Merge branch 'for-3.3' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup
[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         /* A socket spends its whole life in the same cgroup */
385         if (sk->sk_cgrp) {
386                 WARN_ON(1);
387                 return;
388         }
389         if (static_branch(&memcg_socket_limit_enabled)) {
390                 struct mem_cgroup *memcg;
391
392                 BUG_ON(!sk->sk_prot->proto_cgroup);
393
394                 rcu_read_lock();
395                 memcg = mem_cgroup_from_task(current);
396                 if (!mem_cgroup_is_root(memcg)) {
397                         mem_cgroup_get(memcg);
398                         sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
399                 }
400                 rcu_read_unlock();
401         }
402 }
403 EXPORT_SYMBOL(sock_update_memcg);
404
405 void sock_release_memcg(struct sock *sk)
406 {
407         if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
408                 struct mem_cgroup *memcg;
409                 WARN_ON(!sk->sk_cgrp->memcg);
410                 memcg = sk->sk_cgrp->memcg;
411                 mem_cgroup_put(memcg);
412         }
413 }
414
415 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
416 {
417         if (!memcg || mem_cgroup_is_root(memcg))
418                 return NULL;
419
420         return &memcg->tcp_mem.cg_proto;
421 }
422 EXPORT_SYMBOL(tcp_proto_cgroup);
423 #endif /* CONFIG_INET */
424 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
425
426 static void drain_all_stock_async(struct mem_cgroup *memcg);
427
428 static struct mem_cgroup_per_zone *
429 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
430 {
431         return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
432 }
433
434 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
435 {
436         return &memcg->css;
437 }
438
439 static struct mem_cgroup_per_zone *
440 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
441 {
442         int nid = page_to_nid(page);
443         int zid = page_zonenum(page);
444
445         return mem_cgroup_zoneinfo(memcg, nid, zid);
446 }
447
448 static struct mem_cgroup_tree_per_zone *
449 soft_limit_tree_node_zone(int nid, int zid)
450 {
451         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
452 }
453
454 static struct mem_cgroup_tree_per_zone *
455 soft_limit_tree_from_page(struct page *page)
456 {
457         int nid = page_to_nid(page);
458         int zid = page_zonenum(page);
459
460         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
461 }
462
463 static void
464 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
465                                 struct mem_cgroup_per_zone *mz,
466                                 struct mem_cgroup_tree_per_zone *mctz,
467                                 unsigned long long new_usage_in_excess)
468 {
469         struct rb_node **p = &mctz->rb_root.rb_node;
470         struct rb_node *parent = NULL;
471         struct mem_cgroup_per_zone *mz_node;
472
473         if (mz->on_tree)
474                 return;
475
476         mz->usage_in_excess = new_usage_in_excess;
477         if (!mz->usage_in_excess)
478                 return;
479         while (*p) {
480                 parent = *p;
481                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
482                                         tree_node);
483                 if (mz->usage_in_excess < mz_node->usage_in_excess)
484                         p = &(*p)->rb_left;
485                 /*
486                  * We can't avoid mem cgroups that are over their soft
487                  * limit by the same amount
488                  */
489                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
490                         p = &(*p)->rb_right;
491         }
492         rb_link_node(&mz->tree_node, parent, p);
493         rb_insert_color(&mz->tree_node, &mctz->rb_root);
494         mz->on_tree = true;
495 }
496
497 static void
498 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
499                                 struct mem_cgroup_per_zone *mz,
500                                 struct mem_cgroup_tree_per_zone *mctz)
501 {
502         if (!mz->on_tree)
503                 return;
504         rb_erase(&mz->tree_node, &mctz->rb_root);
505         mz->on_tree = false;
506 }
507
508 static void
509 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
510                                 struct mem_cgroup_per_zone *mz,
511                                 struct mem_cgroup_tree_per_zone *mctz)
512 {
513         spin_lock(&mctz->lock);
514         __mem_cgroup_remove_exceeded(memcg, mz, mctz);
515         spin_unlock(&mctz->lock);
516 }
517
518
519 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
520 {
521         unsigned long long excess;
522         struct mem_cgroup_per_zone *mz;
523         struct mem_cgroup_tree_per_zone *mctz;
524         int nid = page_to_nid(page);
525         int zid = page_zonenum(page);
526         mctz = soft_limit_tree_from_page(page);
527
528         /*
529          * Necessary to update all ancestors when hierarchy is used.
530          * because their event counter is not touched.
531          */
532         for (; memcg; memcg = parent_mem_cgroup(memcg)) {
533                 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
534                 excess = res_counter_soft_limit_excess(&memcg->res);
535                 /*
536                  * We have to update the tree if mz is on RB-tree or
537                  * mem is over its softlimit.
538                  */
539                 if (excess || mz->on_tree) {
540                         spin_lock(&mctz->lock);
541                         /* if on-tree, remove it */
542                         if (mz->on_tree)
543                                 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
544                         /*
545                          * Insert again. mz->usage_in_excess will be updated.
546                          * If excess is 0, no tree ops.
547                          */
548                         __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
549                         spin_unlock(&mctz->lock);
550                 }
551         }
552 }
553
554 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
555 {
556         int node, zone;
557         struct mem_cgroup_per_zone *mz;
558         struct mem_cgroup_tree_per_zone *mctz;
559
560         for_each_node_state(node, N_POSSIBLE) {
561                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
562                         mz = mem_cgroup_zoneinfo(memcg, node, zone);
563                         mctz = soft_limit_tree_node_zone(node, zone);
564                         mem_cgroup_remove_exceeded(memcg, mz, mctz);
565                 }
566         }
567 }
568
569 static struct mem_cgroup_per_zone *
570 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
571 {
572         struct rb_node *rightmost = NULL;
573         struct mem_cgroup_per_zone *mz;
574
575 retry:
576         mz = NULL;
577         rightmost = rb_last(&mctz->rb_root);
578         if (!rightmost)
579                 goto done;              /* Nothing to reclaim from */
580
581         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
582         /*
583          * Remove the node now but someone else can add it back,
584          * we will to add it back at the end of reclaim to its correct
585          * position in the tree.
586          */
587         __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
588         if (!res_counter_soft_limit_excess(&mz->mem->res) ||
589                 !css_tryget(&mz->mem->css))
590                 goto retry;
591 done:
592         return mz;
593 }
594
595 static struct mem_cgroup_per_zone *
596 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
597 {
598         struct mem_cgroup_per_zone *mz;
599
600         spin_lock(&mctz->lock);
601         mz = __mem_cgroup_largest_soft_limit_node(mctz);
602         spin_unlock(&mctz->lock);
603         return mz;
604 }
605
606 /*
607  * Implementation Note: reading percpu statistics for memcg.
608  *
609  * Both of vmstat[] and percpu_counter has threshold and do periodic
610  * synchronization to implement "quick" read. There are trade-off between
611  * reading cost and precision of value. Then, we may have a chance to implement
612  * a periodic synchronizion of counter in memcg's counter.
