU.
995          */
996         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
997         /* huge page split is done under lru_lock. so, we have no races. */
998         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
999         if (mem_cgroup_is_root(pc->mem_cgroup))
1000                 return;
1001         VM_BUG_ON(list_empty(&pc->lru));
1002         list_del_init(&pc->lru);
1003 }
1004
1005 void mem_cgroup_del_lru(struct page *page)
1006 {
1007         mem_cgroup_del_lru_list(page, page_lru(page));
1008 }
1009
1010 /*
1011  * Writeback is about to end against a page which has been marked for immediate
1012  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
1013  * inactive list.
1014  */
1015 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1016 {
1017         struct mem_cgroup_per_zone *mz;
1018         struct page_cgroup *pc;
1019         enum lru_list lru = page_lru(page);
1020
1021         if (mem_cgroup_disabled())
1022                 return;
1023
1024         pc = lookup_page_cgroup(page);
1025         /* unused or root page is not rotated. */
1026         if (!PageCgroupUsed(pc))
1027                 return;
1028         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1029         smp_rmb();
1030         if (mem_cgroup_is_root(pc->mem_cgroup))
1031                 return;
1032         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1033         list_move_tail(&pc->lru, &mz->lists[lru]);
1034 }
1035
1036 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1037 {
1038         struct mem_cgroup_per_zone *mz;
1039         struct page_cgroup *pc;
1040
1041         if (mem_cgroup_disabled())
1042                 return;
1043
1044         pc = lookup_page_cgroup(page);
1045         /* unused or root page is not rotated. */
1046         if (!PageCgroupUsed(pc))
1047                 return;
1048         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1049         smp_rmb();
1050         if (mem_cgroup_is_root(pc->mem_cgroup))
1051                 return;
1052         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1053         list_move(&pc->lru, &mz->lists[lru]);
1054 }
1055
1056 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1057 {
1058         struct page_cgroup *pc;
1059         struct mem_cgroup_per_zone *mz;
1060
1061         if (mem_cgroup_disabled())
1062                 return;
1063         pc = lookup_page_cgroup(page);
1064         VM_BUG_ON(PageCgroupAcctLRU(pc));
1065         /*
1066          * putback:                             charge:
1067          * SetPageLRU                           SetPageCgroupUsed
1068          * smp_mb                               smp_mb
1069          * PageCgroupUsed && add to memcg LRU   PageLRU && add to memcg LRU
1070          *
1071          * Ensure that one of the two sides adds the page to the memcg
1072          * LRU during a race.
1073          */
1074         smp_mb();
1075         if (!PageCgroupUsed(pc))
1076                 return;
1077         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1078         smp_rmb();
1079         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1080         /* huge page split is done under lru_lock. so, we have no races. */
1081         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1082         SetPageCgroupAcctLRU(pc);
1083         if (mem_cgroup_is_root(pc->mem_cgroup))
1084                 return;
1085         list_add(&pc->lru, &mz->lists[lru]);
1086 }
1087
1088 /*
1089  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1090  * while it's linked to lru because the page may be reused after it's fully
1091  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1092  * It's done under lock_page and expected that zone->lru_lock isnever held.
1093  */
1094 static void mem_cgroup_lru_del_before_commit(struct page *page)
1095 {
1096         unsigned long flags;
1097         struct zone *zone = page_zone(page);
1098         struct page_cgroup *pc = lookup_page_cgroup(page);
1099
1100         /*
1101          * Doing this check without taking ->lru_lock seems wrong but this
1102          * is safe. Because if page_cgroup's USED bit is unset, the page
1103          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1104          * set, the commit after this will fail, anyway.
1105          * This all charge/uncharge is done under some mutual execustion.
1106          * So, we don't need to taking care of changes in USED bit.
1107          */
1108         if (likely(!PageLRU(page)))
1109                 return;
1110
1111         spin_lock_irqsave(&zone->lru_lock, flags);
1112         /*
1113          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1114          * is guarded by lock_page() because the page is SwapCache.
1115          */
1116         if (!PageCgroupUsed(pc))
1117                 mem_cgroup_del_lru_list(page, page_lru(page));
1118         spin_unlock_irqrestore(&zone->lru_lock, flags);
1119 }
1120
1121 static void mem_cgroup_lru_add_after_commit(struct page *page)
1122 {
1123         unsigned long flags;
1124         struct zone *zone = page_zone(page);
1125         struct page_cgroup *pc = lookup_page_cgroup(page);
1126         /*
1127          * putback:                             charge:
1128          * SetPageLRU                           SetPageCgroupUsed
1129          * smp_mb                               smp_mb
1130          * PageCgroupUsed && add to memcg LRU   PageLRU && add to memcg LRU
1131          *
1132          * Ensure that one of the two sides adds the page to the memcg
1133          * LRU during a race.
1134          */
1135         smp_mb();
1136         /* taking care of that the page is added to LRU while we commit it */
1137         if (likely(!PageLRU(page)))
1138                 return;
1139         spin_lock_irqsave(&zone->lru_lock, flags);
1140         /* link when the page is linked to LRU but page_cgroup isn't */
1141         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1142                 mem_cgroup_add_lru_list(page, page_lru(page));
1143         spin_unlock_irqrestore(&zone->lru_lock, flags);
1144 }
1145
1146
1147 void mem_cgroup_move_lists(struct page *page,
1148                            enum lru_list from, enum lru_list to)
1149 {
1150         if (mem_cgroup_disabled())
1151                 return;
1152         mem_cgroup_del_lru_list(page, from);
1153         mem_cgroup_add_lru_list(page, to);
1154 }
1155
1156 /*
1157  * Checks whether given mem is same or in the root_mem_cgroup's
1158  * hierarchy subtree
1159  */
1160 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1161                 struct mem_cgroup *memcg)
1162 {
1163         if (root_memcg != memcg) {
1164                 return (root_memcg->use_hierarchy &&
1165                         css_is_ancestor(&memcg->css, &root_memcg->css));
1166         }
1167
1168         return true;
1169 }
1170
1171 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1172 {
1173         int ret;
1174         struct mem_cgroup *curr = NULL;
1175         struct task_struct *p;
1176
1177         p = find_lock_task_mm(task);
1178         if (!p)
1179                 return 0;
1180         curr = try_get_mem_cgroup_from_mm(p->mm);
1181         task_unlock(p);
1182         if (!curr)
1183                 return 0;
1184         /*
1185          * We should check use_hierarchy of "memcg" not "curr". Because checking
1186          * use_hierarchy of "curr" here make this function true if hierarchy is
1187          * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1188          * hierarchy(even if use_hierarchy is disabled in "memcg").
1189          */
1190         ret = mem_cgroup_same_or_subtree(memcg, curr);
1191         css_put(&curr->css);
1192         return ret;
1193 }
1194
1195 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1196 {
1197         unsigned long inactive_ratio;
1198         int nid = zone_to_nid(zone);
1199         int zid = zone_idx(zone);
1200         unsigned long inactive;
1201         unsigned long active;
1202         unsigned long gb;
1203
1204         inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1205                                                 BIT(LRU_INACTIVE_ANON));
1206         active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1207                                               BIT(LRU_ACTIVE_ANON));
1208
1209         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1210         if (gb)
1211                 inactive_ratio = int_sqrt(10 * gb);
1212         else
1213                 inactive_ratio = 1;
1214
1215         return inactive * inactive_ratio < active;
1216 }
1217
1218 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1219 {
1220         unsigned long active;
1221         unsigned long inactive;
1222         int zid = zone_idx(zone);
1223         int nid = zone_to_nid(zone);
1224
1225         inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1226                                                 BIT(LRU_INACTIVE_FILE));
1227         active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1228                                               BIT(LRU_ACTIVE_FILE));
1229
1230         return (active > inactive);
1231 }
1232
1233 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1234                                                       struct zone *zone)
1235 {
1236         int nid = zone_to_nid(zone);
1237         int zid = zone_idx(zone);
1238         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1239
1240         return &mz->reclaim_stat;
1241 }
1242
1243 struct zone_reclaim_stat *
1244 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1245 {
1246         struct page_cgroup *pc;
1247         struct mem_cgroup_per_zone *mz;
1248
1249         if (mem_cgroup_disabled())
1250                 return NULL;
1251
1252         pc = lookup_page_cgroup(page);
1253         if (!PageCgroupUsed(pc))
1254                 return NULL;
1255         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1256         smp_rmb();
1257         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1258         return &mz->reclaim_stat;
1259 }
1260
1261 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1262                                         struct list_head *dst,
1263                                         unsigned long *scanned, int order,
1264                                         isolate_mode_t mode,
1265                                         struct zone *z,
1266                                         struct mem_cgroup *mem_cont,
1267                                         int active, int file)
1268 {
1269         unsigned long nr_taken = 0;
1270         struct page *page;
1271         unsigned long scan;
1272         LIST_HEAD(pc_list);
1273         struct list_head *src;
1274         struct page_cgroup *pc, *tmp;
1275         int nid = zone_to_nid(z);
1276         int zid = zone_idx(z);
1277         struct mem_cgroup_per_zone *mz;
1278         int lru = LRU_FILE * file + active;
1279         int ret;
1280
1281         BUG_ON(!mem_cont);
1282         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1283         src = &mz->lists[lru];
1284
1285         scan = 0;
1286         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1287                 if (scan >= nr_to_scan)
1288                         break;
1289
1290                 if (unlikely(!PageCgroupUsed(pc)))
1291                         continue;
1292
1293                 page = lookup_cgroup_page(pc);
1294
1295                 if (unlikely(!PageLRU(page)))
1296                         continue;
1297
1298                 scan++;
1299                 ret = __isolate_lru_page(page, mode, file);
1300                 switch (ret) {
1301                 case 0:
1302                         list_move(&page->lru, dst);
1303                         mem_cgroup_del_lru(page);
1304                         nr_taken += hpage_nr_pages(page);
1305                         break;
1306                 case -EBUSY:
1307                         /* we don't affect global LRU but rotate in our LRU */
1308                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1309                         break;
1310                 default:
1311                         break;
1312                 }
1313         }
1314
1315         *scanned = scan;
1316
1317         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1318                                       0, 0, 0, mode);
1319
1320         return nr_taken;
1321 }
1322
1323 #define mem_cgroup_from_res_counter(counter, member)    \
1324         container_of(counter, struct mem_cgroup, member)
1325
1326 /**
1327  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1328  * @mem: the memory cgroup
1329  *
1330  * Returns the maximum amount of memory @mem can be charged with, in
1331  * pages.
1332  */
1333 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1334 {
1335         unsigned long long margin;
1336
1337         margin = res_counter_margin(&memcg->res);
1338         if (do_swap_account)
1339                 margin = min(margin, res_counter_margin(&memcg->memsw));
1340         return margin >> PAGE_SHIFT;
1341 }
1342
1343 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1344 {
1345         struct cgroup *cgrp = memcg->css.cgroup;
1346
1347         /* root ? */
1348         if (cgrp->parent == NULL)
1349                 return vm_swappiness;
1350
1351         return memcg->swappiness;
1352 }
1353
1354 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1355 {
1356         int cpu;
1357
1358         get_online_cpus();
1359         spin_lock(&memcg->pcp_counter_lock);
1360         for_each_online_cpu(cpu)
1361                 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1362         memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1363         spin_unlock(&memcg->pcp_counter_lock);
1364         put_online_cpus();
1365
1366         synchronize_rcu();
1367 }
1368
1369 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1370 {
1371         int cpu;
1372
1373         if (!memcg)
1374                 return;
1375         get_online_cpus();
1376         spin_lock(&memcg->pcp_counter_lock);
1377         for_each_online_cpu(cpu)
1378                 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1379         memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1380         spin_unlock(&memcg->pcp_counter_lock);
1381         put_online_cpus();
1382 }
1383 /*
1384  * 2 routines for checking "mem" is under move_account() or not.
1385  *
1386  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1387  *                        for avoiding race in accounting. If true,
1388  *                        pc->mem_cgroup may be overwritten.
1389  *
1390  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1391  *                        under hierarchy of moving cgroups. This is for
1392  *                        waiting at hith-memory prressure caused by "move".
1393  */
1394
1395 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1396 {
1397         VM_BUG_ON(!rcu_read_lock_held());
1398         return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1399 }
1400
1401 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1402 {
1403         struct mem_cgroup *from;
1404         struct mem_cgroup *to;
1405         bool ret = false;
1406         /*
1407          * Unlike task_move routines, we access mc.to, mc.from not under
1408          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1409          */
1410         spin_lock(&mc.lock);
1411         from = mc.from;
1412         to = mc.to;
1413         if (!from)
1414                 goto unlock;
1415
1416         ret = mem_cgroup_same_or_subtree(memcg, from)
1417                 || mem_cgroup_same_or_subtree(memcg, to);
1418 unlock:
1419         spin_unlock(&mc.lock);
1420         return ret;
1421 }
1422
1423 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1424 {
1425         if (mc.moving_task && current != mc.moving_task) {
1426                 if (mem_cgroup_under_move(memcg)) {
1427                         DEFINE_WAIT(wait);
1428                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1429                         /* moving charge context might have finished. */
1430                         if (mc.moving_task)
1431                                 schedule();
1432                         finish_wait(&mc.waitq, &wait);
1433                         return true;
1434                 }
1435         }
1436         return false;
1437 }
1438
1439 /**
1440  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1441  * @memcg: The memory cgroup that went over limit
1442  * @p: Task that is going to be killed
1443  *
1444  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1445  * enabled
1446  */
1447 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1448 {
1449         struct cgroup *task_cgrp;
1450         struct cgroup *mem_cgrp;
1451         /*
1452          * Need a buffer in BSS, can't rely on allocations. The code relies
1453          * on the assumption that OOM is serialized for memory controller.
1454          * If this assumption is broken, revisit this code.
1455          */
1456         static char memcg_name[PATH_MAX];
1457         int ret;
1458
1459         if (!memcg || !p)
1460                 return;
1461
1462
1463         rcu_read_lock();
1464
1465         mem_cgrp = memcg->css.cgroup;
1466         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1467
1468         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1469         if (ret < 0) {
1470                 /*
1471                  * Unfortunately, we are unable to convert to a useful name
1472                  * But we'll still print out the usage information
1473                  */
1474                 rcu_read_unlock();
1475                 goto done;
1476         }
1477         rcu_read_unlock();
1478
1479         printk(KERN_INFO "Task in %s killed", memcg_name);
1480
1481         rcu_read_lock();
1482         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1483         if (ret < 0) {
1484                 rcu_read_unlock();
1485                 goto done;
1486         }
1487         rcu_read_unlock();
1488
1489         /*
1490          * Continues from above, so we don't need an KERN_ level
1491          */
1492         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1493 done:
1494
1495         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1496                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1497                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1498                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1499         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1500                 "failcnt %llu\n",
1501                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1502                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1503                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1504 }
1505
1506 /*
1507  * This function returns the number of memcg under hierarchy tree. Returns
1508  * 1(self count) if no children.
1509  */
1510 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1511 {
1512         int num = 0;
1513         struct mem_cgroup *iter;
1514
1515         for_each_mem_cgroup_tree(iter, memcg)
1516                 num++;
1517         return num;
1518 }
1519
1520 /*
1521  * Return the memory (and swap, if configured) limit for a memcg.
1522  */
1523 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1524 {
1525         u64 limit;
1526         u64 memsw;
1527
1528         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1529         limit += total_swap_pages << PAGE_SHIFT;
1530
1531         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1532         /*
1533          * If memsw is finite and limits the amount of swap space available
1534          * to this memcg, return that limit.
1535          */
1536         return min(limit, memsw);
1537 }
1538
1539 /*
1540  * Visit the first child (need not be the first child as per the ordering
1541  * of the cgroup list, since we track last_scanned_child) of @mem and use
1542  * that to reclaim free pages from.
1543  */
1544 static struct mem_cgroup *
1545 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1546 {
1547         struct mem_cgroup *ret = NULL;
1548         struct cgroup_subsys_state *css;
1549         int nextid, found;
1550
1551         if (!root_memcg->use_hierarchy) {
1552                 css_get(&root_memcg->css);
1553                 ret = root_memcg;
1554         }
1555
1556         while (!ret) {
1557                 rcu_read_lock();
1558                 nextid = root_memcg->last_scanned_child + 1;
1559                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1560                                    &found);
1561                 if (css && css_tryget(css))
1562                         ret = container_of(css, struct mem_cgroup, css);
1563
1564                 rcu_read_unlock();
1565                 /* Updates scanning parameter */
1566                 if (!css) {
1567                         /* this means start scan from ID:1 */
1568                         root_memcg->last_scanned_child = 0;
1569                 } else
1570                         root_memcg->last_scanned_child = found;
1571         }
1572
1573         return ret;
1574 }
1575
1576 /**
1577  * test_mem_cgroup_node_reclaimable
1578  * @mem: the target memcg
1579  * @nid: the node ID to be checked.
1580  * @noswap : specify true here if the user wants flle only information.
1581  *
1582  * This function returns whether the specified memcg contains any
1583  * reclaimable pages on a node. Returns true if there are any reclaimable
1584  * pages in the node.
1585  */
1586 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1587                 int nid, bool noswap)
1588 {
1589         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1590                 return true;
1591         if (noswap || !total_swap_pages)
1592                 return false;
1593         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1594                 return true;
1595         return false;
1596
1597 }
1598 #if MAX_NUMNODES > 1
1599
1600 /*
1601  * Always updating the nodemask is not very good - even if we have an empty
1602  * list or the wrong list here, we can start from some node and traverse all
1603  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1604  *
1605  */
1606 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1607 {
1608         int nid;
1609         /*
1610          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1611          * pagein/pageout changes since the last update.
1612          */
1613         if (!atomic_read(&memcg->numainfo_events))
1614                 return;
1615         if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1616                 return;
1617
1618         /* make a nodemask where this memcg uses memory from */
1619         memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1620
1621         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1622
1623                 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1624                         node_clear(nid, memcg->scan_nodes);
1625         }
1626
1627         atomic_set(&memcg->numainfo_events, 0);
1628         atomic_set(&memcg->numainfo_updating, 0);
1629 }
1630
1631 /*
1632  * Selecting a node where we start reclaim from. Because what we need is just
1633  * reducing usage counter, start from anywhere is O,K. Considering
1634  * memory reclaim from current node, there are pros. and cons.
1635  *
1636  * Freeing memory from current node means freeing memory from a node which
1637  * we'll use or we've used. So, it may make LRU bad. And if several threads
1638  * hit limits, it will see a contention on a node. But freeing from remote
1639  * node means more costs for memory reclaim because of memory latency.
1640  *
1641  * Now, we use round-robin. Better algorithm is welcomed.
1642  */
1643 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1644 {
1645         int node;
1646
1647         mem_cgroup_may_update_nodemask(memcg);
1648         node = memcg->last_scanned_node;
1649
1650         node = next_node(node, memcg->scan_nodes);
1651         if (node == MAX_NUMNODES)
1652                 node = first_node(memcg->scan_nodes);
1653         /*
1654          * We call this when we hit limit, not when pages are added to LRU.
1655          * No LRU may hold pages because all pages are UNEVICTABLE or
1656          * memcg is too small and all pages are not on LRU. In that case,
1657          * we use curret node.
1658          */
1659         if (unlikely(node == MAX_NUMNODES))
1660                 node = numa_node_id();
1661
1662         memcg->last_scanned_node = node;
1663         return node;
1664 }
1665
1666 /*
1667  * Check all nodes whether it contains reclaimable pages or not.
1668  * For quick scan, we make use of scan_nodes. This will allow us to skip
1669  * unused nodes. But scan_nodes is lazily updated and may not cotain
1670  * enough new information. We need to do double check.
1671  */
1672 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1673 {
1674         int nid;
1675
1676         /*
1677          * quick check...making use of scan_node.
1678          * We can skip unused nodes.
1679          */
1680         if (!nodes_empty(memcg->scan_nodes)) {
1681                 for (nid = first_node(memcg->scan_nodes);
1682                      nid < MAX_NUMNODES;
1683                      nid = next_node(nid, memcg->scan_nodes)) {
1684
1685                         if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1686                                 return true;
1687                 }
1688         }
1689         /*
1690          * Check rest of nodes.
1691          */
1692         for_each_node_state(nid, N_HIGH_MEMORY) {
1693                 if (node_isset(nid, memcg->scan_nodes))
1694                         continue;
1695                 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1696                         return true;
1697         }
1698         return false;
1699 }
1700
1701 #else
1702 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1703 {
1704         return 0;
1705 }
1706
1707 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1708 {
1709         return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1710 }
1711 #endif
1712
1713 /*
1714  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1715  * we reclaimed from, so that we don't end up penalizing one child extensively
1716  * based on its position in the children list.
1717  *
1718  * root_memcg is the original ancestor that we've been reclaim from.
1719  *
1720  * We give up and return to the caller when we visit root_memcg twice.
1721  * (other groups can be removed while we're walking....)
1722  *
1723  * If shrink==true, for avoiding to free too much, this returns immedieately.
1724  */
1725 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1726                                                 struct zone *zone,
1727                                                 gfp_t gfp_mask,
1728                                                 unsigned long reclaim_options,
1729                                                 unsigned long *total_scanned)
1730 {
1731         struct mem_cgroup *victim;
1732         int ret, total = 0;
1733         int loop = 0;
1734         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1735         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1736         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1737         unsigned long excess;
1738         unsigned long nr_scanned;
1739
1740         excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1741
1742         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1743         if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1744                 noswap = true;
1745
1746         while (1) {
1747                 victim = mem_cgroup_select_victim(root_memcg);
1748                 if (victim == root_memcg) {
1749                         loop++;
1750                         /*
1751                          * We are not draining per cpu cached charges during
1752                          * soft limit reclaim  because global reclaim doesn't
1753                          * care about charges. It tries to free some memory and
1754                          * charges will not give any.
1755                          */
1756                         if (!check_soft && loop >= 1)
1757                                 drain_all_stock_async(root_memcg);
1758                         if (loop >= 2) {
1759                                 /*
1760                                  * If we have not been able to reclaim
1761                                  * anything, it might because there are
1762                                  * no reclaimable pages under this hierarchy
1763                                  */
1764                                 if (!check_soft || !total) {
1765                                         css_put(&victim->css);
1766                                         break;
1767                                 }
1768                                 /*
1769                                  * We want to do more targeted reclaim.
1770                                  * excess >> 2 is not to excessive so as to
1771                                  * reclaim too much, nor too less that we keep
1772                                  * coming back to reclaim from this cgroup
1773                                  */
1774                                 if (total >= (excess >> 2) ||
1775                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1776                                         css_put(&victim->css);
1777                                         break;
1778                                 }
1779                         }
1780                 }
1781                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1782                         /* this cgroup's local usage == 0 */
1783                         css_put(&victim->css);
1784                         continue;
1785                 }
1786                 /* we use swappiness of local cgroup */
1787                 if (check_soft) {
1788                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1789                                 noswap, zone, &nr_scanned);
1790                         *total_scanned += nr_scanned;
1791                 } else
1792                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1793                                                 noswap);
1794                 css_put(&victim->css);
1795                 /*
1796                  * At shrinking usage, we can't check we should stop here or
1797                  * reclaim more. It's depends on callers. last_scanned_child
1798                  * will work enough for keeping fairness under tree.
1799                  */
1800                 if (shrink)
1801                         return ret;
1802                 total += ret;
1803                 if (check_soft) {
1804                         if (!res_counter_soft_limit_excess(&root_memcg->res))
1805                                 return total;
1806                 } else if (mem_cgroup_margin(root_memcg))
1807                         return total;
1808         }
1809         return total;
1810 }
1811
1812 /*
1813  * Check OOM-Killer is already running under our hierarchy.
1814  * If someone is running, return false.
1815  * Has to be called with memcg_oom_lock
1816  */
1817 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1818 {
1819         struct mem_cgroup *iter, *failed = NULL;
1820         bool cond = true;
1821
1822         for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1823                 if (iter->oom_lock) {
1824                         /*
1825                          * this subtree of our hierarchy is already locked
1826                          * so we cannot give a lock.
1827                          */
1828                         failed = iter;
1829                         cond = false;
1830                 } else
1831                         iter->oom_lock = true;
1832         }
1833
1834         if (!failed)
1835                 return true;
1836
1837         /*
1838          * OK, we failed to lock the whole subtree so we have to clean up
1839          * what we set up to the failing subtree
1840          */
1841         cond = true;
1842         for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1843                 if (iter == failed) {
1844                         cond = false;
1845                         continue;
1846                 }
1847                 iter->oom_lock = false;
1848         }
1849         return false;
1850 }
1851
1852 /*
1853  * Has to be called with memcg_oom_lock
1854  */
1855 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1856 {
1857         struct mem_cgroup *iter;
1858
1859         for_each_mem_cgroup_tree(iter, memcg)
1860                 iter->oom_lock = false;
1861         return 0;
1862 }
1863
1864 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1865 {
1866         struct mem_cgroup *iter;
1867
1868         for_each_mem_cgroup_tree(iter, memcg)
1869                 atomic_inc(&iter->under_oom);
1870 }
1871
1872 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1873 {
1874         struct mem_cgroup *iter;
1875
1876         /*
1877          * When a new child is created while the hierarchy is under oom,
1878          * mem_cgroup_oom_lock() may not be called. We have to use
1879          * atomic_add_unless() here.
1880          */
1881         for_each_mem_cgroup_tree(iter, memcg)
1882                 atomic_add_unless(&iter->under_oom, -1, 0);
1883 }
1884
1885 static DEFINE_SPINLOCK(memcg_oom_lock);
1886 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1887
1888 struct oom_wait_info {
1889         struct mem_cgroup *mem;
1890         wait_queue_t    wait;
1891 };
1892
1893 static int memcg_oom_wake_function(wait_queue_t *wait,
1894         unsigned mode, int sync, void *arg)
1895 {
1896         struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1897                           *oom_wait_memcg;
1898         struct oom_wait_info *oom_wait_info;
1899
1900         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1901         oom_wait_memcg = oom_wait_info->mem;
1902
1903         /*
1904          * Both of oom_wait_info->mem and wake_mem are stable under us.
1905          * Then we can use css_is_ancestor without taking care of RCU.
1906          */
1907         if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1908                 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1909                 return 0;
1910         return autoremove_wake_function(wait, mode, sync, arg);
1911 }
1912
1913 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1914 {
1915         /* for filtering, pass "memcg" as argument. */
1916         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1917 }
1918
1919 static void memcg_oom_recover(struct mem_cgroup *memcg)
1920 {
1921         if (memcg && atomic_read(&memcg->under_oom))
1922                 memcg_wakeup_oom(memcg);
1923 }
1924
1925 /*
1926  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1927  */
1928 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1929 {
1930         struct oom_wait_info owait;
1931         bool locked, need_to_kill;
1932
1933         owait.mem = memcg;
1934         owait.wait.flags = 0;
1935         owait.wait.func = memcg_oom_wake_function;
1936         owait.wait.private = current;
1937         INIT_LIST_HEAD(&owait.wait.task_list);
1938         need_to_kill = true;
1939         mem_cgroup_mark_under_oom(memcg);
1940
1941         /* At first, try to OOM lock hierarchy under memcg.*/
1942         spin_lock(&memcg_oom_lock);
1943         locked = mem_cgroup_oom_lock(memcg);
1944         /*
1945          * Even if signal_pending(), we can't quit charge() loop without
1946          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1947          * under OOM is always welcomed, use TASK_KILLABLE here.
1948          */
1949         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1950         if (!locked || memcg->oom_kill_disable)
1951                 need_to_kill = false;
1952         if (locked)
1953                 mem_cgroup_oom_notify(memcg);
1954         spin_unlock(&memcg_oom_lock);
1955
1956         if (need_to_kill) {
1957                 finish_wait(&memcg_oom_waitq, &owait.wait);
1958                 mem_cgroup_out_of_memory(memcg, mask);
1959         } else {
1960                 schedule();
1961                 finish_wait(&memcg_oom_waitq, &owait.wait);
1962         }
1963         spin_lock(&memcg_oom_lock);
1964         if (locked)
1965                 mem_cgroup_oom_unlock(memcg);
1966         memcg_wakeup_oom(memcg);
1967         spin_unlock(&memcg_oom_lock);
1968
1969         mem_cgroup_unmark_under_oom(memcg);
1970
1971         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1972                 return false;
1973         /* Give chance to dying process */
1974         schedule_timeout_uninterruptible(1);
1975         return true;
1976 }
1977
1978 /*
1979  * Currently used to update mapped file statistics, but the routine can be
1980  * generalized to update other statistics as well.
1981  *
1982  * Notes: Race condition
1983  *
1984  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1985  * it tends to be costly. But considering some conditions, we doesn't need
1986  * to do so _always_.
1987  *
1988  * Considering "charge", lock_page_cgroup() is not required because all
1989  * file-stat operations happen after a page is attached to radix-tree. There
1990  * are no race with "charge".
1991  *
1992  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1993  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1994  * if there are race with "uncharge". Statistics itself is properly handled
1995  * by flags.
1996  *
1997  * Considering "move", this is an only case we see a race. To make the race
1998  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1999  * possibility of race condition. If there is, we take a lock.
2000  */
2001
2002 void mem_cgroup_update_page_stat(struct page *page,
2003                                  enum mem_cgroup_page_stat_item idx, int val)
2004 {
2005         struct mem_cgroup *memcg;
2006         struct page_cgroup *pc = lookup_page_cgroup(page);
2007         bool need_unlock = false;
2008         unsigned long uninitialized_var(flags);
2009
2010         if (unlikely(!pc))
2011                 return;
2012
2013         rcu_read_lock();
2014         memcg = pc->mem_cgroup;
2015         if (unlikely(!memcg || !PageCgroupUsed(pc)))
2016                 goto out;
2017         /* pc->mem_cgroup is unstable ? */
2018         if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2019                 /* take a lock against to access pc->mem_cgroup */
2020                 move_lock_page_cgroup(pc, &flags);
2021                 need_unlock = true;
2022                 memcg = pc->mem_cgroup;
2023                 if (!memcg || !PageCgroupUsed(pc))
2024                         goto out;
2025         }
2026
2027         switch (idx) {
2028         case MEMCG_NR_FILE_MAPPED:
2029                 if (val > 0)
2030                         SetPageCgroupFileMapped(pc);
2031                 else if (!page_mapped(page))
2032                         ClearPageCgroupFileMapped(pc);
2033                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2034                 break;
2035         default:
2036                 BUG();
2037         }
2038
2039         this_cpu_add(memcg->stat->count[idx], val);
2040
2041 out:
2042         if (unlikely(need_unlock))
2043                 move_unlock_page_cgroup(pc, &flags);
2044         rcu_read_unlock();
2045         return;
2046 }
2047 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2048
2049 /*
2050  * size of first charge trial. "32" comes from vmscan.c's magic value.
2051  * TODO: maybe necessary to use big numbers in big irons.
2052  */
2053 #define CHARGE_BATCH    32U
2054 struct memcg_stock_pcp {
2055         struct mem_cgroup *cached; /* this never be root cgroup */
2056         unsigned int nr_pages;
2057         struct work_struct work;
2058         unsigned long flags;
2059 #define FLUSHING_CACHED_CHARGE  (0)
2060 };
2061 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2062 static DEFINE_MUTEX(percpu_charge_mutex);
2063
2064 /*
2065  * Try to consume stocked charge on this cpu. If success, one page is consumed
2066  * from local stock and true is returned. If the stock is 0 or charges from a
2067  * cgroup which is not current target, returns false. This stock will be
2068  * refilled.
2069  */
2070 static bool consume_stock(struct mem_cgroup *memcg)
2071 {
2072         struct memcg_stock_pcp *stock;
2073         bool ret = true;
2074
2075         stock = &get_cpu_var(memcg_stock);
2076         if (memcg == stock->cached && stock->nr_pages)
2077                 stock->nr_pages--;
2078         else /* need to call res_counter_charge */
2079                 ret = false;
2080         put_cpu_var(memcg_stock);
2081         return ret;
2082 }
2083
2084 /*
2085  * Returns stocks cached in percpu to res_counter and reset cached information.
2086  */
2087 static void drain_stock(struct memcg_stock_pcp *stock)
2088 {
2089         struct mem_cgroup *old = stock->cached;
2090
2091         if (stock->nr_pages) {
2092                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2093
2094                 res_counter_uncharge(&old->res, bytes);
2095                 if (do_swap_account)
2096                         res_counter_uncharge(&old->memsw, bytes);
2097                 stock->nr_pages = 0;
2098         }
2099         stock->cached = NULL;
2100 }
2101
2102 /*
2103  * This must be called under preempt disabled or must be called by
2104  * a thread which is pinned to local cpu.
2105  */
2106 static void drain_local_stock(struct work_struct *dummy)
2107 {
2108         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2109         drain_stock(stock);
2110         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2111 }
2112
2113 /*
2114  * Cache charges(val) which is from res_counter, to local per_cpu area.
2115  * This will be consumed by consume_stock() function, later.
2116  */
2117 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2118 {
2119         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2120
2121         if (stock->cached != memcg) { /* reset if necessary */
2122                 drain_stock(stock);
2123                 stock->cached = memcg;
2124         }
2125         stock->nr_pages += nr_pages;
2126         put_cpu_var(memcg_stock);
2127 }
2128
2129 /*
2130  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2131  * of the hierarchy under it. sync flag says whether we should block
2132  * until the work is done.
2133  */
2134 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2135 {
2136         int cpu, curcpu;
2137
2138         /* Notify other cpus that system-wide "drain" is running */
2139         get_online_cpus();
2140         curcpu = get_cpu();
2141         for_each_online_cpu(cpu) {
2142                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2143                 struct mem_cgroup *memcg;
2144
2145                 memcg = stock->cached;
2146                 if (!memcg || !stock->nr_pages)
2147                         continue;
2148                 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2149                         continue;
2150                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2151                         if (cpu == curcpu)
2152                                 drain_local_stock(&stock->work);
2153                         else
2154                                 schedule_work_on(cpu, &stock->work);
2155                 }
2156         }
2157         put_cpu();
2158
2159         if (!sync)
2160                 goto out;
2161
2162         for_each_online_cpu(cpu) {
2163                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2164                 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2165                         flush_work(&stock->work);
2166         }
2167 out:
2168         put_online_cpus();
2169 }
2170
2171 /*
2172  * Tries to drain stocked charges in other cpus. This function is asynchronous
2173  * and just put a work per cpu for draining localy on each cpu. Caller can
2174  * expects some charges will be back to res_counter later but cannot wait for
2175  * it.
2176  */
2177 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2178 {
2179         /*
2180          * If someone calls draining, avoid adding more kworker runs.
2181          */
2182         if (!mutex_trylock(&percpu_charge_mutex))
2183                 return;
2184         drain_all_stock(root_memcg, false);
2185         mutex_unlock(&percpu_charge_mutex);
2186 }
2187
2188 /* This is a synchronous drain interface. */
2189 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2190 {
2191         /* called when force_empty is called */
2192         mutex_lock(&percpu_charge_mutex);
2193         drain_all_stock(root_memcg, true);
2194         mutex_unlock(&percpu_charge_mutex);
2195 }
2196
2197 /*
2198  * This function drains percpu counter value from DEAD cpu and
2199  * move it to local cpu. Note that this function can be preempted.