613  *
614  * But this _read() function is used for user interface now. The user accounts
615  * memory usage by memory cgroup and he _always_ requires exact value because
616  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
617  * have to visit all online cpus and make sum. So, for now, unnecessary
618  * synchronization is not implemented. (just implemented for cpu hotplug)
619  *
620  * If there are kernel internal actions which can make use of some not-exact
621  * value, and reading all cpu value can be performance bottleneck in some
622  * common workload, threashold and synchonization as vmstat[] should be
623  * implemented.
624  */
625 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
626                                  enum mem_cgroup_stat_index idx)
627 {
628         long val = 0;
629         int cpu;
630
631         get_online_cpus();
632         for_each_online_cpu(cpu)
633                 val += per_cpu(memcg->stat->count[idx], cpu);
634 #ifdef CONFIG_HOTPLUG_CPU
635         spin_lock(&memcg->pcp_counter_lock);
636         val += memcg->nocpu_base.count[idx];
637         spin_unlock(&memcg->pcp_counter_lock);
638 #endif
639         put_online_cpus();
640         return val;
641 }
642
643 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
644                                          bool charge)
645 {
646         int val = (charge) ? 1 : -1;
647         this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
648 }
649
650 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
651 {
652         this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
653 }
654
655 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
656 {
657         this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
658 }
659
660 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
661                                             enum mem_cgroup_events_index idx)
662 {
663         unsigned long val = 0;
664         int cpu;
665
666         for_each_online_cpu(cpu)
667                 val += per_cpu(memcg->stat->events[idx], cpu);
668 #ifdef CONFIG_HOTPLUG_CPU
669         spin_lock(&memcg->pcp_counter_lock);
670         val += memcg->nocpu_base.events[idx];
671         spin_unlock(&memcg->pcp_counter_lock);
672 #endif
673         return val;
674 }
675
676 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
677                                          bool file, int nr_pages)
678 {
679         preempt_disable();
680
681         if (file)
682                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
683                                 nr_pages);
684         else
685                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
686                                 nr_pages);
687
688         /* pagein of a big page is an event. So, ignore page size */
689         if (nr_pages > 0)
690                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
691         else {
692                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
693                 nr_pages = -nr_pages; /* for event */
694         }
695
696         __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
697
698         preempt_enable();
699 }
700
701 unsigned long
702 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
703                         unsigned int lru_mask)
704 {
705         struct mem_cgroup_per_zone *mz;
706         enum lru_list l;
707         unsigned long ret = 0;
708
709         mz = mem_cgroup_zoneinfo(memcg, nid, zid);
710
711         for_each_lru(l) {
712                 if (BIT(l) & lru_mask)
713                         ret += MEM_CGROUP_ZSTAT(mz, l);
714         }
715         return ret;
716 }
717
718 static unsigned long
719 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
720                         int nid, unsigned int lru_mask)
721 {
722         u64 total = 0;
723         int zid;
724
725         for (zid = 0; zid < MAX_NR_ZONES; zid++)
726                 total += mem_cgroup_zone_nr_lru_pages(memcg,
727                                                 nid, zid, lru_mask);
728
729         return total;
730 }
731
732 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
733                         unsigned int lru_mask)
734 {
735         int nid;
736         u64 total = 0;
737
738         for_each_node_state(nid, N_HIGH_MEMORY)
739                 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
740         return total;
741 }
742
743 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
744 {
745         unsigned long val, next;
746
747         val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
748         next = __this_cpu_read(memcg->stat->targets[target]);
749         /* from time_after() in jiffies.h */
750         return ((long)next - (long)val < 0);
751 }
752
753 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
754 {
755         unsigned long val, next;
756
757         val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
758
759         switch (target) {
760         case MEM_CGROUP_TARGET_THRESH:
761                 next = val + THRESHOLDS_EVENTS_TARGET;
762                 break;
763         case MEM_CGROUP_TARGET_SOFTLIMIT:
764                 next = val + SOFTLIMIT_EVENTS_TARGET;
765                 break;
766         case MEM_CGROUP_TARGET_NUMAINFO:
767                 next = val + NUMAINFO_EVENTS_TARGET;
768                 break;
769         default:
770                 return;
771         }
772
773         __this_cpu_write(memcg->stat->targets[target], next);
774 }
775
776 /*
777  * Check events in order.
778  *
779  */
780 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
781 {
782         preempt_disable();
783         /* threshold event is triggered in finer grain than soft limit */
784         if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
785                 mem_cgroup_threshold(memcg);
786                 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
787                 if (unlikely(__memcg_event_check(memcg,
788                              MEM_CGROUP_TARGET_SOFTLIMIT))) {
789                         mem_cgroup_update_tree(memcg, page);
790                         __mem_cgroup_target_update(memcg,
791                                                    MEM_CGROUP_TARGET_SOFTLIMIT);
792                 }
793 #if MAX_NUMNODES > 1
794                 if (unlikely(__memcg_event_check(memcg,
795                         MEM_CGROUP_TARGET_NUMAINFO))) {
796                         atomic_inc(&memcg->numainfo_events);
797                         __mem_cgroup_target_update(memcg,
798                                 MEM_CGROUP_TARGET_NUMAINFO);
799                 }
800 #endif
801         }
802         preempt_enable();
803 }
804
805 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
806 {
807         return container_of(cgroup_subsys_state(cont,
808                                 mem_cgroup_subsys_id), struct mem_cgroup,
809                                 css);
810 }
811
812 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
813 {
814         /*
815          * mm_update_next_owner() may clear mm->owner to NULL
816          * if it races with swapoff, page migration, etc.
817          * So this can be called with p == NULL.
818          */
819         if (unlikely(!p))
820                 return NULL;
821
822         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
823                                 struct mem_cgroup, css);
824 }
825
826 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
827 {
828         struct mem_cgroup *memcg = NULL;
829
830         if (!mm)
831                 return NULL;
832         /*
833          * Because we have no locks, mm->owner's may be being moved to other
834          * cgroup. We use css_tryget() here even if this looks
835          * pessimistic (rather than adding locks here).
836          */
837         rcu_read_lock();
838         do {
839                 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
840                 if (unlikely(!memcg))
841                         break;
842         } while (!css_tryget(&memcg->css));
843         rcu_read_unlock();
844         return memcg;
845 }
846
847 /* The caller has to guarantee "mem" exists before calling this */
848 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
849 {
850         struct cgroup_subsys_state *css;
851         int found;
852
853         if (!memcg) /* ROOT cgroup has the smallest ID */
854                 return root_mem_cgroup; /*css_put/get against root is ignored*/
855         if (!memcg->use_hierarchy) {
856                 if (css_tryget(&memcg->css))
857                         return memcg;
858                 return NULL;
859         }
860         rcu_read_lock();
861         /*
862          * searching a memory cgroup which has the smallest ID under given
863          * ROOT cgroup. (ID >= 1)
864          */
865         css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
866         if (css && css_tryget(css))
867                 memcg = container_of(css, struct mem_cgroup, css);
868         else
869                 memcg = NULL;
870         rcu_read_unlock();
871         return memcg;
872 }
873
874 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
875                                         struct mem_cgroup *root,
876                                         bool cond)
877 {
878         int nextid = css_id(&iter->css) + 1;
879         int found;
880         int hierarchy_used;
881         struct cgroup_subsys_state *css;
882
883         hierarchy_used = iter->use_hierarchy;
884
885         css_put(&iter->css);
886         /* If no ROOT, walk all, ignore hierarchy */
887         if (!cond || (root && !hierarchy_used))
888                 return NULL;
889
890         if (!root)
891                 root = root_mem_cgroup;
892
893         do {
894                 iter = NULL;
895                 rcu_read_lock();
896
897                 css = css_get_next(&mem_cgroup_subsys, nextid,
898                                 &root->css, &found);
899                 if (css && css_tryget(css))
900                         iter = container_of(css, struct mem_cgroup, css);
901                 rcu_read_unlock();
902                 /* If css is NULL, no more cgroups will be found */
903                 nextid = found + 1;
904         } while (css && !iter);
905
906         return iter;
907 }
908 /*
909  * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
910  * be careful that "break" loop is not allowed. We have reference count.