2200  */
2201 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2202 {
2203         int i;
2204
2205         spin_lock(&memcg->pcp_counter_lock);
2206         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2207                 long x = per_cpu(memcg->stat->count[i], cpu);
2208
2209                 per_cpu(memcg->stat->count[i], cpu) = 0;
2210                 memcg->nocpu_base.count[i] += x;
2211         }
2212         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2213                 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2214
2215                 per_cpu(memcg->stat->events[i], cpu) = 0;
2216                 memcg->nocpu_base.events[i] += x;
2217         }
2218         /* need to clear ON_MOVE value, works as a kind of lock. */
2219         per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2220         spin_unlock(&memcg->pcp_counter_lock);
2221 }
2222
2223 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2224 {
2225         int idx = MEM_CGROUP_ON_MOVE;
2226
2227         spin_lock(&memcg->pcp_counter_lock);
2228         per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2229         spin_unlock(&memcg->pcp_counter_lock);
2230 }
2231
2232 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2233                                         unsigned long action,
2234                                         void *hcpu)
2235 {
2236         int cpu = (unsigned long)hcpu;
2237         struct memcg_stock_pcp *stock;
2238         struct mem_cgroup *iter;
2239
2240         if ((action == CPU_ONLINE)) {
2241                 for_each_mem_cgroup_all(iter)
2242                         synchronize_mem_cgroup_on_move(iter, cpu);
2243                 return NOTIFY_OK;
2244         }
2245
2246         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2247                 return NOTIFY_OK;
2248
2249         for_each_mem_cgroup_all(iter)
2250                 mem_cgroup_drain_pcp_counter(iter, cpu);
2251
2252         stock = &per_cpu(memcg_stock, cpu);
2253         drain_stock(stock);
2254         return NOTIFY_OK;
2255 }
2256
2257
2258 /* See __mem_cgroup_try_charge() for details */
2259 enum {
2260         CHARGE_OK,              /* success */
2261         CHARGE_RETRY,           /* need to retry but retry is not bad */
2262         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2263         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2264         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2265 };
2266
2267 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2268                                 unsigned int nr_pages, bool oom_check)
2269 {
2270         unsigned long csize = nr_pages * PAGE_SIZE;
2271         struct mem_cgroup *mem_over_limit;
2272         struct res_counter *fail_res;
2273         unsigned long flags = 0;
2274         int ret;
2275
2276         ret = res_counter_charge(&memcg->res, csize, &fail_res);
2277
2278         if (likely(!ret)) {
2279                 if (!do_swap_account)
2280                         return CHARGE_OK;
2281                 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2282                 if (likely(!ret))
2283                         return CHARGE_OK;
2284
2285                 res_counter_uncharge(&memcg->res, csize);
2286                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2287                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2288         } else
2289                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2290         /*
2291          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2292          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2293          *
2294          * Never reclaim on behalf of optional batching, retry with a
2295          * single page instead.
2296          */
2297         if (nr_pages == CHARGE_BATCH)
2298                 return CHARGE_RETRY;
2299
2300         if (!(gfp_mask & __GFP_WAIT))
2301                 return CHARGE_WOULDBLOCK;
2302
2303         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2304                                               gfp_mask, flags, NULL);
2305         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2306                 return CHARGE_RETRY;
2307         /*
2308          * Even though the limit is exceeded at this point, reclaim
2309          * may have been able to free some pages.  Retry the charge
2310          * before killing the task.
2311          *
2312          * Only for regular pages, though: huge pages are rather
2313          * unlikely to succeed so close to the limit, and we fall back
2314          * to regular pages anyway in case of failure.
2315          */
2316         if (nr_pages == 1 && ret)
2317                 return CHARGE_RETRY;
2318
2319         /*
2320          * At task move, charge accounts can be doubly counted. So, it's
2321          * better to wait until the end of task_move if something is going on.
2322          */
2323         if (mem_cgroup_wait_acct_move(mem_over_limit))
2324                 return CHARGE_RETRY;
2325
2326         /* If we don't need to call oom-killer at el, return immediately */
2327         if (!oom_check)
2328                 return CHARGE_NOMEM;
2329         /* check OOM */
2330         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2331                 return CHARGE_OOM_DIE;
2332
2333         return CHARGE_RETRY;
2334 }
2335
2336 /*
2337  * Unlike exported interface, "oom" parameter is added. if oom==true,
2338  * oom-killer can be invoked.
2339  */
2340 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2341                                    gfp_t gfp_mask,
2342                                    unsigned int nr_pages,
2343                                    struct mem_cgroup **ptr,
2344                                    bool oom)
2345 {
2346         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2347         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2348         struct mem_cgroup *memcg = NULL;
2349         int ret;
2350
2351         /*
2352          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2353          * in system level. So, allow to go ahead dying process in addition to
2354          * MEMDIE process.
2355          */
2356         if (unlikely(test_thread_flag(TIF_MEMDIE)
2357                      || fatal_signal_pending(current)))
2358                 goto bypass;
2359
2360         /*
2361          * We always charge the cgroup the mm_struct belongs to.
2362          * The mm_struct's mem_cgroup changes on task migration if the
2363          * thread group leader migrates. It's possible that mm is not
2364          * set, if so charge the init_mm (happens for pagecache usage).
2365          */
2366         if (!*ptr && !mm)
2367                 goto bypass;
2368 again:
2369         if (*ptr) { /* css should be a valid one */
2370                 memcg = *ptr;
2371                 VM_BUG_ON(css_is_removed(&memcg->css));
2372                 if (mem_cgroup_is_root(memcg))
2373                         goto done;
2374                 if (nr_pages == 1 && consume_stock(memcg))
2375                         goto done;
2376                 css_get(&memcg->css);
2377         } else {
2378                 struct task_struct *p;
2379
2380                 rcu_read_lock();
2381                 p = rcu_dereference(mm->owner);
2382                 /*
2383                  * Because we don't have task_lock(), "p" can exit.
2384                  * In that case, "memcg" can point to root or p can be NULL with
2385                  * race with swapoff. Then, we have small risk of mis-accouning.
2386                  * But such kind of mis-account by race always happens because
2387                  * we don't have cgroup_mutex(). It's overkill and we allo that
2388                  * small race, here.
2389                  * (*) swapoff at el will charge against mm-struct not against
2390                  * task-struct. So, mm->owner can be NULL.
2391                  */
2392                 memcg = mem_cgroup_from_task(p);
2393                 if (!memcg || mem_cgroup_is_root(memcg)) {
2394                         rcu_read_unlock();
2395                         goto done;
2396                 }
2397                 if (nr_pages == 1 && consume_stock(memcg)) {
2398                         /*
2399                          * It seems dagerous to access memcg without css_get().
2400                          * But considering how consume_stok works, it's not
2401                          * necessary. If consume_stock success, some charges
2402                          * from this memcg are cached on this cpu. So, we
2403                          * don't need to call css_get()/css_tryget() before
2404                          * calling consume_stock().
2405                          */
2406                         rcu_read_unlock();
2407                         goto done;
2408                 }
2409                 /* after here, we may be blocked. we need to get refcnt */
2410                 if (!css_tryget(&memcg->css)) {
2411                         rcu_read_unlock();
2412                         goto again;
2413                 }
2414                 rcu_read_unlock();
2415         }
2416
2417         do {
2418                 bool oom_check;
2419
2420                 /* If killed, bypass charge */
2421                 if (fatal_signal_pending(current)) {
2422                         css_put(&memcg->css);
2423                         goto bypass;
2424                 }
2425
2426                 oom_check = false;
2427                 if (oom && !nr_oom_retries) {
2428                         oom_check = true;
2429                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2430                 }
2431
2432                 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2433                 switch (ret) {
2434                 case CHARGE_OK:
2435                         break;
2436                 case CHARGE_RETRY: /* not in OOM situation but retry */
2437                         batch = nr_pages;
2438                         css_put(&memcg->css);
2439                         memcg = NULL;
2440                         goto again;
2441                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2442                         css_put(&memcg->css);
2443                         goto nomem;
2444                 case CHARGE_NOMEM: /* OOM routine works */
2445                         if (!oom) {
2446                                 css_put(&memcg->css);
2447                                 goto nomem;
2448                         }
2449                         /* If oom, we never return -ENOMEM */
2450                         nr_oom_retries--;
2451                         break;
2452                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2453                         css_put(&memcg->css);
2454                         goto bypass;
2455                 }
2456         } while (ret != CHARGE_OK);
2457
2458         if (batch > nr_pages)
2459                 refill_stock(memcg, batch - nr_pages);
2460         css_put(&memcg->css);
2461 done:
2462         *ptr = memcg;
2463         return 0;
2464 nomem:
2465         *ptr = NULL;
2466         return -ENOMEM;
2467 bypass:
2468         *ptr = NULL;
2469         return 0;
2470 }
2471
2472 /*
2473  * Somemtimes we have to undo a charge we got by try_charge().
2474  * This function is for that and do uncharge, put css's refcnt.
2475  * gotten by try_charge().
2476  */
2477 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2478                                        unsigned int nr_pages)
2479 {
2480         if (!mem_cgroup_is_root(memcg)) {
2481                 unsigned long bytes = nr_pages * PAGE_SIZE;
2482
2483                 res_counter_uncharge(&memcg->res, bytes);
2484                 if (do_swap_account)
2485                         res_counter_uncharge(&memcg->memsw, bytes);
2486         }
2487 }
2488
2489 /*
2490  * A helper function to get mem_cgroup from ID. must be called under
2491  * rcu_read_lock(). The caller must check css_is_removed() or some if
2492  * it's concern. (dropping refcnt from swap can be called against removed
2493  * memcg.)
2494  */
2495 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2496 {
2497         struct cgroup_subsys_state *css;
2498
2499         /* ID 0 is unused ID */
2500         if (!id)
2501                 return NULL;
2502         css = css_lookup(&mem_cgroup_subsys, id);
2503         if (!css)
2504                 return NULL;
2505         return container_of(css, struct mem_cgroup, css);
2506 }
2507
2508 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2509 {
2510         struct mem_cgroup *memcg = NULL;
2511         struct page_cgroup *pc;
2512         unsigned short id;
2513         swp_entry_t ent;
2514
2515         VM_BUG_ON(!PageLocked(page));
2516
2517         pc = lookup_page_cgroup(page);
2518         lock_page_cgroup(pc);
2519         if (PageCgroupUsed(pc)) {
2520                 memcg = pc->mem_cgroup;
2521                 if (memcg && !css_tryget(&memcg->css))
2522                         memcg = NULL;
2523         } else if (PageSwapCache(page)) {
2524                 ent.val = page_private(page);
2525                 id = lookup_swap_cgroup(ent);
2526                 rcu_read_lock();
2527                 memcg = mem_cgroup_lookup(id);
2528                 if (memcg && !css_tryget(&memcg->css))
2529                         memcg = NULL;
2530                 rcu_read_unlock();
2531         }
2532         unlock_page_cgroup(pc);
2533         return memcg;
2534 }
2535
2536 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2537                                        struct page *page,
2538                                        unsigned int nr_pages,
2539                                        struct page_cgroup *pc,
2540                                        enum charge_type ctype)
2541 {
2542         lock_page_cgroup(pc);
2543         if (unlikely(PageCgroupUsed(pc))) {
2544                 unlock_page_cgroup(pc);
2545                 __mem_cgroup_cancel_charge(memcg, nr_pages);
2546                 return;
2547         }
2548         /*
2549          * we don't need page_cgroup_lock about tail pages, becase they are not
2550          * accessed by any other context at this point.
2551          */
2552         pc->mem_cgroup = memcg;
2553         /*
2554          * We access a page_cgroup asynchronously without lock_page_cgroup().
2555          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2556          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2557          * before USED bit, we need memory barrier here.
2558          * See mem_cgroup_add_lru_list(), etc.
2559          */
2560         smp_wmb();
2561         switch (ctype) {
2562         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2563         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2564                 SetPageCgroupCache(pc);
2565                 SetPageCgroupUsed(pc);
2566                 break;
2567         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2568                 ClearPageCgroupCache(pc);
2569                 SetPageCgroupUsed(pc);
2570                 break;
2571         default:
2572                 break;
2573         }
2574
2575         mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2576         unlock_page_cgroup(pc);
2577         /*
2578          * "charge_statistics" updated event counter. Then, check it.
2579          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2580          * if they exceeds softlimit.
2581          */
2582         memcg_check_events(memcg, page);
2583 }
2584
2585 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2586
2587 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2588                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2589 /*
2590  * Because tail pages are not marked as "used", set it. We're under
2591  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2592  */
2593 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2594 {
2595         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2596         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2597         unsigned long flags;
2598
2599         if (mem_cgroup_disabled())
2600                 return;
2601         /*
2602          * We have no races with charge/uncharge but will have races with
2603          * page state accounting.
2604          */
2605         move_lock_page_cgroup(head_pc, &flags);
2606
2607         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2608         smp_wmb(); /* see __commit_charge() */
2609         if (PageCgroupAcctLRU(head_pc)) {
2610                 enum lru_list lru;
2611                 struct mem_cgroup_per_zone *mz;
2612
2613                 /*
2614                  * LRU flags cannot be copied because we need to add tail
2615                  *.page to LRU by generic call and our hook will be called.
2616                  * We hold lru_lock, then, reduce counter directly.
2617                  */
2618                 lru = page_lru(head);
2619                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2620                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2621         }
2622         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2623         move_unlock_page_cgroup(head_pc, &flags);
2624 }
2625 #endif
2626
2627 /**
2628  * mem_cgroup_move_account - move account of the page
2629  * @page: the page
2630  * @nr_pages: number of regular pages (>1 for huge pages)
2631  * @pc: page_cgroup of the page.
2632  * @from: mem_cgroup which the page is moved from.
2633  * @to: mem_cgroup which the page is moved to. @from != @to.
2634  * @uncharge: whether we should call uncharge and css_put against @from.
2635  *
2636  * The caller must confirm following.
2637  * - page is not on LRU (isolate_page() is useful.)
2638  * - compound_lock is held when nr_pages > 1
2639  *
2640  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2641  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2642  * true, this function does "uncharge" from old cgroup, but it doesn't if
2643  * @uncharge is false, so a caller should do "uncharge".
2644  */
2645 static int mem_cgroup_move_account(struct page *page,
2646                                    unsigned int nr_pages,
2647                                    struct page_cgroup *pc,
2648                                    struct mem_cgroup *from,
2649                                    struct mem_cgroup *to,
2650                                    bool uncharge)
2651 {
2652         unsigned long flags;
2653         int ret;
2654
2655         VM_BUG_ON(from == to);
2656         VM_BUG_ON(PageLRU(page));
2657         /*
2658          * The page is isolated from LRU. So, collapse function
2659          * will not handle this page. But page splitting can happen.
2660          * Do this check under compound_page_lock(). The caller should
2661          * hold it.
2662          */
2663         ret = -EBUSY;
2664         if (nr_pages > 1 && !PageTransHuge(page))
2665                 goto out;
2666
2667         lock_page_cgroup(pc);
2668
2669         ret = -EINVAL;
2670         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2671                 goto unlock;
2672
2673         move_lock_page_cgroup(pc, &flags);
2674
2675         if (PageCgroupFileMapped(pc)) {
2676                 /* Update mapped_file data for mem_cgroup */
2677                 preempt_disable();
2678                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2679                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2680                 preempt_enable();
2681         }
2682         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2683         if (uncharge)
2684                 /* This is not "cancel", but cancel_charge does all we need. */
2685                 __mem_cgroup_cancel_charge(from, nr_pages);
2686
2687         /* caller should have done css_get */
2688         pc->mem_cgroup = to;
2689         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2690         /*
2691          * We charges against "to" which may not have any tasks. Then, "to"
2692          * can be under rmdir(). But in current implementation, caller of
2693          * this function is just force_empty() and move charge, so it's
2694          * guaranteed that "to" is never removed. So, we don't check rmdir
2695          * status here.
2696          */
2697         move_unlock_page_cgroup(pc, &flags);
2698         ret = 0;
2699 unlock:
2700         unlock_page_cgroup(pc);
2701         /*
2702          * check events
2703          */
2704         memcg_check_events(to, page);
2705         memcg_check_events(from, page);
2706 out:
2707         return ret;
2708 }
2709
2710 /*
2711  * move charges to its parent.
2712  */
2713
2714 static int mem_cgroup_move_parent(struct page *page,
2715                                   struct page_cgroup *pc,
2716                                   struct mem_cgroup *child,
2717                                   gfp_t gfp_mask)
2718 {
2719         struct cgroup *cg = child->css.cgroup;
2720         struct cgroup *pcg = cg->parent;
2721         struct mem_cgroup *parent;
2722         unsigned int nr_pages;
2723         unsigned long uninitialized_var(flags);
2724         int ret;
2725
2726         /* Is ROOT ? */
2727         if (!pcg)
2728                 return -EINVAL;
2729
2730         ret = -EBUSY;
2731         if (!get_page_unless_zero(page))
2732                 goto out;
2733         if (isolate_lru_page(page))
2734                 goto put;
2735
2736         nr_pages = hpage_nr_pages(page);
2737
2738         parent = mem_cgroup_from_cont(pcg);
2739         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2740         if (ret || !parent)
2741                 goto put_back;
2742
2743         if (nr_pages > 1)
2744                 flags = compound_lock_irqsave(page);
2745
2746         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2747         if (ret)
2748                 __mem_cgroup_cancel_charge(parent, nr_pages);
2749
2750         if (nr_pages > 1)
2751                 compound_unlock_irqrestore(page, flags);
2752 put_back:
2753         putback_lru_page(page);
2754 put:
2755         put_page(page);
2756 out:
2757         return ret;
2758 }
2759
2760 /*
2761  * Charge the memory controller for page usage.
2762  * Return
2763  * 0 if the charge was successful
2764  * < 0 if the cgroup is over its limit
2765  */
2766 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2767                                 gfp_t gfp_mask, enum charge_type ctype)
2768 {
2769         struct mem_cgroup *memcg = NULL;
2770         unsigned int nr_pages = 1;
2771         struct page_cgroup *pc;
2772         bool oom = true;
2773         int ret;
2774
2775         if (PageTransHuge(page)) {
2776                 nr_pages <<= compound_order(page);
2777                 VM_BUG_ON(!PageTransHuge(page));
2778                 /*
2779                  * Never OOM-kill a process for a huge page.  The
2780                  * fault handler will fall back to regular pages.
2781                  */
2782                 oom = false;
2783         }
2784
2785         pc = lookup_page_cgroup(page);
2786         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2787
2788         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2789         if (ret || !memcg)
2790                 return ret;
2791
2792         __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2793         return 0;
2794 }
2795
2796 int mem_cgroup_newpage_charge(struct page *page,
2797                               struct mm_struct *mm, gfp_t gfp_mask)
2798 {
2799         if (mem_cgroup_disabled())
2800                 return 0;
2801         /*
2802          * If already mapped, we don't have to account.
2803          * If page cache, page->mapping has address_space.
2804          * But page->mapping may have out-of-use anon_vma pointer,
2805          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2806          * is NULL.
2807          */
2808         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2809                 return 0;
2810         if (unlikely(!mm))
2811                 mm = &init_mm;
2812         return mem_cgroup_charge_common(page, mm, gfp_mask,
2813                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2814 }
2815
2816 static void
2817 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2818                                         enum charge_type ctype);
2819
2820 static void
2821 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2822                                         enum charge_type ctype)
2823 {
2824         struct page_cgroup *pc = lookup_page_cgroup(page);
2825         /*
2826          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2827          * is already on LRU. It means the page may on some other page_cgroup's
2828          * LRU. Take care of it.
2829          */
2830         mem_cgroup_lru_del_before_commit(page);
2831         __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2832         mem_cgroup_lru_add_after_commit(page);
2833         return;
2834 }
2835
2836 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2837                                 gfp_t gfp_mask)
2838 {
2839         struct mem_cgroup *memcg = NULL;
2840         int ret;
2841
2842         if (mem_cgroup_disabled())
2843                 return 0;
2844         if (PageCompound(page))
2845                 return 0;
2846
2847         if (unlikely(!mm))
2848                 mm = &init_mm;
2849
2850         if (page_is_file_cache(page)) {
2851                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2852                 if (ret || !memcg)
2853                         return ret;
2854
2855                 /*
2856                  * FUSE reuses pages without going through the final
2857                  * put that would remove them from the LRU list, make
2858                  * sure that they get relinked properly.
2859                  */
2860                 __mem_cgroup_commit_charge_lrucare(page, memcg,
2861                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2862                 return ret;
2863         }
2864         /* shmem */
2865         if (PageSwapCache(page)) {
2866                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2867                 if (!ret)
2868                         __mem_cgroup_commit_charge_swapin(page, memcg,
2869                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2870         } else
2871                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2872                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2873
2874         return ret;
2875 }
2876
2877 /*
2878  * While swap-in, try_charge -> commit or cancel, the page is locked.
2879  * And when try_charge() successfully returns, one refcnt to memcg without
2880  * struct page_cgroup is acquired. This refcnt will be consumed by
2881  * "commit()" or removed by "cancel()"
2882  */
2883 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2884                                  struct page *page,
2885                                  gfp_t mask, struct mem_cgroup **ptr)
2886 {
2887         struct mem_cgroup *memcg;
2888         int ret;
2889
2890         *ptr = NULL;
2891
2892         if (mem_cgroup_disabled())
2893                 return 0;
2894
2895         if (!do_swap_account)
2896                 goto charge_cur_mm;
2897         /*
2898          * A racing thread's fault, or swapoff, may have already updated
2899          * the pte, and even removed page from swap cache: in those cases
2900          * do_swap_page()'s pte_same() test will fail; but there's also a
2901          * KSM case which does need to charge the page.
2902          */
2903         if (!PageSwapCache(page))
2904                 goto charge_cur_mm;
2905         memcg = try_get_mem_cgroup_from_page(page);
2906         if (!memcg)
2907                 goto charge_cur_mm;
2908         *ptr = memcg;
2909         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2910         css_put(&memcg->css);
2911         return ret;
2912 charge_cur_mm:
2913         if (unlikely(!mm))
2914                 mm = &init_mm;
2915         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2916 }
2917
2918 static void
2919 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2920                                         enum charge_type ctype)
2921 {
2922         if (mem_cgroup_disabled())
2923                 return;
2924         if (!ptr)
2925                 return;
2926         cgroup_exclude_rmdir(&ptr->css);
2927
2928         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2929         /*
2930          * Now swap is on-memory. This means this page may be
2931          * counted both as mem and swap....double count.
2932          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2933          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2934          * may call delete_from_swap_cache() before reach here.
2935          */
2936         if (do_swap_account && PageSwapCache(page)) {
2937                 swp_entry_t ent = {.val = page_private(page)};
2938                 unsigned short id;
2939                 struct mem_cgroup *memcg;
2940
2941                 id = swap_cgroup_record(ent, 0);
2942                 rcu_read_lock();
2943                 memcg = mem_cgroup_lookup(id);
2944                 if (memcg) {
2945                         /*
2946                          * This recorded memcg can be obsolete one. So, avoid
2947                          * calling css_tryget
2948                          */
2949                         if (!mem_cgroup_is_root(memcg))
2950                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2951                         mem_cgroup_swap_statistics(memcg, false);
2952                         mem_cgroup_put(memcg);
2953                 }
2954                 rcu_read_unlock();
2955         }
2956         /*
2957          * At swapin, we may charge account against cgroup which has no tasks.
2958          * So, rmdir()->pre_destroy() can be called while we do this charge.
2959          * In that case, we need to call pre_destroy() again. check it here.
2960          */
2961         cgroup_release_and_wakeup_rmdir(&ptr->css);
2962 }
2963
2964 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2965 {
2966         __mem_cgroup_commit_charge_swapin(page, ptr,
2967                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2968 }
2969
2970 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2971 {
2972         if (mem_cgroup_disabled())
2973                 return;
2974         if (!memcg)
2975                 return;
2976         __mem_cgroup_cancel_charge(memcg, 1);
2977 }
2978
2979 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2980                                    unsigned int nr_pages,
2981                                    const enum charge_type ctype)
2982 {
2983         struct memcg_batch_info *batch = NULL;
2984         bool uncharge_memsw = true;
2985
2986         /* If swapout, usage of swap doesn't decrease */
2987         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2988                 uncharge_memsw = false;
2989
2990         batch = &current->memcg_batch;
2991         /*
2992          * In usual, we do css_get() when we remember memcg pointer.
2993          * But in this case, we keep res->usage until end of a series of
2994          * uncharges. Then, it's ok to ignore memcg's refcnt.
2995          */
2996         if (!batch->memcg)
2997                 batch->memcg = memcg;
2998         /*
2999          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3000          * In those cases, all pages freed continuously can be expected to be in
3001          * the same cgroup and we have chance to coalesce uncharges.
3002          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3003          * because we want to do uncharge as soon as possible.
3004          */
3005
3006         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3007                 goto direct_uncharge;
3008
3009         if (nr_pages > 1)
3010                 goto direct_uncharge;
3011
3012         /*
3013          * In typical case, batch->memcg == mem. This means we can
3014          * merge a series of uncharges to an uncharge of res_counter.
3015          * If not, we uncharge res_counter ony by one.
3016          */
3017         if (batch->memcg != memcg)
3018                 goto direct_uncharge;
3019         /* remember freed charge and uncharge it later */
3020         batch->nr_pages++;
3021         if (uncharge_memsw)
3022                 batch->memsw_nr_pages++;
3023         return;
3024 direct_uncharge:
3025         res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3026         if (uncharge_memsw)
3027                 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3028         if (unlikely(batch->memcg != memcg))
3029                 memcg_oom_recover(memcg);
3030         return;
3031 }
3032
3033 /*
3034  * uncharge if !page_mapped(page)
3035  */
3036 static struct mem_cgroup *
3037 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3038 {
3039         struct mem_cgroup *memcg = NULL;
3040         unsigned int nr_pages = 1;
3041         struct page_cgroup *pc;
3042
3043         if (mem_cgroup_disabled())
3044                 return NULL;
3045
3046         if (PageSwapCache(page))
3047                 return NULL;
3048
3049         if (PageTransHuge(page)) {
3050                 nr_pages <<= compound_order(page);
3051                 VM_BUG_ON(!PageTransHuge(page));
3052         }
3053         /*
3054          * Check if our page_cgroup is valid
3055          */
3056         pc = lookup_page_cgroup(page);
3057         if (unlikely(!pc || !PageCgroupUsed(pc)))
3058                 return NULL;
3059
3060         lock_page_cgroup(pc);
3061
3062         memcg = pc->mem_cgroup;
3063
3064         if (!PageCgroupUsed(pc))
3065                 goto unlock_out;
3066
3067         switch (ctype) {
3068         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3069         case MEM_CGROUP_CHARGE_TYPE_DROP:
3070                 /* See mem_cgroup_prepare_migration() */
3071                 if (page_mapped(page) || PageCgroupMigration(pc))
3072                         goto unlock_out;
3073                 break;
3074         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3075                 if (!PageAnon(page)) {  /* Shared memory */
3076                         if (page->mapping && !page_is_file_cache(page))
3077                                 goto unlock_out;
3078                 } else if (page_mapped(page)) /* Anon */
3079                                 goto unlock_out;
3080                 break;
3081         default:
3082                 break;
3083         }
3084
3085         mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3086
3087         ClearPageCgroupUsed(pc);
3088         /*
3089          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3090          * freed from LRU. This is safe because uncharged page is expected not
3091          * to be reused (freed soon). Exception is SwapCache, it's handled by
3092          * special functions.
3093          */
3094
3095         unlock_page_cgroup(pc);
3096         /*
3097          * even after unlock, we have memcg->res.usage here and this memcg
3098          * will never be freed.
3099          */
3100         memcg_check_events(memcg, page);
3101         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3102                 mem_cgroup_swap_statistics(memcg, true);
3103                 mem_cgroup_get(memcg);
3104         }
3105         if (!mem_cgroup_is_root(memcg))
3106                 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3107
3108         return memcg;
3109
3110 unlock_out:
3111         unlock_page_cgroup(pc);
3112         return NULL;
3113 }
3114
3115 void mem_cgroup_uncharge_page(struct page *page)
3116 {
3117         /* early check. */
3118         if (page_mapped(page))
3119                 return;
3120         if (page->mapping && !PageAnon(page))
3121                 return;
3122         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3123 }
3124
3125 void mem_cgroup_uncharge_cache_page(struct page *page)
3126 {
3127         VM_BUG_ON(page_mapped(page));
3128         VM_BUG_ON(page->mapping);
3129         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3130 }
3131
3132 /*
3133  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3134  * In that cases, pages are freed continuously and we can expect pages
3135  * are in the same memcg. All these calls itself limits the number of
3136  * pages freed at once, then uncharge_start/end() is called properly.
3137  * This may be called prural(2) times in a context,
3138  */
3139
3140 void mem_cgroup_uncharge_start(void)
3141 {
3142         current->memcg_batch.do_batch++;
3143         /* We can do nest. */
3144         if (current->memcg_batch.do_batch == 1) {
3145                 current->memcg_batch.memcg = NULL;
3146                 current->memcg_batch.nr_pages = 0;
3147                 current->memcg_batch.memsw_nr_pages = 0;
3148         }
3149 }
3150
3151 void mem_cgroup_uncharge_end(void)
3152 {
3153         struct memcg_batch_info *batch = &current->memcg_batch;
3154
3155         if (!batch->do_batch)
3156                 return;
3157
3158         batch->do_batch--;
3159         if (batch->do_batch) /* If stacked, do nothing. */
3160                 return;
3161
3162         if (!batch->memcg)
3163                 return;
3164         /*
3165          * This "batch->memcg" is valid without any css_get/put etc...
3166          * bacause we hide charges behind us.
3167          */
3168         if (batch->nr_pages)
3169                 res_counter_uncharge(&batch->memcg->res,
3170                                      batch->nr_pages * PAGE_SIZE);
3171         if (batch->memsw_nr_pages)
3172                 res_counter_uncharge(&batch->memcg->memsw,
3173                                      batch->memsw_nr_pages * PAGE_SIZE);
3174         memcg_oom_recover(batch->memcg);
3175         /* forget this pointer (for sanity check) */
3176         batch->memcg = NULL;
3177 }
3178
3179 #ifdef CONFIG_SWAP
3180 /*
3181  * called after __delete_from_swap_cache() and drop "page" account.
3182  * memcg information is recorded to swap_cgroup of "ent"
3183  */
3184 void
3185 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3186 {
3187         struct mem_cgroup *memcg;
3188         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3189
3190         if (!swapout) /* this was a swap cache but the swap is unused ! */
3191                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3192
3193         memcg = __mem_cgroup_uncharge_common(page, ctype);
3194
3195         /*
3196          * record memcg information,  if swapout && memcg != NULL,
3197          * mem_cgroup_get() was called in uncharge().
3198          */
3199         if (do_swap_account && swapout && memcg)
3200                 swap_cgroup_record(ent, css_id(&memcg->css));
3201 }
3202 #endif
3203
3204 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3205 /*
3206  * called from swap_entry_free(). remove record in swap_cgroup and
3207  * uncharge "memsw" account.
3208  */
3209 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3210 {
3211         struct mem_cgroup *memcg;
3212         unsigned short id;
3213
3214         if (!do_swap_account)
3215                 return;
3216
3217         id = swap_cgroup_record(ent, 0);
3218         rcu_read_lock();
3219         memcg = mem_cgroup_lookup(id);
3220         if (memcg) {
3221                 /*
3222                  * We uncharge this because swap is freed.
3223                  * This memcg can be obsolete one. We avoid calling css_tryget
3224                  */
3225                 if (!mem_cgroup_is_root(memcg))
3226                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3227                 mem_cgroup_swap_statistics(memcg, false);
3228                 mem_cgroup_put(memcg);
3229         }
3230         rcu_read_unlock();
3231 }
3232
3233 /**
3234  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3235  * @entry: swap entry to be moved
3236  * @from:  mem_cgroup which the entry is moved from
3237  * @to:  mem_cgroup which the entry is moved to
3238  * @need_fixup: whether we should fixup res_counters and refcounts.
3239  *
3240  * It succeeds only when the swap_cgroup's record for this entry is the same
3241  * as the mem_cgroup's id of @from.
3242  *
3243  * Returns 0 on success, -EINVAL on failure.
3244  *
3245  * The caller must have charged to @to, IOW, called res_counter_charge() about
3246  * both res and memsw, and called css_get().
3247  */
3248 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3249                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3250 {
3251         unsigned short old_id, new_id;
3252
3253         old_id = css_id(&from->css);
3254         new_id = css_id(&to->css);
3255
3256         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3257                 mem_cgroup_swap_statistics(from, false);
3258                 mem_cgroup_swap_statistics(to, true);
3259                 /*
3260                  * This function is only called from task migration context now.
3261                  * It postpones res_counter and refcount handling till the end
3262                  * of task migration(mem_cgroup_clear_mc()) for performance
3263                  * improvement. But we cannot postpone mem_cgroup_get(to)
3264                  * because if the process that has been moved to @to does
3265                  * swap-in, the refcount of @to might be decreased to 0.
3266                  */
3267                 mem_cgroup_get(to);
3268                 if (need_fixup) {
3269                         if (!mem_cgroup_is_root(from))
3270                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3271                         mem_cgroup_put(from);
3272                         /*
3273                          * we charged both to->res and to->memsw, so we should
3274                          * uncharge to->res.
3275                          */
3276                         if (!mem_cgroup_is_root(to))
3277                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3278                 }
3279                 return 0;
3280         }
3281         return -EINVAL;
3282 }
3283 #else
3284 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3285                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3286 {
3287         return -EINVAL;
3288 }
3289 #endif
3290
3291 /*
3292  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3293  * page belongs to.
3294  */
3295 int mem_cgroup_prepare_migration(struct page *page,
3296         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3297 {
3298         struct mem_cgroup *memcg = NULL;
3299         struct page_cgroup *pc;
3300         enum charge_type ctype;
3301         int ret = 0;
3302
3303         *ptr = NULL;
3304
3305         VM_BUG_ON(PageTransHuge(page));
3306         if (mem_cgroup_disabled())
3307                 return 0;
3308
3309         pc = lookup_page_cgroup(page);
3310         lock_page_cgroup(pc);
3311         if (PageCgroupUsed(pc)) {
3312                 memcg = pc->mem_cgroup;
3313                 css_get(&memcg->css);
3314                 /*
3315                  * At migrating an anonymous page, its mapcount goes down
3316                  * to 0 and uncharge() will be called. But, even if it's fully
3317                  * unmapped, migration may fail and this page has to be
3318                  * charged again. We set MIGRATION flag here and delay uncharge
3319                  * until end_migration() is called
3320                  *
3321                  * Corner Case Thinking
3322                  * A)
3323                  * When the old page was mapped as Anon and it's unmap-and-freed
3324                  * while migration was ongoing.