911  * Instead of that modify "cond" to be false and "continue" to exit the loop.
912  */
913 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
914         for (iter = mem_cgroup_start_loop(root);\
915              iter != NULL;\
916              iter = mem_cgroup_get_next(iter, root, cond))
917
918 #define for_each_mem_cgroup_tree(iter, root) \
919         for_each_mem_cgroup_tree_cond(iter, root, true)
920
921 #define for_each_mem_cgroup_all(iter) \
922         for_each_mem_cgroup_tree_cond(iter, NULL, true)
923
924
925 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
926 {
927         return (memcg == root_mem_cgroup);
928 }
929
930 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
931 {
932         struct mem_cgroup *memcg;
933
934         if (!mm)
935                 return;
936
937         rcu_read_lock();
938         memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
939         if (unlikely(!memcg))
940                 goto out;
941
942         switch (idx) {
943         case PGMAJFAULT:
944                 mem_cgroup_pgmajfault(memcg, 1);
945                 break;
946         case PGFAULT:
947                 mem_cgroup_pgfault(memcg, 1);
948                 break;
949         default:
950                 BUG();
951         }
952 out:
953         rcu_read_unlock();
954 }
955 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
956
957 /*
958  * Following LRU functions are allowed to be used without PCG_LOCK.
959  * Operations are called by routine of global LRU independently from memcg.
960  * What we have to take care of here is validness of pc->mem_cgroup.
961  *
962  * Changes to pc->mem_cgroup happens when
963  * 1. charge
964  * 2. moving account
965  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
966  * It is added to LRU before charge.
967  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
968  * When moving account, the page is not on LRU. It's isolated.
969  */
970
971 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
972 {
973         struct page_cgroup *pc;
974         struct mem_cgroup_per_zone *mz;
975
976         if (mem_cgroup_disabled())
977                 return;
978         pc = lookup_page_cgroup(page);
979         /* can happen while we handle swapcache. */
980         if (!TestClearPageCgroupAcctLRU(pc))
981                 return;
982         VM_BUG_ON(!pc->mem_cgroup);
983         /*
984          * We don't check PCG_USED bit. It's cleared when the "page" is finally
985          * removed from global LRU.
986          */
987         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
988         /* huge page split is done under lru_lock. so, we have no races. */
989         MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
990         if (mem_cgroup_is_root(pc->mem_cgroup))
991                 return;
992         VM_BUG_ON(list_empty(&pc->lru));
993         list_del_init(&pc->lru);
994 }
995
996 void mem_cgroup_del_lru(struct page *page)
997 {
998         mem_cgroup_del_lru_list(page, page_lru(page));
999 }
1000
1001 /*
1002  * Writeback is about to end against a page which has been marked for immediate
1003  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
1004  * inactive list.
1005  */
1006 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1007 {
1008         struct mem_cgroup_per_zone *mz;
1009         struct page_cgroup *pc;
1010         enum lru_list lru = page_lru(page);
1011
1012         if (mem_cgroup_disabled())
1013                 return;
1014
1015         pc = lookup_page_cgroup(page);
1016         /* unused or root page is not rotated. */
1017         if (!PageCgroupUsed(pc))
1018                 return;
1019         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1020         smp_rmb();
1021         if (mem_cgroup_is_root(pc->mem_cgroup))
1022                 return;
1023         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1024         list_move_tail(&pc->lru, &mz->lists[lru]);
1025 }
1026
1027 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1028 {
1029         struct mem_cgroup_per_zone *mz;
1030         struct page_cgroup *pc;
1031
1032         if (mem_cgroup_disabled())
1033                 return;
1034
1035         pc = lookup_page_cgroup(page);
1036         /* unused or root page is not rotated. */
1037         if (!PageCgroupUsed(pc))
1038                 return;
1039         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1040         smp_rmb();
1041         if (mem_cgroup_is_root(pc->mem_cgroup))
1042                 return;
1043         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1044         list_move(&pc->lru, &mz->lists[lru]);
1045 }
1046
1047 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1048 {
1049         struct page_cgroup *pc;
1050         struct mem_cgroup_per_zone *mz;
1051
1052         if (mem_cgroup_disabled())
1053                 return;
1054         pc = lookup_page_cgroup(page);
1055         VM_BUG_ON(PageCgroupAcctLRU(pc));
1056         /*
1057          * putback:                             charge:
1058          * SetPageLRU                           SetPageCgroupUsed
1059          * smp_mb                               smp_mb
1060          * PageCgroupUsed && add to memcg LRU   PageLRU && add to memcg LRU
1061          *
1062          * Ensure that one of the two sides adds the page to the memcg
1063          * LRU during a race.
1064          */
1065         smp_mb();
1066         if (!PageCgroupUsed(pc))
1067                 return;
1068         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1069         smp_rmb();
1070         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1071         /* huge page split is done under lru_lock. so, we have no races. */
1072         MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1073         SetPageCgroupAcctLRU(pc);
1074         if (mem_cgroup_is_root(pc->mem_cgroup))
1075                 return;
1076         list_add(&pc->lru, &mz->lists[lru]);
1077 }
1078
1079 /*
1080  * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1081  * while it's linked to lru because the page may be reused after it's fully
1082  * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1083  * It's done under lock_page and expected that zone->lru_lock isnever held.
1084  */
1085 static void mem_cgroup_lru_del_before_commit(struct page *page)
1086 {
1087         unsigned long flags;
1088         struct zone *zone = page_zone(page);
1089         struct page_cgroup *pc = lookup_page_cgroup(page);
1090
1091         /*
1092          * Doing this check without taking ->lru_lock seems wrong but this
1093          * is safe. Because if page_cgroup's USED bit is unset, the page
1094          * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1095          * set, the commit after this will fail, anyway.
1096          * This all charge/uncharge is done under some mutual execustion.
1097          * So, we don't need to taking care of changes in USED bit.
1098          */
1099         if (likely(!PageLRU(page)))
1100                 return;
1101
1102         spin_lock_irqsave(&zone->lru_lock, flags);
1103         /*
1104          * Forget old LRU when this page_cgroup is *not* used. This Used bit
1105          * is guarded by lock_page() because the page is SwapCache.