3325                  * If unmap finds the old page, uncharge() of it will be delayed
3326                  * until end_migration(). If unmap finds a new page, it's
3327                  * uncharged when it make mapcount to be 1->0. If unmap code
3328                  * finds swap_migration_entry, the new page will not be mapped
3329                  * and end_migration() will find it(mapcount==0).
3330                  *
3331                  * B)
3332                  * When the old page was mapped but migraion fails, the kernel
3333                  * remaps it. A charge for it is kept by MIGRATION flag even
3334                  * if mapcount goes down to 0. We can do remap successfully
3335                  * without charging it again.
3336                  *
3337                  * C)
3338                  * The "old" page is under lock_page() until the end of
3339                  * migration, so, the old page itself will not be swapped-out.
3340                  * If the new page is swapped out before end_migraton, our
3341                  * hook to usual swap-out path will catch the event.
3342                  */
3343                 if (PageAnon(page))
3344                         SetPageCgroupMigration(pc);
3345         }
3346         unlock_page_cgroup(pc);
3347         /*
3348          * If the page is not charged at this point,
3349          * we return here.
3350          */
3351         if (!memcg)
3352                 return 0;
3353
3354         *ptr = memcg;
3355         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3356         css_put(&memcg->css);/* drop extra refcnt */
3357         if (ret || *ptr == NULL) {
3358                 if (PageAnon(page)) {
3359                         lock_page_cgroup(pc);
3360                         ClearPageCgroupMigration(pc);
3361                         unlock_page_cgroup(pc);
3362                         /*
3363                          * The old page may be fully unmapped while we kept it.
3364                          */
3365                         mem_cgroup_uncharge_page(page);
3366                 }
3367                 return -ENOMEM;
3368         }
3369         /*
3370          * We charge new page before it's used/mapped. So, even if unlock_page()
3371          * is called before end_migration, we can catch all events on this new
3372          * page. In the case new page is migrated but not remapped, new page's
3373          * mapcount will be finally 0 and we call uncharge in end_migration().
3374          */
3375         pc = lookup_page_cgroup(newpage);
3376         if (PageAnon(page))
3377                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3378         else if (page_is_file_cache(page))
3379                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3380         else
3381                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3382         __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3383         return ret;
3384 }
3385
3386 /* remove redundant charge if migration failed*/
3387 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3388         struct page *oldpage, struct page *newpage, bool migration_ok)
3389 {
3390         struct page *used, *unused;
3391         struct page_cgroup *pc;
3392
3393         if (!memcg)
3394                 return;
3395         /* blocks rmdir() */
3396         cgroup_exclude_rmdir(&memcg->css);
3397         if (!migration_ok) {
3398                 used = oldpage;
3399                 unused = newpage;
3400         } else {
3401                 used = newpage;
3402                 unused = oldpage;
3403         }
3404         /*
3405          * We disallowed uncharge of pages under migration because mapcount
3406          * of the page goes down to zero, temporarly.
3407          * Clear the flag and check the page should be charged.
3408          */
3409         pc = lookup_page_cgroup(oldpage);
3410         lock_page_cgroup(pc);
3411         ClearPageCgroupMigration(pc);
3412         unlock_page_cgroup(pc);
3413
3414         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3415
3416         /*
3417          * If a page is a file cache, radix-tree replacement is very atomic
3418          * and we can skip this check. When it was an Anon page, its mapcount
3419          * goes down to 0. But because we added MIGRATION flage, it's not
3420          * uncharged yet. There are several case but page->mapcount check
3421          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3422          * check. (see prepare_charge() also)
3423          */
3424         if (PageAnon(used))
3425                 mem_cgroup_uncharge_page(used);
3426         /*
3427          * At migration, we may charge account against cgroup which has no
3428          * tasks.
3429          * So, rmdir()->pre_destroy() can be called while we do this charge.
3430          * In that case, we need to call pre_destroy() again. check it here.
3431          */
3432         cgroup_release_and_wakeup_rmdir(&memcg->css);
3433 }
3434
3435 /*
3436  * At replace page cache, newpage is not under any memcg but it's on
3437  * LRU. So, this function doesn't touch res_counter but handles LRU
3438  * in correct way. Both pages are locked so we cannot race with uncharge.
3439  */
3440 void mem_cgroup_replace_page_cache(struct page *oldpage,
3441                                   struct page *newpage)
3442 {
3443         struct mem_cgroup *memcg;
3444         struct page_cgroup *pc;
3445         struct zone *zone;
3446         enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3447         unsigned long flags;
3448
3449         if (mem_cgroup_disabled())
3450                 return;
3451
3452         pc = lookup_page_cgroup(oldpage);
3453         /* fix accounting on old pages */
3454         lock_page_cgroup(pc);
3455         memcg = pc->mem_cgroup;
3456         mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3457         ClearPageCgroupUsed(pc);
3458         unlock_page_cgroup(pc);
3459
3460         if (PageSwapBacked(oldpage))
3461                 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3462
3463         zone = page_zone(newpage);
3464         pc = lookup_page_cgroup(newpage);
3465         /*
3466          * Even if newpage->mapping was NULL before starting replacement,
3467          * the newpage may be on LRU(or pagevec for LRU) already. We lock
3468          * LRU while we overwrite pc->mem_cgroup.
3469          */
3470         spin_lock_irqsave(&zone->lru_lock, flags);
3471         if (PageLRU(newpage))
3472                 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3473         __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3474         if (PageLRU(newpage))
3475                 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3476         spin_unlock_irqrestore(&zone->lru_lock, flags);
3477 }
3478
3479 #ifdef CONFIG_DEBUG_VM
3480 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3481 {
3482         struct page_cgroup *pc;
3483
3484         pc = lookup_page_cgroup(page);
3485         if (likely(pc) && PageCgroupUsed(pc))
3486                 return pc;
3487         return NULL;
3488 }
3489
3490 bool mem_cgroup_bad_page_check(struct page *page)
3491 {
3492         if (mem_cgroup_disabled())
3493                 return false;
3494
3495         return lookup_page_cgroup_used(page) != NULL;
3496 }
3497
3498 void mem_cgroup_print_bad_page(struct page *page)
3499 {
3500         struct page_cgroup *pc;
3501
3502         pc = lookup_page_cgroup_used(page);
3503         if (pc) {
3504                 int ret = -1;
3505                 char *path;
3506
3507                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3508                        pc, pc->flags, pc->mem_cgroup);
3509
3510                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3511                 if (path) {
3512                         rcu_read_lock();
3513                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3514                                                         path, PATH_MAX);
3515                         rcu_read_unlock();
3516                 }
3517
3518                 printk(KERN_CONT "(%s)\n",
3519                                 (ret < 0) ? "cannot get the path" : path);
3520                 kfree(path);
3521         }
3522 }
3523 #endif
3524
3525 static DEFINE_MUTEX(set_limit_mutex);
3526
3527 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3528                                 unsigned long long val)
3529 {
3530         int retry_count;
3531         u64 memswlimit, memlimit;
3532         int ret = 0;
3533         int children = mem_cgroup_count_children(memcg);
3534         u64 curusage, oldusage;
3535         int enlarge;
3536
3537         /*
3538          * For keeping hierarchical_reclaim simple, how long we should retry
3539          * is depends on callers. We set our retry-count to be function
3540          * of # of children which we should visit in this loop.
3541          */
3542         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3543
3544         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3545
3546         enlarge = 0;
3547         while (retry_count) {
3548                 if (signal_pending(current)) {
3549                         ret = -EINTR;
3550                         break;
3551                 }
3552                 /*
3553                  * Rather than hide all in some function, I do this in
3554                  * open coded manner. You see what this really does.
3555                  * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3556                  */
3557                 mutex_lock(&set_limit_mutex);
3558                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3559                 if (memswlimit < val) {
3560                         ret = -EINVAL;
3561                         mutex_unlock(&set_limit_mutex);
3562                         break;
3563                 }
3564
3565                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3566                 if (memlimit < val)
3567                         enlarge = 1;
3568
3569                 ret = res_counter_set_limit(&memcg->res, val);
3570                 if (!ret) {
3571                         if (memswlimit == val)
3572                                 memcg->memsw_is_minimum = true;
3573                         else
3574                                 memcg->memsw_is_minimum = false;
3575                 }
3576                 mutex_unlock(&set_limit_mutex);
3577
3578                 if (!ret)
3579                         break;
3580
3581                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3582                                                 MEM_CGROUP_RECLAIM_SHRINK,
3583                                                 NULL);
3584                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3585                 /* Usage is reduced ? */
3586                 if (curusage >= oldusage)
3587                         retry_count--;
3588                 else
3589                         oldusage = curusage;
3590         }
3591         if (!ret && enlarge)
3592                 memcg_oom_recover(memcg);
3593
3594         return ret;
3595 }
3596
3597 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3598                                         unsigned long long val)
3599 {
3600         int retry_count;
3601         u64 memlimit, memswlimit, oldusage, curusage;
3602         int children = mem_cgroup_count_children(memcg);
3603         int ret = -EBUSY;
3604         int enlarge = 0;
3605
3606         /* see mem_cgroup_resize_res_limit */
3607         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3608         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3609         while (retry_count) {
3610                 if (signal_pending(current)) {
3611                         ret = -EINTR;
3612                         break;
3613                 }
3614                 /*
3615                  * Rather than hide all in some function, I do this in
3616                  * open coded manner. You see what this really does.
3617                  * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3618                  */
3619                 mutex_lock(&set_limit_mutex);
3620                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3621                 if (memlimit > val) {
3622                         ret = -EINVAL;
3623                         mutex_unlock(&set_limit_mutex);
3624                         break;
3625                 }
3626                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3627                 if (memswlimit < val)
3628                         enlarge = 1;
3629                 ret = res_counter_set_limit(&memcg->memsw, val);
3630                 if (!ret) {
3631                         if (memlimit == val)
3632                                 memcg->memsw_is_minimum = true;
3633                         else
3634                                 memcg->memsw_is_minimum = false;
3635                 }
3636                 mutex_unlock(&set_limit_mutex);
3637
3638                 if (!ret)
3639                         break;
3640
3641                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3642                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3643                                                 MEM_CGROUP_RECLAIM_SHRINK,
3644                                                 NULL);
3645                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3646                 /* Usage is reduced ? */
3647                 if (curusage >= oldusage)
3648                         retry_count--;
3649                 else
3650                         oldusage = curusage;
3651         }
3652         if (!ret && enlarge)
3653                 memcg_oom_recover(memcg);
3654         return ret;
3655 }
3656
3657 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3658                                             gfp_t gfp_mask,
3659                                             unsigned long *total_scanned)
3660 {
3661         unsigned long nr_reclaimed = 0;
3662         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3663         unsigned long reclaimed;
3664         int loop = 0;
3665         struct mem_cgroup_tree_per_zone *mctz;
3666         unsigned long long excess;
3667         unsigned long nr_scanned;
3668
3669         if (order > 0)
3670                 return 0;
3671
3672         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3673         /*
3674          * This loop can run a while, specially if mem_cgroup's continuously
3675          * keep exceeding their soft limit and putting the system under
3676          * pressure
3677          */
3678         do {
3679                 if (next_mz)
3680                         mz = next_mz;
3681                 else
3682                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3683                 if (!mz)
3684                         break;
3685
3686                 nr_scanned = 0;
3687                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3688                                                 gfp_mask,
3689                                                 MEM_CGROUP_RECLAIM_SOFT,
3690                                                 &nr_scanned);
3691                 nr_reclaimed += reclaimed;
3692                 *total_scanned += nr_scanned;
3693                 spin_lock(&mctz->lock);
3694
3695                 /*
3696                  * If we failed to reclaim anything from this memory cgroup
3697                  * it is time to move on to the next cgroup
3698                  */
3699                 next_mz = NULL;
3700                 if (!reclaimed) {
3701                         do {
3702                                 /*
3703                                  * Loop until we find yet another one.
3704                                  *
3705                                  * By the time we get the soft_limit lock
3706                                  * again, someone might have aded the
3707                                  * group back on the RB tree. Iterate to
3708                                  * make sure we get a different mem.
3709                                  * mem_cgroup_largest_soft_limit_node returns
3710                                  * NULL if no other cgroup is present on
3711                                  * the tree
3712                                  */
3713                                 next_mz =
3714                                 __mem_cgroup_largest_soft_limit_node(mctz);
3715                                 if (next_mz == mz)
3716                                         css_put(&next_mz->mem->css);
3717                                 else /* next_mz == NULL or other memcg */
3718                                         break;
3719                         } while (1);
3720                 }
3721                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3722                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3723                 /*
3724                  * One school of thought says that we should not add
3725                  * back the node to the tree if reclaim returns 0.
3726                  * But our reclaim could return 0, simply because due
3727                  * to priority we are exposing a smaller subset of
3728                  * memory to reclaim from. Consider this as a longer
3729                  * term TODO.
3730                  */
3731                 /* If excess == 0, no tree ops */
3732                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3733                 spin_unlock(&mctz->lock);
3734                 css_put(&mz->mem->css);
3735                 loop++;
3736                 /*
3737                  * Could not reclaim anything and there are no more
3738                  * mem cgroups to try or we seem to be looping without
3739                  * reclaiming anything.
3740                  */
3741                 if (!nr_reclaimed &&
3742                         (next_mz == NULL ||
3743                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3744                         break;
3745         } while (!nr_reclaimed);
3746         if (next_mz)
3747                 css_put(&next_mz->mem->css);
3748         return nr_reclaimed;
3749 }
3750
3751 /*
3752  * This routine traverse page_cgroup in given list and drop them all.
3753  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3754  */
3755 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3756                                 int node, int zid, enum lru_list lru)
3757 {
3758         struct zone *zone;
3759         struct mem_cgroup_per_zone *mz;
3760         struct page_cgroup *pc, *busy;
3761         unsigned long flags, loop;
3762         struct list_head *list;
3763         int ret = 0;
3764
3765         zone = &NODE_DATA(node)->node_zones[zid];
3766         mz = mem_cgroup_zoneinfo(memcg, node, zid);
3767         list = &mz->lists[lru];
3768
3769         loop = MEM_CGROUP_ZSTAT(mz, lru);
3770         /* give some margin against EBUSY etc...*/
3771         loop += 256;
3772         busy = NULL;
3773         while (loop--) {
3774                 struct page *page;
3775
3776                 ret = 0;
3777                 spin_lock_irqsave(&zone->lru_lock, flags);
3778                 if (list_empty(list)) {
3779                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3780                         break;
3781                 }
3782                 pc = list_entry(list->prev, struct page_cgroup, lru);
3783                 if (busy == pc) {
3784                         list_move(&pc->lru, list);
3785                         busy = NULL;
3786                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3787                         continue;
3788                 }
3789                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3790
3791                 page = lookup_cgroup_page(pc);
3792
3793                 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3794                 if (ret == -ENOMEM)
3795                         break;
3796
3797                 if (ret == -EBUSY || ret == -EINVAL) {
3798                         /* found lock contention or "pc" is obsolete. */
3799                         busy = pc;
3800                         cond_resched();
3801                 } else
3802                         busy = NULL;
3803         }
3804
3805         if (!ret && !list_empty(list))
3806                 return -EBUSY;
3807         return ret;
3808 }
3809
3810 /*
3811  * make mem_cgroup's charge to be 0 if there is no task.
3812  * This enables deleting this mem_cgroup.
3813  */
3814 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3815 {
3816         int ret;
3817         int node, zid, shrink;
3818         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3819         struct cgroup *cgrp = memcg->css.cgroup;
3820
3821         css_get(&memcg->css);
3822
3823         shrink = 0;
3824         /* should free all ? */
3825         if (free_all)
3826                 goto try_to_free;
3827 move_account:
3828         do {
3829                 ret = -EBUSY;
3830                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3831                         goto out;
3832                 ret = -EINTR;
3833                 if (signal_pending(current))
3834                         goto out;
3835                 /* This is for making all *used* pages to be on LRU. */
3836                 lru_add_drain_all();
3837                 drain_all_stock_sync(memcg);
3838                 ret = 0;
3839                 mem_cgroup_start_move(memcg);
3840                 for_each_node_state(node, N_HIGH_MEMORY) {
3841                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3842                                 enum lru_list l;
3843                                 for_each_lru(l) {
3844                                         ret = mem_cgroup_force_empty_list(memcg,
3845                                                         node, zid, l);
3846                                         if (ret)
3847                                                 break;
3848                                 }
3849                         }
3850                         if (ret)
3851                                 break;
3852                 }
3853                 mem_cgroup_end_move(memcg);
3854                 memcg_oom_recover(memcg);
3855                 /* it seems parent cgroup doesn't have enough mem */
3856                 if (ret == -ENOMEM)
3857                         goto try_to_free;
3858                 cond_resched();
3859         /* "ret" should also be checked to ensure all lists are empty. */
3860         } while (memcg->res.usage > 0 || ret);
3861 out:
3862         css_put(&memcg->css);
3863         return ret;
3864
3865 try_to_free:
3866         /* returns EBUSY if there is a task or if we come here twice. */
3867         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3868                 ret = -EBUSY;
3869                 goto out;
3870         }
3871         /* we call try-to-free pages for make this cgroup empty */
3872         lru_add_drain_all();
3873         /* try to free all pages in this cgroup */
3874         shrink = 1;
3875         while (nr_retries && memcg->res.usage > 0) {
3876                 int progress;
3877
3878                 if (signal_pending(current)) {
3879                         ret = -EINTR;
3880                         goto out;
3881                 }
3882                 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3883                                                 false);
3884                 if (!progress) {
3885                         nr_retries--;
3886                         /* maybe some writeback is necessary */
3887                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3888                 }
3889
3890         }
3891         lru_add_drain();
3892         /* try move_account...there may be some *locked* pages. */
3893         goto move_account;
3894 }
3895
3896 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3897 {
3898         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3899 }
3900
3901
3902 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3903 {
3904         return mem_cgroup_from_cont(cont)->use_hierarchy;
3905 }
3906
3907 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3908                                         u64 val)
3909 {
3910         int retval = 0;
3911         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3912         struct cgroup *parent = cont->parent;
3913         struct mem_cgroup *parent_memcg = NULL;
3914
3915         if (parent)
3916                 parent_memcg = mem_cgroup_from_cont(parent);
3917
3918         cgroup_lock();
3919         /*
3920          * If parent's use_hierarchy is set, we can't make any modifications
3921          * in the child subtrees. If it is unset, then the change can
3922          * occur, provided the current cgroup has no children.