1106          */
1107         if (!PageCgroupUsed(pc))
1108                 mem_cgroup_del_lru_list(page, page_lru(page));
1109         spin_unlock_irqrestore(&zone->lru_lock, flags);
1110 }
1111
1112 static void mem_cgroup_lru_add_after_commit(struct page *page)
1113 {
1114         unsigned long flags;
1115         struct zone *zone = page_zone(page);
1116         struct page_cgroup *pc = lookup_page_cgroup(page);
1117         /*
1118          * putback:                             charge:
1119          * SetPageLRU                           SetPageCgroupUsed
1120          * smp_mb                               smp_mb
1121          * PageCgroupUsed && add to memcg LRU   PageLRU && add to memcg LRU
1122          *
1123          * Ensure that one of the two sides adds the page to the memcg
1124          * LRU during a race.
1125          */
1126         smp_mb();
1127         /* taking care of that the page is added to LRU while we commit it */
1128         if (likely(!PageLRU(page)))
1129                 return;
1130         spin_lock_irqsave(&zone->lru_lock, flags);
1131         /* link when the page is linked to LRU but page_cgroup isn't */
1132         if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1133                 mem_cgroup_add_lru_list(page, page_lru(page));
1134         spin_unlock_irqrestore(&zone->lru_lock, flags);
1135 }
1136
1137
1138 void mem_cgroup_move_lists(struct page *page,
1139                            enum lru_list from, enum lru_list to)
1140 {
1141         if (mem_cgroup_disabled())
1142                 return;
1143         mem_cgroup_del_lru_list(page, from);
1144         mem_cgroup_add_lru_list(page, to);
1145 }
1146
1147 /*
1148  * Checks whether given mem is same or in the root_mem_cgroup's
1149  * hierarchy subtree
1150  */
1151 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1152                 struct mem_cgroup *memcg)
1153 {
1154         if (root_memcg != memcg) {
1155                 return (root_memcg->use_hierarchy &&
1156                         css_is_ancestor(&memcg->css, &root_memcg->css));
1157         }
1158
1159         return true;
1160 }
1161
1162 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1163 {
1164         int ret;
1165         struct mem_cgroup *curr = NULL;
1166         struct task_struct *p;
1167
1168         p = find_lock_task_mm(task);
1169         if (!p)
1170                 return 0;
1171         curr = try_get_mem_cgroup_from_mm(p->mm);
1172         task_unlock(p);
1173         if (!curr)
1174                 return 0;
1175         /*
1176          * We should check use_hierarchy of "memcg" not "curr". Because checking
1177          * use_hierarchy of "curr" here make this function true if hierarchy is
1178          * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1179          * hierarchy(even if use_hierarchy is disabled in "memcg").
1180          */
1181         ret = mem_cgroup_same_or_subtree(memcg, curr);
1182         css_put(&curr->css);
1183         return ret;
1184 }
1185
1186 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1187 {
1188         unsigned long inactive_ratio;
1189         int nid = zone_to_nid(zone);
1190         int zid = zone_idx(zone);
1191         unsigned long inactive;
1192         unsigned long active;
1193         unsigned long gb;
1194
1195         inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1196                                                 BIT(LRU_INACTIVE_ANON));
1197         active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1198                                               BIT(LRU_ACTIVE_ANON));
1199
1200         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1201         if (gb)
1202                 inactive_ratio = int_sqrt(10 * gb);
1203         else
1204                 inactive_ratio = 1;
1205
1206         return inactive * inactive_ratio < active;
1207 }
1208
1209 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1210 {
1211         unsigned long active;
1212         unsigned long inactive;
1213         int zid = zone_idx(zone);
1214         int nid = zone_to_nid(zone);
1215
1216         inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1217                                                 BIT(LRU_INACTIVE_FILE));
1218         active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1219                                               BIT(LRU_ACTIVE_FILE));
1220
1221         return (active > inactive);
1222 }
1223
1224 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1225                                                       struct zone *zone)
1226 {
1227         int nid = zone_to_nid(zone);
1228         int zid = zone_idx(zone);
1229         struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1230
1231         return &mz->reclaim_stat;
1232 }
1233
1234 struct zone_reclaim_stat *
1235 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1236 {
1237         struct page_cgroup *pc;
1238         struct mem_cgroup_per_zone *mz;
1239
1240         if (mem_cgroup_disabled())
1241                 return NULL;
1242
1243         pc = lookup_page_cgroup(page);
1244         if (!PageCgroupUsed(pc))
1245                 return NULL;
1246         /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1247         smp_rmb();
1248         mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1249         return &mz->reclaim_stat;
1250 }
1251
1252 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1253                                         struct list_head *dst,
1254                                         unsigned long *scanned, int order,
1255                                         isolate_mode_t mode,
1256                                         struct zone *z,
1257                                         struct mem_cgroup *mem_cont,
1258                                         int active, int file)
1259 {
1260         unsigned long nr_taken = 0;
1261         struct page *page;
1262         unsigned long scan;
1263         LIST_HEAD(pc_list);
1264         struct list_head *src;
1265         struct page_cgroup *pc, *tmp;
1266         int nid = zone_to_nid(z);
1267         int zid = zone_idx(z);
1268         struct mem_cgroup_per_zone *mz;
1269         int lru = LRU_FILE * file + active;
1270         int ret;
1271
1272         BUG_ON(!mem_cont);
1273         mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1274         src = &mz->lists[lru];
1275
1276         scan = 0;
1277         list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1278                 if (scan >= nr_to_scan)
1279                         break;
1280
1281                 if (unlikely(!PageCgroupUsed(pc)))
1282                         continue;
1283
1284                 page = lookup_cgroup_page(pc);
1285
1286                 if (unlikely(!PageLRU(page)))
1287                         continue;
1288
1289                 scan++;
1290                 ret = __isolate_lru_page(page, mode, file);
1291                 switch (ret) {
1292                 case 0:
1293                         list_move(&page->lru, dst);
1294                         mem_cgroup_del_lru(page);
1295                         nr_taken += hpage_nr_pages(page);
1296                         break;
1297                 case -EBUSY:
1298                         /* we don't affect global LRU but rotate in our LRU */
1299                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1300                         break;
1301                 default:
1302                         break;
1303                 }
1304         }
1305
1306         *scanned = scan;
1307
1308         trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1309                                       0, 0, 0, mode);
1310
1311         return nr_taken;
1312 }
1313
1314 #define mem_cgroup_from_res_counter(counter, member)    \
1315         container_of(counter, struct mem_cgroup, member)
1316
1317 /**
1318  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1319  * @mem: the memory cgroup
1320  *
1321  * Returns the maximum amount of memory @mem can be charged with, in
1322  * pages.
1323  */
1324 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1325 {
1326         unsigned long long margin;
1327
1328         margin = res_counter_margin(&memcg->res);
1329         if (do_swap_account)
1330                 margin = min(margin, res_counter_margin(&memcg->memsw));
1331         return margin >> PAGE_SHIFT;
1332 }
1333
1334 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1335 {
1336         struct cgroup *cgrp = memcg->css.cgroup;
1337
1338         /* root ? */
1339         if (cgrp->parent == NULL)
1340                 return vm_swappiness;
1341
1342         return memcg->swappiness;
1343 }
1344
1345 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1346 {
1347         int cpu;
1348
1349         get_online_cpus();
1350         spin_lock(&memcg->pcp_counter_lock);
1351         for_each_online_cpu(cpu)
1352                 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1353         memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1354         spin_unlock(&memcg->pcp_counter_lock);
1355         put_online_cpus();
1356
1357         synchronize_rcu();
1358 }
1359
1360 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1361 {
1362         int cpu;
1363
1364         if (!memcg)
1365                 return;
1366         get_online_cpus();
1367         spin_lock(&memcg->pcp_counter_lock);
1368         for_each_online_cpu(cpu)
1369                 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1370         memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1371         spin_unlock(&memcg->pcp_counter_lock);
1372         put_online_cpus();
1373 }
1374 /*
1375  * 2 routines for checking "mem" is under move_account() or not.
1376  *
1377  * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1378  *                        for avoiding race in accounting. If true,
1379  *                        pc->mem_cgroup may be overwritten.
1380  *
1381  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1382  *                        under hierarchy of moving cgroups. This is for
1383  *                        waiting at hith-memory prressure caused by "move".
1384  */
1385
1386 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1387 {
1388         VM_BUG_ON(!rcu_read_lock_held());
1389         return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1390 }
1391
1392 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1393 {
1394         struct mem_cgroup *from;
1395         struct mem_cgroup *to;
1396         bool ret = false;
1397         /*
1398          * Unlike task_move routines, we access mc.to, mc.from not under
1399          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1400          */
1401         spin_lock(&mc.lock);
1402         from = mc.from;
1403         to = mc.to;
1404         if (!from)
1405                 goto unlock;
1406
1407         ret = mem_cgroup_same_or_subtree(memcg, from)
1408                 || mem_cgroup_same_or_subtree(memcg, to);
1409 unlock:
1410         spin_unlock(&mc.lock);
1411         return ret;
1412 }
1413
1414 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1415 {
1416         if (mc.moving_task && current != mc.moving_task) {
1417                 if (mem_cgroup_under_move(memcg)) {
1418                         DEFINE_WAIT(wait);
1419                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1420                         /* moving charge context might have finished. */
1421                         if (mc.moving_task)
1422                                 schedule();
1423                         finish_wait(&mc.waitq, &wait);
1424                         return true;
1425                 }
1426         }
1427         return false;
1428 }
1429
1430 /**
1431  * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1432  * @memcg: The memory cgroup that went over limit
1433  * @p: Task that is going to be killed
1434  *
1435  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1436  * enabled
1437  */
1438 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1439 {
1440         struct cgroup *task_cgrp;
1441         struct cgroup *mem_cgrp;
1442         /*
1443          * Need a buffer in BSS, can't rely on allocations. The code relies
1444          * on the assumption that OOM is serialized for memory controller.
1445          * If this assumption is broken, revisit this code.
1446          */
1447         static char memcg_name[PATH_MAX];
1448         int ret;
1449
1450         if (!memcg || !p)
1451                 return;
1452
1453
1454         rcu_read_lock();
1455
1456         mem_cgrp = memcg->css.cgroup;
1457         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1458
1459         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1460         if (ret < 0) {
1461                 /*
1462                  * Unfortunately, we are unable to convert to a useful name
1463                  * But we'll still print out the usage information
1464                  */
1465                 rcu_read_unlock();
1466                 goto done;
1467         }
1468         rcu_read_unlock();
1469
1470         printk(KERN_INFO "Task in %s killed", memcg_name);
1471
1472         rcu_read_lock();
1473         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1474         if (ret < 0) {
1475                 rcu_read_unlock();
1476                 goto done;
1477         }
1478         rcu_read_unlock();
1479
1480         /*
1481          * Continues from above, so we don't need an KERN_ level
1482          */
1483         printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1484 done:
1485
1486         printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1487                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1488                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1489                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1490         printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1491                 "failcnt %llu\n",
1492                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1493                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1494                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1495 }
1496
1497 /*
1498  * This function returns the number of memcg under hierarchy tree. Returns
1499  * 1(self count) if no children.
1500  */
1501 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1502 {
1503         int num = 0;
1504         struct mem_cgroup *iter;
1505
1506         for_each_mem_cgroup_tree(iter, memcg)
1507                 num++;
1508         return num;
1509 }
1510
1511 /*
1512  * Return the memory (and swap, if configured) limit for a memcg.
1513  */
1514 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1515 {
1516         u64 limit;
1517         u64 memsw;
1518
1519         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1520         limit += total_swap_pages << PAGE_SHIFT;
1521
1522         memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1523         /*
1524          * If memsw is finite and limits the amount of swap space available
1525          * to this memcg, return that limit.
1526          */
1527         return min(limit, memsw);
1528 }
1529
1530 /*
1531  * Visit the first child (need not be the first child as per the ordering
1532  * of the cgroup list, since we track last_scanned_child) of @mem and use
1533  * that to reclaim free pages from.
1534  */
1535 static struct mem_cgroup *
1536 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1537 {
1538         struct mem_cgroup *ret = NULL;
1539         struct cgroup_subsys_state *css;
1540         int nextid, found;
1541
1542         if (!root_memcg->use_hierarchy) {
1543                 css_get(&root_memcg->css);
1544                 ret = root_memcg;
1545         }
1546
1547         while (!ret) {
1548                 rcu_read_lock();
1549                 nextid = root_memcg->last_scanned_child + 1;
1550                 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1551                                    &found);
1552                 if (css && css_tryget(css))
1553                         ret = container_of(css, struct mem_cgroup, css);
1554
1555                 rcu_read_unlock();
1556                 /* Updates scanning parameter */
1557                 if (!css) {
1558                         /* this means start scan from ID:1 */
1559                         root_memcg->last_scanned_child = 0;
1560                 } else
1561                         root_memcg->last_scanned_child = found;
1562         }
1563
1564         return ret;
1565 }
1566
1567 /**
1568  * test_mem_cgroup_node_reclaimable
1569  * @mem: the target memcg
1570  * @nid: the node ID to be checked.
1571  * @noswap : specify true here if the user wants flle only information.
1572  *
1573  * This function returns whether the specified memcg contains any
1574  * reclaimable pages on a node. Returns true if there are any reclaimable
1575  * pages in the node.
1576  */
1577 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1578                 int nid, bool noswap)
1579 {
1580         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1581                 return true;
1582         if (noswap || !total_swap_pages)
1583                 return false;
1584         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1585                 return true;
1586         return false;
1587
1588 }
1589 #if MAX_NUMNODES > 1
1590
1591 /*
1592  * Always updating the nodemask is not very good - even if we have an empty
1593  * list or the wrong list here, we can start from some node and traverse all
1594  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1595  *
1596  */
1597 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1598 {
1599         int nid;
1600         /*
1601          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1602          * pagein/pageout changes since the last update.
1603          */
1604         if (!atomic_read(&memcg->numainfo_events))
1605                 return;
1606         if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1607                 return;
1608
1609         /* make a nodemask where this memcg uses memory from */
1610         memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1611
1612         for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1613
1614                 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1615                         node_clear(nid, memcg->scan_nodes);
1616         }
1617
1618         atomic_set(&memcg->numainfo_events, 0);
1619         atomic_set(&memcg->numainfo_updating, 0);
1620 }
1621
1622 /*
1623  * Selecting a node where we start reclaim from. Because what we need is just
1624  * reducing usage counter, start from anywhere is O,K. Considering
1625  * memory reclaim from current node, there are pros. and cons.