3923          *
3924          * For the root cgroup, parent_mem is NULL, we allow value to be
3925          * set if there are no children.
3926          */
3927         if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3928                                 (val == 1 || val == 0)) {
3929                 if (list_empty(&cont->children))
3930                         memcg->use_hierarchy = val;
3931                 else
3932                         retval = -EBUSY;
3933         } else
3934                 retval = -EINVAL;
3935         cgroup_unlock();
3936
3937         return retval;
3938 }
3939
3940
3941 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3942                                                enum mem_cgroup_stat_index idx)
3943 {
3944         struct mem_cgroup *iter;
3945         long val = 0;
3946
3947         /* Per-cpu values can be negative, use a signed accumulator */
3948         for_each_mem_cgroup_tree(iter, memcg)
3949                 val += mem_cgroup_read_stat(iter, idx);
3950
3951         if (val < 0) /* race ? */
3952                 val = 0;
3953         return val;
3954 }
3955
3956 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3957 {
3958         u64 val;
3959
3960         if (!mem_cgroup_is_root(memcg)) {
3961                 if (!swap)
3962                         return res_counter_read_u64(&memcg->res, RES_USAGE);
3963                 else
3964                         return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3965         }
3966
3967         val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3968         val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3969
3970         if (swap)
3971                 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3972
3973         return val << PAGE_SHIFT;
3974 }
3975
3976 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3977 {
3978         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3979         u64 val;
3980         int type, name;
3981
3982         type = MEMFILE_TYPE(cft->private);
3983         name = MEMFILE_ATTR(cft->private);
3984         switch (type) {
3985         case _MEM:
3986                 if (name == RES_USAGE)
3987                         val = mem_cgroup_usage(memcg, false);
3988                 else
3989                         val = res_counter_read_u64(&memcg->res, name);
3990                 break;
3991         case _MEMSWAP:
3992                 if (name == RES_USAGE)
3993                         val = mem_cgroup_usage(memcg, true);
3994                 else
3995                         val = res_counter_read_u64(&memcg->memsw, name);
3996                 break;
3997         default:
3998                 BUG();
3999                 break;
4000         }
4001         return val;
4002 }
4003 /*
4004  * The user of this function is...
4005  * RES_LIMIT.
4006  */
4007 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4008                             const char *buffer)
4009 {
4010         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4011         int type, name;
4012         unsigned long long val;
4013         int ret;
4014
4015         type = MEMFILE_TYPE(cft->private);
4016         name = MEMFILE_ATTR(cft->private);
4017         switch (name) {
4018         case RES_LIMIT:
4019                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4020                         ret = -EINVAL;
4021                         break;
4022                 }
4023                 /* This function does all necessary parse...reuse it */
4024                 ret = res_counter_memparse_write_strategy(buffer, &val);
4025                 if (ret)
4026                         break;
4027                 if (type == _MEM)
4028                         ret = mem_cgroup_resize_limit(memcg, val);
4029                 else
4030                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
4031                 break;
4032         case RES_SOFT_LIMIT:
4033                 ret = res_counter_memparse_write_strategy(buffer, &val);
4034                 if (ret)
4035                         break;
4036                 /*
4037                  * For memsw, soft limits are hard to implement in terms
4038                  * of semantics, for now, we support soft limits for
4039                  * control without swap
4040                  */
4041                 if (type == _MEM)
4042                         ret = res_counter_set_soft_limit(&memcg->res, val);
4043                 else
4044                         ret = -EINVAL;
4045                 break;
4046         default:
4047                 ret = -EINVAL; /* should be BUG() ? */
4048                 break;
4049         }
4050         return ret;
4051 }
4052
4053 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4054                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4055 {
4056         struct cgroup *cgroup;
4057         unsigned long long min_limit, min_memsw_limit, tmp;
4058
4059         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4060         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4061         cgroup = memcg->css.cgroup;
4062         if (!memcg->use_hierarchy)
4063                 goto out;
4064
4065         while (cgroup->parent) {
4066                 cgroup = cgroup->parent;
4067                 memcg = mem_cgroup_from_cont(cgroup);
4068                 if (!memcg->use_hierarchy)
4069                         break;
4070                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4071                 min_limit = min(min_limit, tmp);
4072                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4073                 min_memsw_limit = min(min_memsw_limit, tmp);
4074         }
4075 out:
4076         *mem_limit = min_limit;
4077         *memsw_limit = min_memsw_limit;
4078         return;
4079 }
4080
4081 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4082 {
4083         struct mem_cgroup *memcg;
4084         int type, name;
4085
4086         memcg = mem_cgroup_from_cont(cont);
4087         type = MEMFILE_TYPE(event);
4088         name = MEMFILE_ATTR(event);
4089         switch (name) {
4090         case RES_MAX_USAGE:
4091                 if (type == _MEM)
4092                         res_counter_reset_max(&memcg->res);
4093                 else
4094                         res_counter_reset_max(&memcg->memsw);
4095                 break;
4096         case RES_FAILCNT:
4097                 if (type == _MEM)
4098                         res_counter_reset_failcnt(&memcg->res);
4099                 else
4100                         res_counter_reset_failcnt(&memcg->memsw);
4101                 break;
4102         }
4103
4104         return 0;
4105 }
4106
4107 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4108                                         struct cftype *cft)
4109 {
4110         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4111 }
4112
4113 #ifdef CONFIG_MMU
4114 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4115                                         struct cftype *cft, u64 val)
4116 {
4117         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4118
4119         if (val >= (1 << NR_MOVE_TYPE))
4120                 return -EINVAL;
4121         /*
4122          * We check this value several times in both in can_attach() and
4123          * attach(), so we need cgroup lock to prevent this value from being
4124          * inconsistent.
4125          */
4126         cgroup_lock();
4127         memcg->move_charge_at_immigrate = val;
4128         cgroup_unlock();
4129
4130         return 0;
4131 }
4132 #else
4133 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4134                                         struct cftype *cft, u64 val)
4135 {
4136         return -ENOSYS;
4137 }
4138 #endif
4139
4140
4141 /* For read statistics */
4142 enum {
4143         MCS_CACHE,
4144         MCS_RSS,
4145         MCS_FILE_MAPPED,
4146         MCS_PGPGIN,
4147         MCS_PGPGOUT,
4148         MCS_SWAP,
4149         MCS_PGFAULT,
4150         MCS_PGMAJFAULT,
4151         MCS_INACTIVE_ANON,
4152         MCS_ACTIVE_ANON,
4153         MCS_INACTIVE_FILE,
4154         MCS_ACTIVE_FILE,
4155         MCS_UNEVICTABLE,
4156         NR_MCS_STAT,
4157 };
4158
4159 struct mcs_total_stat {
4160         s64 stat[NR_MCS_STAT];
4161 };
4162
4163 struct {
4164         char *local_name;
4165         char *total_name;
4166 } memcg_stat_strings[NR_MCS_STAT] = {
4167         {"cache", "total_cache"},
4168         {"rss", "total_rss"},
4169         {"mapped_file", "total_mapped_file"},
4170         {"pgpgin", "total_pgpgin"},
4171         {"pgpgout", "total_pgpgout"},
4172         {"swap", "total_swap"},
4173         {"pgfault", "total_pgfault"},
4174         {"pgmajfault", "total_pgmajfault"},
4175         {"inactive_anon", "total_inactive_anon"},
4176         {"active_anon", "total_active_anon"},
4177         {"inactive_file", "total_inactive_file"},
4178         {"active_file", "total_active_file"},
4179         {"unevictable", "total_unevictable"}
4180 };
4181
4182
4183 static void
4184 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4185 {
4186         s64 val;
4187
4188         /* per cpu stat */
4189         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4190         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4191         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4192         s->stat[MCS_RSS] += val * PAGE_SIZE;
4193         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4194         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4195         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4196         s->stat[MCS_PGPGIN] += val;
4197         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4198         s->stat[MCS_PGPGOUT] += val;
4199         if (do_swap_account) {
4200                 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4201                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4202         }
4203         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4204         s->stat[MCS_PGFAULT] += val;
4205         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4206         s->stat[MCS_PGMAJFAULT] += val;
4207
4208         /* per zone stat */
4209         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4210         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4211         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4212         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4213         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4214         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4215         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4216         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4217         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4218         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4219 }
4220
4221 static void
4222 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4223 {
4224         struct mem_cgroup *iter;
4225
4226         for_each_mem_cgroup_tree(iter, memcg)
4227                 mem_cgroup_get_local_stat(iter, s);
4228 }
4229
4230 #ifdef CONFIG_NUMA
4231 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4232 {
4233         int nid;
4234         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4235         unsigned long node_nr;
4236         struct cgroup *cont = m->private;
4237         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4238
4239         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4240         seq_printf(m, "total=%lu", total_nr);
4241         for_each_node_state(nid, N_HIGH_MEMORY) {
4242                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4243                 seq_printf(m, " N%d=%lu", nid, node_nr);
4244         }
4245         seq_putc(m, '\n');
4246
4247         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4248         seq_printf(m, "file=%lu", file_nr);
4249         for_each_node_state(nid, N_HIGH_MEMORY) {
4250                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4251                                 LRU_ALL_FILE);
4252                 seq_printf(m, " N%d=%lu", nid, node_nr);
4253         }
4254         seq_putc(m, '\n');
4255
4256         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4257         seq_printf(m, "anon=%lu", anon_nr);
4258         for_each_node_state(nid, N_HIGH_MEMORY) {
4259                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4260                                 LRU_ALL_ANON);
4261                 seq_printf(m, " N%d=%lu", nid, node_nr);
4262         }
4263         seq_putc(m, '\n');
4264
4265         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4266         seq_printf(m, "unevictable=%lu", unevictable_nr);
4267         for_each_node_state(nid, N_HIGH_MEMORY) {
4268                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4269                                 BIT(LRU_UNEVICTABLE));
4270                 seq_printf(m, " N%d=%lu", nid, node_nr);
4271         }
4272         seq_putc(m, '\n');
4273         return 0;
4274 }
4275 #endif /* CONFIG_NUMA */
4276
4277 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4278                                  struct cgroup_map_cb *cb)
4279 {
4280         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4281         struct mcs_total_stat mystat;
4282         int i;
4283
4284         memset(&mystat, 0, sizeof(mystat));
4285         mem_cgroup_get_local_stat(mem_cont, &mystat);
4286
4287
4288         for (i = 0; i < NR_MCS_STAT; i++) {
4289                 if (i == MCS_SWAP && !do_swap_account)
4290                         continue;
4291                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4292         }
4293
4294         /* Hierarchical information */
4295         {
4296                 unsigned long long limit, memsw_limit;
4297                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4298                 cb->fill(cb, "hierarchical_memory_limit", limit);
4299                 if (do_swap_account)
4300                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4301         }
4302
4303         memset(&mystat, 0, sizeof(mystat));
4304         mem_cgroup_get_total_stat(mem_cont, &mystat);
4305         for (i = 0; i < NR_MCS_STAT; i++) {
4306                 if (i == MCS_SWAP && !do_swap_account)
4307                         continue;
4308                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4309         }
4310
4311 #ifdef CONFIG_DEBUG_VM
4312         {
4313                 int nid, zid;
4314                 struct mem_cgroup_per_zone *mz;
4315                 unsigned long recent_rotated[2] = {0, 0};
4316                 unsigned long recent_scanned[2] = {0, 0};
4317
4318                 for_each_online_node(nid)
4319                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4320                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4321
4322                                 recent_rotated[0] +=
4323                                         mz->reclaim_stat.recent_rotated[0];
4324                                 recent_rotated[1] +=
4325                                         mz->reclaim_stat.recent_rotated[1];
4326                                 recent_scanned[0] +=
4327                                         mz->reclaim_stat.recent_scanned[0];
4328                                 recent_scanned[1] +=
4329                                         mz->reclaim_stat.recent_scanned[1];
4330                         }
4331                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4332                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4333                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4334                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4335         }
4336 #endif
4337
4338         return 0;
4339 }
4340
4341 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4342 {
4343         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4344
4345         return mem_cgroup_swappiness(memcg);
4346 }
4347
4348 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4349                                        u64 val)
4350 {
4351         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4352         struct mem_cgroup *parent;
4353
4354         if (val > 100)
4355                 return -EINVAL;
4356
4357         if (cgrp->parent == NULL)
4358                 return -EINVAL;
4359
4360         parent = mem_cgroup_from_cont(cgrp->parent);
4361
4362         cgroup_lock();
4363
4364         /* If under hierarchy, only empty-root can set this value */
4365         if ((parent->use_hierarchy) ||
4366             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4367                 cgroup_unlock();
4368                 return -EINVAL;
4369         }
4370
4371         memcg->swappiness = val;
4372
4373         cgroup_unlock();
4374
4375         return 0;
4376 }
4377
4378 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4379 {
4380         struct mem_cgroup_threshold_ary *t;
4381         u64 usage;
4382         int i;
4383
4384         rcu_read_lock();
4385         if (!swap)
4386                 t = rcu_dereference(memcg->thresholds.primary);
4387         else
4388                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4389
4390         if (!t)
4391                 goto unlock;
4392
4393         usage = mem_cgroup_usage(memcg, swap);
4394
4395         /*
4396          * current_threshold points to threshold just below usage.
4397          * If it's not true, a threshold was crossed after last
4398          * call of __mem_cgroup_threshold().
4399          */
4400         i = t->current_threshold;
4401
4402         /*
4403          * Iterate backward over array of thresholds starting from
4404          * current_threshold and check if a threshold is crossed.
4405          * If none of thresholds below usage is crossed, we read
4406          * only one element of the array here.
4407          */
4408         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4409                 eventfd_signal(t->entries[i].eventfd, 1);
4410
4411         /* i = current_threshold + 1 */
4412         i++;
4413
4414         /*
4415          * Iterate forward over array of thresholds starting from
4416          * current_threshold+1 and check if a threshold is crossed.
4417          * If none of thresholds above usage is crossed, we read
4418          * only one element of the array here.
4419          */
4420         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4421                 eventfd_signal(t->entries[i].eventfd, 1);
4422
4423         /* Update current_threshold */
4424         t->current_threshold = i - 1;
4425 unlock:
4426         rcu_read_unlock();
4427 }
4428
4429 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4430 {
4431         while (memcg) {
4432                 __mem_cgroup_threshold(memcg, false);
4433                 if (do_swap_account)
4434                         __mem_cgroup_threshold(memcg, true);
4435
4436                 memcg = parent_mem_cgroup(memcg);
4437         }
4438 }
4439
4440 static int compare_thresholds(const void *a, const void *b)
4441 {
4442         const struct mem_cgroup_threshold *_a = a;
4443         const struct mem_cgroup_threshold *_b = b;
4444
4445         return _a->threshold - _b->threshold;
4446 }
4447
4448 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4449 {
4450         struct mem_cgroup_eventfd_list *ev;
4451
4452         list_for_each_entry(ev, &memcg->oom_notify, list)
4453                 eventfd_signal(ev->eventfd, 1);
4454         return 0;
4455 }
4456
4457 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4458 {
4459         struct mem_cgroup *iter;
4460
4461         for_each_mem_cgroup_tree(iter, memcg)
4462                 mem_cgroup_oom_notify_cb(iter);
4463 }
4464
4465 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4466         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4467 {
4468         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4469         struct mem_cgroup_thresholds *thresholds;
4470         struct mem_cgroup_threshold_ary *new;
4471         int type = MEMFILE_TYPE(cft->private);
4472         u64 threshold, usage;
4473         int i, size, ret;
4474
4475         ret = res_counter_memparse_write_strategy(args, &threshold);
4476         if (ret)
4477                 return ret;
4478
4479         mutex_lock(&memcg->thresholds_lock);
4480
4481         if (type == _MEM)
4482                 thresholds = &memcg->thresholds;
4483         else if (type == _MEMSWAP)
4484                 thresholds = &memcg->memsw_thresholds;
4485         else
4486                 BUG();
4487
4488         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4489
4490         /* Check if a threshold crossed before adding a new one */
4491         if (thresholds->primary)
4492                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4493
4494         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4495
4496         /* Allocate memory for new array of thresholds */
4497         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4498                         GFP_KERNEL);
4499         if (!new) {
4500                 ret = -ENOMEM;
4501                 goto unlock;
4502         }
4503         new->size = size;
4504
4505         /* Copy thresholds (if any) to new array */
4506         if (thresholds->primary) {
4507                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4508                                 sizeof(struct mem_cgroup_threshold));
4509         }
4510
4511         /* Add new threshold */
4512         new->entries[size - 1].eventfd = eventfd;
4513         new->entries[size - 1].threshold = threshold;
4514
4515         /* Sort thresholds. Registering of new threshold isn't time-critical */
4516         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4517                         compare_thresholds, NULL);
4518
4519         /* Find current threshold */
4520         new->current_threshold = -1;
4521         for (i = 0; i < size; i++) {
4522                 if (new->entries[i].threshold < usage) {
4523                         /*
4524                          * new->current_threshold will not be used until
4525                          * rcu_assign_pointer(), so it's safe to increment
4526                          * it here.