1626  *
1627  * Freeing memory from current node means freeing memory from a node which
1628  * we'll use or we've used. So, it may make LRU bad. And if several threads
1629  * hit limits, it will see a contention on a node. But freeing from remote
1630  * node means more costs for memory reclaim because of memory latency.
1631  *
1632  * Now, we use round-robin. Better algorithm is welcomed.
1633  */
1634 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1635 {
1636         int node;
1637
1638         mem_cgroup_may_update_nodemask(memcg);
1639         node = memcg->last_scanned_node;
1640
1641         node = next_node(node, memcg->scan_nodes);
1642         if (node == MAX_NUMNODES)
1643                 node = first_node(memcg->scan_nodes);
1644         /*
1645          * We call this when we hit limit, not when pages are added to LRU.
1646          * No LRU may hold pages because all pages are UNEVICTABLE or
1647          * memcg is too small and all pages are not on LRU. In that case,
1648          * we use curret node.
1649          */
1650         if (unlikely(node == MAX_NUMNODES))
1651                 node = numa_node_id();
1652
1653         memcg->last_scanned_node = node;
1654         return node;
1655 }
1656
1657 /*
1658  * Check all nodes whether it contains reclaimable pages or not.
1659  * For quick scan, we make use of scan_nodes. This will allow us to skip
1660  * unused nodes. But scan_nodes is lazily updated and may not cotain
1661  * enough new information. We need to do double check.
1662  */
1663 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1664 {
1665         int nid;
1666
1667         /*
1668          * quick check...making use of scan_node.
1669          * We can skip unused nodes.
1670          */
1671         if (!nodes_empty(memcg->scan_nodes)) {
1672                 for (nid = first_node(memcg->scan_nodes);
1673                      nid < MAX_NUMNODES;
1674                      nid = next_node(nid, memcg->scan_nodes)) {
1675
1676                         if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1677                                 return true;
1678                 }
1679         }
1680         /*
1681          * Check rest of nodes.
1682          */
1683         for_each_node_state(nid, N_HIGH_MEMORY) {
1684                 if (node_isset(nid, memcg->scan_nodes))
1685                         continue;
1686                 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1687                         return true;
1688         }
1689         return false;
1690 }
1691
1692 #else
1693 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1694 {
1695         return 0;
1696 }
1697
1698 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1699 {
1700         return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1701 }
1702 #endif
1703
1704 /*
1705  * Scan the hierarchy if needed to reclaim memory. We remember the last child
1706  * we reclaimed from, so that we don't end up penalizing one child extensively
1707  * based on its position in the children list.
1708  *
1709  * root_memcg is the original ancestor that we've been reclaim from.
1710  *
1711  * We give up and return to the caller when we visit root_memcg twice.
1712  * (other groups can be removed while we're walking....)
1713  *
1714  * If shrink==true, for avoiding to free too much, this returns immedieately.
1715  */
1716 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1717                                                 struct zone *zone,
1718                                                 gfp_t gfp_mask,
1719                                                 unsigned long reclaim_options,
1720                                                 unsigned long *total_scanned)
1721 {
1722         struct mem_cgroup *victim;
1723         int ret, total = 0;
1724         int loop = 0;
1725         bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1726         bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1727         bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1728         unsigned long excess;
1729         unsigned long nr_scanned;
1730
1731         excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1732
1733         /* If memsw_is_minimum==1, swap-out is of-no-use. */
1734         if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1735                 noswap = true;
1736
1737         while (1) {
1738                 victim = mem_cgroup_select_victim(root_memcg);
1739                 if (victim == root_memcg) {
1740                         loop++;
1741                         /*
1742                          * We are not draining per cpu cached charges during
1743                          * soft limit reclaim  because global reclaim doesn't
1744                          * care about charges. It tries to free some memory and
1745                          * charges will not give any.
1746                          */
1747                         if (!check_soft && loop >= 1)
1748                                 drain_all_stock_async(root_memcg);
1749                         if (loop >= 2) {
1750                                 /*
1751                                  * If we have not been able to reclaim
1752                                  * anything, it might because there are
1753                                  * no reclaimable pages under this hierarchy
1754                                  */
1755                                 if (!check_soft || !total) {
1756                                         css_put(&victim->css);
1757                                         break;
1758                                 }
1759                                 /*
1760                                  * We want to do more targeted reclaim.
1761                                  * excess >> 2 is not to excessive so as to
1762                                  * reclaim too much, nor too less that we keep
1763                                  * coming back to reclaim from this cgroup
1764                                  */
1765                                 if (total >= (excess >> 2) ||
1766                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1767                                         css_put(&victim->css);
1768                                         break;
1769                                 }
1770                         }
1771                 }
1772                 if (!mem_cgroup_reclaimable(victim, noswap)) {
1773                         /* this cgroup's local usage == 0 */
1774                         css_put(&victim->css);
1775                         continue;
1776                 }
1777                 /* we use swappiness of local cgroup */
1778                 if (check_soft) {
1779                         ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1780                                 noswap, zone, &nr_scanned);
1781                         *total_scanned += nr_scanned;
1782                 } else
1783                         ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1784                                                 noswap);
1785                 css_put(&victim->css);
1786                 /*
1787                  * At shrinking usage, we can't check we should stop here or
1788                  * reclaim more. It's depends on callers. last_scanned_child
1789                  * will work enough for keeping fairness under tree.
1790                  */
1791                 if (shrink)
1792                         return ret;
1793                 total += ret;
1794                 if (check_soft) {
1795                         if (!res_counter_soft_limit_excess(&root_memcg->res))
1796                                 return total;
1797                 } else if (mem_cgroup_margin(root_memcg))
1798                         return total;
1799         }
1800         return total;
1801 }
1802
1803 /*
1804  * Check OOM-Killer is already running under our hierarchy.
1805  * If someone is running, return false.
1806  * Has to be called with memcg_oom_lock
1807  */
1808 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1809 {
1810         struct mem_cgroup *iter, *failed = NULL;
1811         bool cond = true;
1812
1813         for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1814                 if (iter->oom_lock) {
1815                         /*
1816                          * this subtree of our hierarchy is already locked
1817                          * so we cannot give a lock.
1818                          */
1819                         failed = iter;
1820                         cond = false;
1821                 } else
1822                         iter->oom_lock = true;
1823         }
1824
1825         if (!failed)
1826                 return true;
1827
1828         /*
1829          * OK, we failed to lock the whole subtree so we have to clean up
1830          * what we set up to the failing subtree
1831          */
1832         cond = true;
1833         for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1834                 if (iter == failed) {
1835                         cond = false;
1836                         continue;
1837                 }
1838                 iter->oom_lock = false;
1839         }
1840         return false;
1841 }
1842
1843 /*
1844  * Has to be called with memcg_oom_lock
1845  */
1846 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1847 {
1848         struct mem_cgroup *iter;
1849
1850         for_each_mem_cgroup_tree(iter, memcg)
1851                 iter->oom_lock = false;
1852         return 0;
1853 }
1854
1855 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1856 {
1857         struct mem_cgroup *iter;
1858
1859         for_each_mem_cgroup_tree(iter, memcg)
1860                 atomic_inc(&iter->under_oom);
1861 }
1862
1863 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1864 {
1865         struct mem_cgroup *iter;
1866
1867         /*
1868          * When a new child is created while the hierarchy is under oom,
1869          * mem_cgroup_oom_lock() may not be called. We have to use
1870          * atomic_add_unless() here.