4527                          */
4528                         ++new->current_threshold;
4529                 }
4530         }
4531
4532         /* Free old spare buffer and save old primary buffer as spare */
4533         kfree(thresholds->spare);
4534         thresholds->spare = thresholds->primary;
4535
4536         rcu_assign_pointer(thresholds->primary, new);
4537
4538         /* To be sure that nobody uses thresholds */
4539         synchronize_rcu();
4540
4541 unlock:
4542         mutex_unlock(&memcg->thresholds_lock);
4543
4544         return ret;
4545 }
4546
4547 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4548         struct cftype *cft, struct eventfd_ctx *eventfd)
4549 {
4550         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4551         struct mem_cgroup_thresholds *thresholds;
4552         struct mem_cgroup_threshold_ary *new;
4553         int type = MEMFILE_TYPE(cft->private);
4554         u64 usage;
4555         int i, j, size;
4556
4557         mutex_lock(&memcg->thresholds_lock);
4558         if (type == _MEM)
4559                 thresholds = &memcg->thresholds;
4560         else if (type == _MEMSWAP)
4561                 thresholds = &memcg->memsw_thresholds;
4562         else
4563                 BUG();
4564
4565         /*
4566          * Something went wrong if we trying to unregister a threshold
4567          * if we don't have thresholds
4568          */
4569         BUG_ON(!thresholds);
4570
4571         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4572
4573         /* Check if a threshold crossed before removing */
4574         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4575
4576         /* Calculate new number of threshold */
4577         size = 0;
4578         for (i = 0; i < thresholds->primary->size; i++) {
4579                 if (thresholds->primary->entries[i].eventfd != eventfd)
4580                         size++;
4581         }
4582
4583         new = thresholds->spare;
4584
4585         /* Set thresholds array to NULL if we don't have thresholds */
4586         if (!size) {
4587                 kfree(new);
4588                 new = NULL;
4589                 goto swap_buffers;
4590         }
4591
4592         new->size = size;
4593
4594         /* Copy thresholds and find current threshold */
4595         new->current_threshold = -1;
4596         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4597                 if (thresholds->primary->entries[i].eventfd == eventfd)
4598                         continue;
4599
4600                 new->entries[j] = thresholds->primary->entries[i];
4601                 if (new->entries[j].threshold < usage) {
4602                         /*
4603                          * new->current_threshold will not be used
4604                          * until rcu_assign_pointer(), so it's safe to increment
4605                          * it here.
4606                          */
4607                         ++new->current_threshold;
4608                 }
4609                 j++;
4610         }
4611
4612 swap_buffers:
4613         /* Swap primary and spare array */
4614         thresholds->spare = thresholds->primary;
4615         rcu_assign_pointer(thresholds->primary, new);
4616
4617         /* To be sure that nobody uses thresholds */
4618         synchronize_rcu();
4619
4620         mutex_unlock(&memcg->thresholds_lock);
4621 }
4622
4623 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4624         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4625 {
4626         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4627         struct mem_cgroup_eventfd_list *event;
4628         int type = MEMFILE_TYPE(cft->private);
4629
4630         BUG_ON(type != _OOM_TYPE);
4631         event = kmalloc(sizeof(*event), GFP_KERNEL);
4632         if (!event)
4633                 return -ENOMEM;
4634
4635         spin_lock(&memcg_oom_lock);
4636
4637         event->eventfd = eventfd;
4638         list_add(&event->list, &memcg->oom_notify);
4639
4640         /* already in OOM ? */
4641         if (atomic_read(&memcg->under_oom))
4642                 eventfd_signal(eventfd, 1);
4643         spin_unlock(&memcg_oom_lock);
4644
4645         return 0;
4646 }
4647
4648 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4649         struct cftype *cft, struct eventfd_ctx *eventfd)
4650 {
4651         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4652         struct mem_cgroup_eventfd_list *ev, *tmp;
4653         int type = MEMFILE_TYPE(cft->private);
4654
4655         BUG_ON(type != _OOM_TYPE);
4656
4657         spin_lock(&memcg_oom_lock);
4658
4659         list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4660                 if (ev->eventfd == eventfd) {
4661                         list_del(&ev->list);
4662                         kfree(ev);
4663                 }
4664         }
4665
4666         spin_unlock(&memcg_oom_lock);
4667 }
4668
4669 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4670         struct cftype *cft,  struct cgroup_map_cb *cb)
4671 {
4672         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4673
4674         cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4675
4676         if (atomic_read(&memcg->under_oom))
4677                 cb->fill(cb, "under_oom", 1);
4678         else
4679                 cb->fill(cb, "under_oom", 0);
4680         return 0;
4681 }
4682
4683 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4684         struct cftype *cft, u64 val)
4685 {
4686         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4687         struct mem_cgroup *parent;
4688
4689         /* cannot set to root cgroup and only 0 and 1 are allowed */
4690         if (!cgrp->parent || !((val == 0) || (val == 1)))
4691                 return -EINVAL;
4692
4693         parent = mem_cgroup_from_cont(cgrp->parent);
4694
4695         cgroup_lock();
4696         /* oom-kill-disable is a flag for subhierarchy. */
4697         if ((parent->use_hierarchy) ||
4698             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4699                 cgroup_unlock();
4700                 return -EINVAL;
4701         }
4702         memcg->oom_kill_disable = val;
4703         if (!val)
4704                 memcg_oom_recover(memcg);
4705         cgroup_unlock();
4706         return 0;
4707 }
4708
4709 #ifdef CONFIG_NUMA
4710 static const struct file_operations mem_control_numa_stat_file_operations = {
4711         .read = seq_read,
4712         .llseek = seq_lseek,
4713         .release = single_release,
4714 };
4715
4716 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4717 {
4718         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4719
4720         file->f_op = &mem_control_numa_stat_file_operations;
4721         return single_open(file, mem_control_numa_stat_show, cont);
4722 }
4723 #endif /* CONFIG_NUMA */
4724
4725 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4726 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4727 {
4728         /*
4729          * Part of this would be better living in a separate allocation
4730          * function, leaving us with just the cgroup tree population work.
4731          * We, however, depend on state such as network's proto_list that
4732          * is only initialized after cgroup creation. I found the less
4733          * cumbersome way to deal with it to defer it all to populate time
4734          */
4735         return mem_cgroup_sockets_init(cont, ss);
4736 };
4737
4738 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4739                                 struct cgroup *cont)
4740 {
4741         mem_cgroup_sockets_destroy(cont, ss);
4742 }
4743 #else
4744 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4745 {
4746         return 0;
4747 }
4748
4749 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4750                                 struct cgroup *cont)
4751 {
4752 }
4753 #endif
4754
4755 static struct cftype mem_cgroup_files[] = {
4756         {
4757                 .name = "usage_in_bytes",
4758                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4759                 .read_u64 = mem_cgroup_read,
4760                 .register_event = mem_cgroup_usage_register_event,
4761                 .unregister_event = mem_cgroup_usage_unregister_event,
4762         },
4763         {
4764                 .name = "max_usage_in_bytes",
4765                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4766                 .trigger = mem_cgroup_reset,
4767                 .read_u64 = mem_cgroup_read,
4768         },
4769         {
4770                 .name = "limit_in_bytes",
4771                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4772                 .write_string = mem_cgroup_write,
4773                 .read_u64 = mem_cgroup_read,
4774         },
4775         {
4776                 .name = "soft_limit_in_bytes",
4777                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4778                 .write_string = mem_cgroup_write,
4779                 .read_u64 = mem_cgroup_read,
4780         },
4781         {
4782                 .name = "failcnt",
4783                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4784                 .trigger = mem_cgroup_reset,
4785                 .read_u64 = mem_cgroup_read,
4786         },
4787         {
4788                 .name = "stat",
4789                 .read_map = mem_control_stat_show,
4790         },
4791         {
4792                 .name = "force_empty",
4793                 .trigger = mem_cgroup_force_empty_write,
4794         },
4795         {
4796                 .name = "use_hierarchy",
4797                 .write_u64 = mem_cgroup_hierarchy_write,
4798                 .read_u64 = mem_cgroup_hierarchy_read,
4799         },
4800         {
4801                 .name = "swappiness",
4802                 .read_u64 = mem_cgroup_swappiness_read,
4803                 .write_u64 = mem_cgroup_swappiness_write,
4804         },
4805         {
4806                 .name = "move_charge_at_immigrate",
4807                 .read_u64 = mem_cgroup_move_charge_read,
4808                 .write_u64 = mem_cgroup_move_charge_write,
4809         },
4810         {
4811                 .name = "oom_control",
4812                 .read_map = mem_cgroup_oom_control_read,
4813                 .write_u64 = mem_cgroup_oom_control_write,
4814                 .register_event = mem_cgroup_oom_register_event,
4815                 .unregister_event = mem_cgroup_oom_unregister_event,
4816                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4817         },
4818 #ifdef CONFIG_NUMA
4819         {
4820                 .name = "numa_stat",
4821                 .open = mem_control_numa_stat_open,
4822                 .mode = S_IRUGO,
4823         },
4824 #endif
4825 };
4826
4827 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4828 static struct cftype memsw_cgroup_files[] = {
4829         {
4830                 .name = "memsw.usage_in_bytes",
4831                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4832                 .read_u64 = mem_cgroup_read,
4833                 .register_event = mem_cgroup_usage_register_event,
4834                 .unregister_event = mem_cgroup_usage_unregister_event,
4835         },
4836         {
4837                 .name = "memsw.max_usage_in_bytes",
4838                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4839                 .trigger = mem_cgroup_reset,
4840                 .read_u64 = mem_cgroup_read,
4841         },
4842         {
4843                 .name = "memsw.limit_in_bytes",
4844                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4845                 .write_string = mem_cgroup_write,
4846                 .read_u64 = mem_cgroup_read,
4847         },
4848         {
4849                 .name = "memsw.failcnt",
4850                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4851                 .trigger = mem_cgroup_reset,
4852                 .read_u64 = mem_cgroup_read,
4853         },
4854 };
4855
4856 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4857 {
4858         if (!do_swap_account)
4859                 return 0;
4860         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4861                                 ARRAY_SIZE(memsw_cgroup_files));
4862 };
4863 #else
4864 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4865 {
4866         return 0;
4867 }
4868 #endif
4869
4870 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4871 {
4872         struct mem_cgroup_per_node *pn;
4873         struct mem_cgroup_per_zone *mz;
4874         enum lru_list l;
4875         int zone, tmp = node;
4876         /*
4877          * This routine is called against possible nodes.
4878          * But it's BUG to call kmalloc() against offline node.
4879          *
4880          * TODO: this routine can waste much memory for nodes which will
4881          *       never be onlined. It's better to use memory hotplug callback
4882          *       function.
4883          */
4884         if (!node_state(node, N_NORMAL_MEMORY))
4885                 tmp = -1;
4886         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4887         if (!pn)
4888                 return 1;
4889
4890         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4891                 mz = &pn->zoneinfo[zone];
4892                 for_each_lru(l)
4893                         INIT_LIST_HEAD(&mz->lists[l]);
4894                 mz->usage_in_excess = 0;
4895                 mz->on_tree = false;
4896                 mz->mem = memcg;
4897         }
4898         memcg->info.nodeinfo[node] = pn;
4899         return 0;
4900 }
4901
4902 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4903 {
4904         kfree(memcg->info.nodeinfo[node]);
4905 }
4906
4907 static struct mem_cgroup *mem_cgroup_alloc(void)
4908 {
4909         struct mem_cgroup *mem;
4910         int size = sizeof(struct mem_cgroup);
4911
4912         /* Can be very big if MAX_NUMNODES is very big */
4913         if (size < PAGE_SIZE)
4914                 mem = kzalloc(size, GFP_KERNEL);
4915         else
4916                 mem = vzalloc(size);
4917
4918         if (!mem)
4919                 return NULL;
4920
4921         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4922         if (!mem->stat)
4923                 goto out_free;
4924         spin_lock_init(&mem->pcp_counter_lock);
4925         return mem;
4926
4927 out_free:
4928         if (size < PAGE_SIZE)
4929                 kfree(mem);
4930         else
4931                 vfree(mem);
4932         return NULL;
4933 }
4934
4935 /*
4936  * At destroying mem_cgroup, references from swap_cgroup can remain.
4937  * (scanning all at force_empty is too costly...)
4938  *
4939  * Instead of clearing all references at force_empty, we remember
4940  * the number of reference from swap_cgroup and free mem_cgroup when
4941  * it goes down to 0.
4942  *
4943  * Removal of cgroup itself succeeds regardless of refs from swap.
4944  */
4945
4946 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4947 {
4948         int node;
4949
4950         mem_cgroup_remove_from_trees(memcg);
4951         free_css_id(&mem_cgroup_subsys, &memcg->css);
4952
4953         for_each_node_state(node, N_POSSIBLE)
4954                 free_mem_cgroup_per_zone_info(memcg, node);
4955
4956         free_percpu(memcg->stat);
4957         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4958                 kfree(memcg);
4959         else
4960                 vfree(memcg);
4961 }
4962
4963 static void mem_cgroup_get(struct mem_cgroup *memcg)
4964 {
4965         atomic_inc(&memcg->refcnt);
4966 }
4967
4968 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4969 {
4970         if (atomic_sub_and_test(count, &memcg->refcnt)) {
4971                 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4972                 __mem_cgroup_free(memcg);
4973                 if (parent)
4974                         mem_cgroup_put(parent);
4975         }
4976 }
4977
4978 static void mem_cgroup_put(struct mem_cgroup *memcg)
4979 {
4980         __mem_cgroup_put(memcg, 1);
4981 }
4982
4983 /*
4984  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4985  */
4986 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4987 {
4988         if (!memcg->res.parent)
4989                 return NULL;
4990         return mem_cgroup_from_res_counter(memcg->res.parent, res);
4991 }
4992 EXPORT_SYMBOL(parent_mem_cgroup);
4993
4994 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4995 static void __init enable_swap_cgroup(void)
4996 {
4997         if (!mem_cgroup_disabled() && really_do_swap_account)
4998                 do_swap_account = 1;
4999 }
5000 #else
5001 static void __init enable_swap_cgroup(void)
5002 {
5003 }
5004 #endif
5005
5006 static int mem_cgroup_soft_limit_tree_init(void)
5007 {
5008         struct mem_cgroup_tree_per_node *rtpn;
5009         struct mem_cgroup_tree_per_zone *rtpz;
5010         int tmp, node, zone;
5011
5012         for_each_node_state(node, N_POSSIBLE) {
5013                 tmp = node;
5014                 if (!node_state(node, N_NORMAL_MEMORY))
5015                         tmp = -1;
5016                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5017                 if (!rtpn)
5018                         return 1;
5019
5020                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5021
5022                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5023                         rtpz = &rtpn->rb_tree_per_zone[zone];
5024                         rtpz->rb_root = RB_ROOT;
5025                         spin_lock_init(&rtpz->lock);
5026                 }
5027         }
5028         return 0;
5029 }
5030
5031 static struct cgroup_subsys_state * __ref
5032 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5033 {
5034         struct mem_cgroup *memcg, *parent;
5035         long error = -ENOMEM;
5036         int node;
5037
5038         memcg = mem_cgroup_alloc();
5039         if (!memcg)
5040                 return ERR_PTR(error);
5041
5042         for_each_node_state(node, N_POSSIBLE)
5043                 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5044                         goto free_out;
5045
5046         /* root ? */
5047         if (cont->parent == NULL) {
5048                 int cpu;
5049                 enable_swap_cgroup();
5050                 parent = NULL;
5051                 if (mem_cgroup_soft_limit_tree_init())
5052                         goto free_out;
5053                 root_mem_cgroup = memcg;
5054                 for_each_possible_cpu(cpu) {
5055                         struct memcg_stock_pcp *stock =
5056                                                 &per_cpu(memcg_stock, cpu);
5057                         INIT_WORK(&stock->work, drain_local_stock);
5058                 }
5059                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5060         } else {
5061                 parent = mem_cgroup_from_cont(cont->parent);
5062                 memcg->use_hierarchy = parent->use_hierarchy;
5063                 memcg->oom_kill_disable = parent->oom_kill_disable;
5064         }
5065
5066         if (parent && parent->use_hierarchy) {
5067                 res_counter_init(&memcg->res, &parent->res);
5068                 res_counter_init(&memcg->memsw, &parent->memsw);
5069                 /*
5070                  * We increment refcnt of the parent to ensure that we can
5071                  * safely access it on res_counter_charge/uncharge.
5072                  * This refcnt will be decremented when freeing this
5073                  * mem_cgroup(see mem_cgroup_put).