1871          */
1872         for_each_mem_cgroup_tree(iter, memcg)
1873                 atomic_add_unless(&iter->under_oom, -1, 0);
1874 }
1875
1876 static DEFINE_SPINLOCK(memcg_oom_lock);
1877 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1878
1879 struct oom_wait_info {
1880         struct mem_cgroup *mem;
1881         wait_queue_t    wait;
1882 };
1883
1884 static int memcg_oom_wake_function(wait_queue_t *wait,
1885         unsigned mode, int sync, void *arg)
1886 {
1887         struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1888                           *oom_wait_memcg;
1889         struct oom_wait_info *oom_wait_info;
1890
1891         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1892         oom_wait_memcg = oom_wait_info->mem;
1893
1894         /*
1895          * Both of oom_wait_info->mem and wake_mem are stable under us.
1896          * Then we can use css_is_ancestor without taking care of RCU.
1897          */
1898         if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1899                 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1900                 return 0;
1901         return autoremove_wake_function(wait, mode, sync, arg);
1902 }
1903
1904 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1905 {
1906         /* for filtering, pass "memcg" as argument. */
1907         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1908 }
1909
1910 static void memcg_oom_recover(struct mem_cgroup *memcg)
1911 {
1912         if (memcg && atomic_read(&memcg->under_oom))
1913                 memcg_wakeup_oom(memcg);
1914 }
1915
1916 /*
1917  * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1918  */
1919 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1920 {
1921         struct oom_wait_info owait;
1922         bool locked, need_to_kill;
1923
1924         owait.mem = memcg;
1925         owait.wait.flags = 0;
1926         owait.wait.func = memcg_oom_wake_function;
1927         owait.wait.private = current;
1928         INIT_LIST_HEAD(&owait.wait.task_list);
1929         need_to_kill = true;
1930         mem_cgroup_mark_under_oom(memcg);
1931
1932         /* At first, try to OOM lock hierarchy under memcg.*/
1933         spin_lock(&memcg_oom_lock);
1934         locked = mem_cgroup_oom_lock(memcg);
1935         /*
1936          * Even if signal_pending(), we can't quit charge() loop without
1937          * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1938          * under OOM is always welcomed, use TASK_KILLABLE here.
1939          */
1940         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1941         if (!locked || memcg->oom_kill_disable)
1942                 need_to_kill = false;
1943         if (locked)
1944                 mem_cgroup_oom_notify(memcg);
1945         spin_unlock(&memcg_oom_lock);
1946
1947         if (need_to_kill) {
1948                 finish_wait(&memcg_oom_waitq, &owait.wait);
1949                 mem_cgroup_out_of_memory(memcg, mask);
1950         } else {
1951                 schedule();
1952                 finish_wait(&memcg_oom_waitq, &owait.wait);
1953         }
1954         spin_lock(&memcg_oom_lock);
1955         if (locked)
1956                 mem_cgroup_oom_unlock(memcg);
1957         memcg_wakeup_oom(memcg);
1958         spin_unlock(&memcg_oom_lock);
1959
1960         mem_cgroup_unmark_under_oom(memcg);
1961
1962         if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1963                 return false;
1964         /* Give chance to dying process */
1965         schedule_timeout_uninterruptible(1);
1966         return true;
1967 }
1968
1969 /*
1970  * Currently used to update mapped file statistics, but the routine can be
1971  * generalized to update other statistics as well.
1972  *
1973  * Notes: Race condition
1974  *
1975  * We usually use page_cgroup_lock() for accessing page_cgroup member but
1976  * it tends to be costly. But considering some conditions, we doesn't need
1977  * to do so _always_.
1978  *
1979  * Considering "charge", lock_page_cgroup() is not required because all
1980  * file-stat operations happen after a page is attached to radix-tree. There
1981  * are no race with "charge".
1982  *
1983  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1984  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1985  * if there are race with "uncharge". Statistics itself is properly handled
1986  * by flags.
1987  *
1988  * Considering "move", this is an only case we see a race. To make the race
1989  * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1990  * possibility of race condition. If there is, we take a lock.
1991  */
1992
1993 void mem_cgroup_update_page_stat(struct page *page,
1994                                  enum mem_cgroup_page_stat_item idx, int val)
1995 {
1996         struct mem_cgroup *memcg;
1997         struct page_cgroup *pc = lookup_page_cgroup(page);
1998         bool need_unlock = false;
1999         unsigned long uninitialized_var(flags);
2000
2001         if (unlikely(!pc))
2002                 return;
2003
2004         rcu_read_lock();
2005         memcg = pc->mem_cgroup;
2006         if (unlikely(!memcg || !PageCgroupUsed(pc)))
2007                 goto out;
2008         /* pc->mem_cgroup is unstable ? */
2009         if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2010                 /* take a lock against to access pc->mem_cgroup */
2011                 move_lock_page_cgroup(pc, &flags);
2012                 need_unlock = true;
2013                 memcg = pc->mem_cgroup;
2014                 if (!memcg || !PageCgroupUsed(pc))
2015                         goto out;
2016         }
2017
2018         switch (idx) {
2019         case MEMCG_NR_FILE_MAPPED:
2020                 if (val > 0)
2021                         SetPageCgroupFileMapped(pc);
2022                 else if (!page_mapped(page))
2023                         ClearPageCgroupFileMapped(pc);
2024                 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2025                 break;
2026         default:
2027                 BUG();
2028         }
2029
2030         this_cpu_add(memcg->stat->count[idx], val);
2031
2032 out:
2033         if (unlikely(need_unlock))
2034                 move_unlock_page_cgroup(pc, &flags);
2035         rcu_read_unlock();
2036         return;
2037 }
2038 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2039
2040 /*
2041  * size of first charge trial. "32" comes from vmscan.c's magic value.
2042  * TODO: maybe necessary to use big numbers in big irons.
2043  */
2044 #define CHARGE_BATCH    32U
2045 struct memcg_stock_pcp {
2046         struct mem_cgroup *cached; /* this never be root cgroup */
2047         unsigned int nr_pages;
2048         struct work_struct work;
2049         unsigned long flags;
2050 #define FLUSHING_CACHED_CHARGE  (0)
2051 };
2052 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2053 static DEFINE_MUTEX(percpu_charge_mutex);
2054
2055 /*
2056  * Try to consume stocked charge on this cpu. If success, one page is consumed
2057  * from local stock and true is returned. If the stock is 0 or charges from a
2058  * cgroup which is not current target, returns false. This stock will be
2059  * refilled.
2060  */
2061 static bool consume_stock(struct mem_cgroup *memcg)
2062 {
2063         struct memcg_stock_pcp *stock;
2064         bool ret = true;
2065
2066         stock = &get_cpu_var(memcg_stock);
2067         if (memcg == stock->cached && stock->nr_pages)
2068                 stock->nr_pages--;
2069         else /* need to call res_counter_charge */
2070                 ret = false;
2071         put_cpu_var(memcg_stock);
2072         return ret;
2073 }
2074
2075 /*
2076  * Returns stocks cached in percpu to res_counter and reset cached information.