5074                  */
5075                 mem_cgroup_get(parent);
5076         } else {
5077                 res_counter_init(&memcg->res, NULL);
5078                 res_counter_init(&memcg->memsw, NULL);
5079         }
5080         memcg->last_scanned_child = 0;
5081         memcg->last_scanned_node = MAX_NUMNODES;
5082         INIT_LIST_HEAD(&memcg->oom_notify);
5083
5084         if (parent)
5085                 memcg->swappiness = mem_cgroup_swappiness(parent);
5086         atomic_set(&memcg->refcnt, 1);
5087         memcg->move_charge_at_immigrate = 0;
5088         mutex_init(&memcg->thresholds_lock);
5089         return &memcg->css;
5090 free_out:
5091         __mem_cgroup_free(memcg);
5092         return ERR_PTR(error);
5093 }
5094
5095 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5096                                         struct cgroup *cont)
5097 {
5098         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5099
5100         return mem_cgroup_force_empty(memcg, false);
5101 }
5102
5103 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5104                                 struct cgroup *cont)
5105 {
5106         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5107
5108         kmem_cgroup_destroy(ss, cont);
5109
5110         mem_cgroup_put(memcg);
5111 }
5112
5113 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5114                                 struct cgroup *cont)
5115 {
5116         int ret;
5117
5118         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5119                                 ARRAY_SIZE(mem_cgroup_files));
5120
5121         if (!ret)
5122                 ret = register_memsw_files(cont, ss);
5123
5124         if (!ret)
5125                 ret = register_kmem_files(cont, ss);
5126
5127         return ret;
5128 }
5129
5130 #ifdef CONFIG_MMU
5131 /* Handlers for move charge at task migration. */
5132 #define PRECHARGE_COUNT_AT_ONCE 256
5133 static int mem_cgroup_do_precharge(unsigned long count)
5134 {
5135         int ret = 0;
5136         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5137         struct mem_cgroup *memcg = mc.to;
5138
5139         if (mem_cgroup_is_root(memcg)) {
5140                 mc.precharge += count;
5141                 /* we don't need css_get for root */
5142                 return ret;
5143         }
5144         /* try to charge at once */
5145         if (count > 1) {
5146                 struct res_counter *dummy;
5147                 /*
5148                  * "memcg" cannot be under rmdir() because we've already checked
5149                  * by cgroup_lock_live_cgroup() that it is not removed and we
5150                  * are still under the same cgroup_mutex. So we can postpone
5151                  * css_get().
5152                  */
5153                 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5154                         goto one_by_one;
5155                 if (do_swap_account && res_counter_charge(&memcg->memsw,
5156                                                 PAGE_SIZE * count, &dummy)) {
5157                         res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5158                         goto one_by_one;
5159                 }
5160                 mc.precharge += count;
5161                 return ret;
5162         }
5163 one_by_one:
5164         /* fall back to one by one charge */
5165         while (count--) {
5166                 if (signal_pending(current)) {
5167                         ret = -EINTR;
5168                         break;
5169                 }
5170                 if (!batch_count--) {
5171                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5172                         cond_resched();
5173                 }
5174                 ret = __mem_cgroup_try_charge(NULL,
5175                                         GFP_KERNEL, 1, &memcg, false);
5176                 if (ret || !memcg)
5177                         /* mem_cgroup_clear_mc() will do uncharge later */
5178                         return -ENOMEM;
5179                 mc.precharge++;
5180         }
5181         return ret;
5182 }
5183
5184 /**
5185  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5186  * @vma: the vma the pte to be checked belongs
5187  * @addr: the address corresponding to the pte to be checked
5188  * @ptent: the pte to be checked
5189  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5190  *
5191  * Returns
5192  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5193  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5194  *     move charge. if @target is not NULL, the page is stored in target->page
5195  *     with extra refcnt got(Callers should handle it).
5196  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5197  *     target for charge migration. if @target is not NULL, the entry is stored
5198  *     in target->ent.
5199  *
5200  * Called with pte lock held.
5201  */
5202 union mc_target {
5203         struct page     *page;
5204         swp_entry_t     ent;
5205 };
5206
5207 enum mc_target_type {
5208         MC_TARGET_NONE, /* not used */
5209         MC_TARGET_PAGE,
5210         MC_TARGET_SWAP,
5211 };
5212
5213 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5214                                                 unsigned long addr, pte_t ptent)
5215 {
5216         struct page *page = vm_normal_page(vma, addr, ptent);
5217
5218         if (!page || !page_mapped(page))
5219                 return NULL;
5220         if (PageAnon(page)) {
5221                 /* we don't move shared anon */
5222                 if (!move_anon() || page_mapcount(page) > 2)
5223                         return NULL;
5224         } else if (!move_file())
5225                 /* we ignore mapcount for file pages */
5226                 return NULL;
5227         if (!get_page_unless_zero(page))
5228                 return NULL;
5229
5230         return page;
5231 }
5232
5233 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5234                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5235 {
5236         int usage_count;
5237         struct page *page = NULL;
5238         swp_entry_t ent = pte_to_swp_entry(ptent);
5239
5240         if (!move_anon() || non_swap_entry(ent))
5241                 return NULL;
5242         usage_count = mem_cgroup_count_swap_user(ent, &page);
5243         if (usage_count > 1) { /* we don't move shared anon */
5244                 if (page)
5245                         put_page(page);
5246                 return NULL;
5247         }
5248         if (do_swap_account)
5249                 entry->val = ent.val;
5250
5251         return page;
5252 }
5253
5254 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5255                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5256 {
5257         struct page *page = NULL;
5258         struct inode *inode;
5259         struct address_space *mapping;
5260         pgoff_t pgoff;
5261
5262         if (!vma->vm_file) /* anonymous vma */
5263                 return NULL;
5264         if (!move_file())
5265                 return NULL;
5266
5267         inode = vma->vm_file->f_path.dentry->d_inode;
5268         mapping = vma->vm_file->f_mapping;
5269         if (pte_none(ptent))
5270                 pgoff = linear_page_index(vma, addr);
5271         else /* pte_file(ptent) is true */
5272                 pgoff = pte_to_pgoff(ptent);
5273
5274         /* page is moved even if it's not RSS of this task(page-faulted). */
5275         page = find_get_page(mapping, pgoff);
5276
5277 #ifdef CONFIG_SWAP
5278         /* shmem/tmpfs may report page out on swap: account for that too. */
5279         if (radix_tree_exceptional_entry(page)) {
5280                 swp_entry_t swap = radix_to_swp_entry(page);
5281                 if (do_swap_account)
5282                         *entry = swap;
5283                 page = find_get_page(&swapper_space, swap.val);
5284         }
5285 #endif
5286         return page;
5287 }
5288
5289 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5290                 unsigned long addr, pte_t ptent, union mc_target *target)
5291 {
5292         struct page *page = NULL;
5293         struct page_cgroup *pc;
5294         int ret = 0;
5295         swp_entry_t ent = { .val = 0 };
5296
5297         if (pte_present(ptent))
5298                 page = mc_handle_present_pte(vma, addr, ptent);
5299         else if (is_swap_pte(ptent))
5300                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5301         else if (pte_none(ptent) || pte_file(ptent))
5302                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5303
5304         if (!page && !ent.val)
5305                 return 0;
5306         if (page) {
5307                 pc = lookup_page_cgroup(page);
5308                 /*
5309                  * Do only loose check w/o page_cgroup lock.
5310                  * mem_cgroup_move_account() checks the pc is valid or not under
5311                  * the lock.
5312                  */
5313                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5314                         ret = MC_TARGET_PAGE;
5315                         if (target)
5316                                 target->page = page;
5317                 }
5318                 if (!ret || !target)
5319                         put_page(page);
5320         }
5321         /* There is a swap entry and a page doesn't exist or isn't charged */
5322         if (ent.val && !ret &&
5323                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5324                 ret = MC_TARGET_SWAP;
5325                 if (target)
5326                         target->ent = ent;
5327         }
5328         return ret;
5329 }
5330
5331 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5332                                         unsigned long addr, unsigned long end,
5333                                         struct mm_walk *walk)
5334 {
5335         struct vm_area_struct *vma = walk->private;
5336         pte_t *pte;
5337         spinlock_t *ptl;
5338
5339         split_huge_page_pmd(walk->mm, pmd);
5340
5341         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5342         for (; addr != end; pte++, addr += PAGE_SIZE)
5343                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5344                         mc.precharge++; /* increment precharge temporarily */
5345         pte_unmap_unlock(pte - 1, ptl);
5346         cond_resched();
5347
5348         return 0;
5349 }
5350
5351 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5352 {
5353         unsigned long precharge;
5354         struct vm_area_struct *vma;
5355
5356         down_read(&mm->mmap_sem);
5357         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5358                 struct mm_walk mem_cgroup_count_precharge_walk = {
5359                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5360                         .mm = mm,
5361                         .private = vma,
5362                 };
5363                 if (is_vm_hugetlb_page(vma))
5364                         continue;
5365                 walk_page_range(vma->vm_start, vma->vm_end,
5366                                         &mem_cgroup_count_precharge_walk);
5367         }
5368         up_read(&mm->mmap_sem);
5369
5370         precharge = mc.precharge;
5371         mc.precharge = 0;
5372
5373         return precharge;
5374 }
5375
5376 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5377 {
5378         unsigned long precharge = mem_cgroup_count_precharge(mm);
5379
5380         VM_BUG_ON(mc.moving_task);
5381         mc.moving_task = current;
5382         return mem_cgroup_do_precharge(precharge);
5383 }
5384
5385 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5386 static void __mem_cgroup_clear_mc(void)
5387 {
5388         struct mem_cgroup *from = mc.from;
5389         struct mem_cgroup *to = mc.to;
5390
5391         /* we must uncharge all the leftover precharges from mc.to */
5392         if (mc.precharge) {
5393                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5394                 mc.precharge = 0;
5395         }
5396         /*
5397          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5398          * we must uncharge here.
5399          */
5400         if (mc.moved_charge) {
5401                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5402                 mc.moved_charge = 0;
5403         }
5404         /* we must fixup refcnts and charges */
5405         if (mc.moved_swap) {
5406                 /* uncharge swap account from the old cgroup */
5407                 if (!mem_cgroup_is_root(mc.from))
5408                         res_counter_uncharge(&mc.from->memsw,
5409                                                 PAGE_SIZE * mc.moved_swap);
5410                 __mem_cgroup_put(mc.from, mc.moved_swap);
5411
5412                 if (!mem_cgroup_is_root(mc.to)) {
5413                         /*
5414                          * we charged both to->res and to->memsw, so we should
5415                          * uncharge to->res.
5416                          */
5417                         res_counter_uncharge(&mc.to->res,
5418                                                 PAGE_SIZE * mc.moved_swap);
5419                 }
5420                 /* we've already done mem_cgroup_get(mc.to) */
5421                 mc.moved_swap = 0;
5422         }
5423         memcg_oom_recover(from);
5424         memcg_oom_recover(to);
5425         wake_up_all(&mc.waitq);
5426 }
5427
5428 static void mem_cgroup_clear_mc(void)
5429 {
5430         struct mem_cgroup *from = mc.from;
5431
5432         /*
5433          * we must clear moving_task before waking up waiters at the end of
5434          * task migration.
5435          */
5436         mc.moving_task = NULL;
5437         __mem_cgroup_clear_mc();
5438         spin_lock(&mc.lock);
5439         mc.from = NULL;
5440         mc.to = NULL;
5441         spin_unlock(&mc.lock);
5442         mem_cgroup_end_move(from);
5443 }
5444
5445 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5446                                 struct cgroup *cgroup,
5447                                 struct cgroup_taskset *tset)
5448 {
5449         struct task_struct *p = cgroup_taskset_first(tset);
5450         int ret = 0;
5451         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5452
5453         if (memcg->move_charge_at_immigrate) {
5454                 struct mm_struct *mm;
5455                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5456
5457                 VM_BUG_ON(from == memcg);
5458
5459                 mm = get_task_mm(p);
5460                 if (!mm)
5461                         return 0;
5462                 /* We move charges only when we move a owner of the mm */
5463                 if (mm->owner == p) {
5464                         VM_BUG_ON(mc.from);
5465                         VM_BUG_ON(mc.to);
5466                         VM_BUG_ON(mc.precharge);
5467                         VM_BUG_ON(mc.moved_charge);
5468                         VM_BUG_ON(mc.moved_swap);
5469                         mem_cgroup_start_move(from);
5470                         spin_lock(&mc.lock);
5471                         mc.from = from;
5472                         mc.to = memcg;
5473                         spin_unlock(&mc.lock);
5474                         /* We set mc.moving_task later */
5475
5476                         ret = mem_cgroup_precharge_mc(mm);
5477                         if (ret)
5478                                 mem_cgroup_clear_mc();
5479                 }
5480                 mmput(mm);
5481         }
5482         return ret;
5483 }
5484
5485 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5486                                 struct cgroup *cgroup,
5487                                 struct cgroup_taskset *tset)
5488 {
5489         mem_cgroup_clear_mc();
5490 }
5491
5492 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5493                                 unsigned long addr, unsigned long end,
5494                                 struct mm_walk *walk)
5495 {
5496         int ret = 0;
5497         struct vm_area_struct *vma = walk->private;
5498         pte_t *pte;
5499         spinlock_t *ptl;
5500
5501         split_huge_page_pmd(walk->mm, pmd);
5502 retry:
5503         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5504         for (; addr != end; addr += PAGE_SIZE) {
5505                 pte_t ptent = *(pte++);
5506                 union mc_target target;
5507                 int type;
5508                 struct page *page;
5509                 struct page_cgroup *pc;
5510                 swp_entry_t ent;
5511
5512                 if (!mc.precharge)
5513                         break;
5514
5515                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5516                 switch (type) {
5517                 case MC_TARGET_PAGE:
5518                         page = target.page;
5519                         if (isolate_lru_page(page))
5520                                 goto put;
5521                         pc = lookup_page_cgroup(page);
5522                         if (!mem_cgroup_move_account(page, 1, pc,
5523                                                      mc.from, mc.to, false)) {
5524                                 mc.precharge--;
5525                                 /* we uncharge from mc.from later. */
5526                                 mc.moved_charge++;
5527                         }
5528                         putback_lru_page(page);
5529 put:                    /* is_target_pte_for_mc() gets the page */
5530                         put_page(page);
5531                         break;
5532                 case MC_TARGET_SWAP:
5533                         ent = target.ent;
5534                         if (!mem_cgroup_move_swap_account(ent,
5535                                                 mc.from, mc.to, false)) {
5536                                 mc.precharge--;
5537                                 /* we fixup refcnts and charges later. */
5538                                 mc.moved_swap++;
5539                         }
5540                         break;
5541                 default:
5542                         break;
5543                 }
5544         }
5545         pte_unmap_unlock(pte - 1, ptl);
5546         cond_resched();
5547
5548         if (addr != end) {
5549                 /*
5550                  * We have consumed all precharges we got in can_attach().
5551                  * We try charge one by one, but don't do any additional
5552                  * charges to mc.to if we have failed in charge once in attach()
5553                  * phase.
5554                  */
5555                 ret = mem_cgroup_do_precharge(1);
5556                 if (!ret)
5557                         goto retry;
5558         }
5559
5560         return ret;
5561 }
5562
5563 static void mem_cgroup_move_charge(struct mm_struct *mm)
5564 {
5565         struct vm_area_struct *vma;
5566
5567         lru_add_drain_all();
5568 retry:
5569         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5570                 /*
5571                  * Someone who are holding the mmap_sem might be waiting in
5572                  * waitq. So we cancel all extra charges, wake up all waiters,
5573                  * and retry. Because we cancel precharges, we might not be able
5574                  * to move enough charges, but moving charge is a best-effort
5575                  * feature anyway, so it wouldn't be a big problem.
5576                  */
5577                 __mem_cgroup_clear_mc();
5578                 cond_resched();
5579                 goto retry;
5580         }
5581         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5582                 int ret;
5583                 struct mm_walk mem_cgroup_move_charge_walk = {
5584                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5585                         .mm = mm,
5586                         .private = vma,
5587                 };
5588                 if (is_vm_hugetlb_page(vma))
5589                         continue;
5590                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5591                                                 &mem_cgroup_move_charge_walk);
5592                 if (ret)
5593                         /*
5594                          * means we have consumed all precharges and failed in
5595                          * doing additional charge. Just abandon here.
5596                          */
5597                         break;
5598         }
5599         up_read(&mm->mmap_sem);
5600 }
5601
5602 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5603                                 struct cgroup *cont,
5604                                 struct cgroup_taskset *tset)
5605 {
5606         struct task_struct *p = cgroup_taskset_first(tset);
5607         struct mm_struct *mm = get_task_mm(p);
5608
5609         if (mm) {
5610                 if (mc.to)
5611                         mem_cgroup_move_charge(mm);
5612                 put_swap_token(mm);
5613                 mmput(mm);
5614         }
5615         if (mc.to)
5616                 mem_cgroup_clear_mc();
5617 }
5618 #else   /* !CONFIG_MMU */
5619 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5620                                 struct cgroup *cgroup,
5621                                 struct cgroup_taskset *tset)
5622 {
5623         return 0;
5624 }
5625 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5626                                 struct cgroup *cgroup,
5627                                 struct cgroup_taskset *tset)
5628 {
5629 }
5630 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5631                                 struct cgroup *cont,
5632                                 struct cgroup_taskset *tset)
5633 {
5634 }
5635 #endif
5636
5637 struct cgroup_subsys mem_cgroup_subsys = {
5638         .name = "memory",
5639         .subsys_id = mem_cgroup_subsys_id,
5640         .create = mem_cgroup_create,
5641         .pre_destroy = mem_cgroup_pre_destroy,
5642         .destroy = mem_cgroup_destroy,
5643         .populate = mem_cgroup_populate,
5644         .can_attach = mem_cgroup_can_attach,
5645         .cancel_attach = mem_cgroup_cancel_attach,
5646         .attach = mem_cgroup_move_task,
5647         .early_init = 0,
5648         .use_id = 1,
5649 };
5650
5651 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5652 static int __init enable_swap_account(char *s)
5653 {
5654         /* consider enabled if no parameter or 1 is given */
5655         if (!strcmp(s, "1"))
5656                 really_do_swap_account = 1;
5657         else if (!strcmp(s, "0"))
5658                 really_do_swap_account = 0;
5659         return 1;
5660 }
5661 __setup("swapaccount=", enable_swap_account);
5662
5663 #endif