2077  */
2078 static void drain_stock(struct memcg_stock_pcp *stock)
2079 {
2080         struct mem_cgroup *old = stock->cached;
2081
2082         if (stock->nr_pages) {
2083                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2084
2085                 res_counter_uncharge(&old->res, bytes);
2086                 if (do_swap_account)
2087                         res_counter_uncharge(&old->memsw, bytes);
2088                 stock->nr_pages = 0;
2089         }
2090         stock->cached = NULL;
2091 }
2092
2093 /*
2094  * This must be called under preempt disabled or must be called by
2095  * a thread which is pinned to local cpu.
2096  */
2097 static void drain_local_stock(struct work_struct *dummy)
2098 {
2099         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2100         drain_stock(stock);
2101         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2102 }
2103
2104 /*
2105  * Cache charges(val) which is from res_counter, to local per_cpu area.
2106  * This will be consumed by consume_stock() function, later.
2107  */
2108 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2109 {
2110         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2111
2112         if (stock->cached != memcg) { /* reset if necessary */
2113                 drain_stock(stock);
2114                 stock->cached = memcg;
2115         }
2116         stock->nr_pages += nr_pages;
2117         put_cpu_var(memcg_stock);
2118 }
2119
2120 /*
2121  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2122  * of the hierarchy under it. sync flag says whether we should block
2123  * until the work is done.
2124  */
2125 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2126 {
2127         int cpu, curcpu;
2128
2129         /* Notify other cpus that system-wide "drain" is running */
2130         get_online_cpus();
2131         curcpu = get_cpu();
2132         for_each_online_cpu(cpu) {
2133                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2134                 struct mem_cgroup *memcg;
2135
2136                 memcg = stock->cached;
2137                 if (!memcg || !stock->nr_pages)
2138                         continue;
2139                 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2140                         continue;
2141                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2142                         if (cpu == curcpu)
2143                                 drain_local_stock(&stock->work);
2144                         else
2145                                 schedule_work_on(cpu, &stock->work);
2146                 }
2147         }
2148         put_cpu();
2149
2150         if (!sync)
2151                 goto out;
2152
2153         for_each_online_cpu(cpu) {
2154                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2155                 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2156                         flush_work(&stock->work);
2157         }
2158 out:
2159         put_online_cpus();
2160 }
2161
2162 /*
2163  * Tries to drain stocked charges in other cpus. This function is asynchronous
2164  * and just put a work per cpu for draining localy on each cpu. Caller can
2165  * expects some charges will be back to res_counter later but cannot wait for
2166  * it.
2167  */
2168 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2169 {
2170         /*
2171          * If someone calls draining, avoid adding more kworker runs.
2172          */
2173         if (!mutex_trylock(&percpu_charge_mutex))
2174                 return;
2175         drain_all_stock(root_memcg, false);
2176         mutex_unlock(&percpu_charge_mutex);
2177 }
2178
2179 /* This is a synchronous drain interface. */
2180 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2181 {
2182         /* called when force_empty is called */
2183         mutex_lock(&percpu_charge_mutex);
2184         drain_all_stock(root_memcg, true);
2185         mutex_unlock(&percpu_charge_mutex);
2186 }
2187
2188 /*
2189  * This function drains percpu counter value from DEAD cpu and
2190  * move it to local cpu. Note that this function can be preempted.
2191  */
2192 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2193 {
2194         int i;
2195
2196         spin_lock(&memcg->pcp_counter_lock);
2197         for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2198                 long x = per_cpu(memcg->stat->count[i], cpu);
2199
2200                 per_cpu(memcg->stat->count[i], cpu) = 0;
2201                 memcg->nocpu_base.count[i] += x;
2202         }
2203         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2204                 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2205
2206                 per_cpu(memcg->stat->events[i], cpu) = 0;
2207                 memcg->nocpu_base.events[i] += x;
2208         }
2209         /* need to clear ON_MOVE value, works as a kind of lock. */
2210         per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2211         spin_unlock(&memcg->pcp_counter_lock);
2212 }
2213
2214 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2215 {
2216         int idx = MEM_CGROUP_ON_MOVE;
2217
2218         spin_lock(&memcg->pcp_counter_lock);
2219         per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2220         spin_unlock(&memcg->pcp_counter_lock);
2221 }
2222
2223 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2224                                         unsigned long action,
2225                                         void *hcpu)
2226 {
2227         int cpu = (unsigned long)hcpu;
2228         struct memcg_stock_pcp *stock;
2229         struct mem_cgroup *iter;
2230
2231         if ((action == CPU_ONLINE)) {
2232                 for_each_mem_cgroup_all(iter)
2233                         synchronize_mem_cgroup_on_move(iter, cpu);
2234                 return NOTIFY_OK;
2235         }
2236
2237         if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2238                 return NOTIFY_OK;
2239
2240         for_each_mem_cgroup_all(iter)
2241                 mem_cgroup_drain_pcp_counter(iter, cpu);
2242
2243         stock = &per_cpu(memcg_stock, cpu);
2244         drain_stock(stock);
2245         return NOTIFY_OK;
2246 }
2247
2248
2249 /* See __mem_cgroup_try_charge() for details */
2250 enum {
2251         CHARGE_OK,              /* success */
2252         CHARGE_RETRY,           /* need to retry but retry is not bad */
2253         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2254         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2255         CHARGE_OOM_DIE,         /* the current is killed because of OOM */
2256 };
2257
2258 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2259                                 unsigned int nr_pages, bool oom_check)
2260 {
2261         unsigned long csize = nr_pages * PAGE_SIZE;
2262         struct mem_cgroup *mem_over_limit;
2263         struct res_counter *fail_res;
2264         unsigned long flags = 0;
2265         int ret;
2266
2267         ret = res_counter_charge(&memcg->res, csize, &fail_res);
2268
2269         if (likely(!ret)) {
2270                 if (!do_swap_account)
2271                         return CHARGE_OK;
2272                 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2273                 if (likely(!ret))
2274                         return CHARGE_OK;
2275
2276                 res_counter_uncharge(&memcg->res, csize);
2277                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2278                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2279         } else
2280                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2281         /*
2282          * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2283          * of regular pages (CHARGE_BATCH), or a single regular page (1).
2284          *
2285          * Never reclaim on behalf of optional batching, retry with a
2286          * single page instead.
2287          */
2288         if (nr_pages == CHARGE_BATCH)
2289                 return CHARGE_RETRY;
2290
2291         if (!(gfp_mask & __GFP_WAIT))
2292                 return CHARGE_WOULDBLOCK;
2293
2294         ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2295                                               gfp_mask, flags, NULL);
2296         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2297                 return CHARGE_RETRY;
2298         /*
2299          * Even though the limit is exceeded at this point, reclaim
2300          * may have been able to free some pages.  Retry the charge
2301          * before killing the task.
2302          *
2303          * Only for regular pages, though: huge pages are rather
2304          * unlikely to succeed so close to the limit, and we fall back
2305          * to regular pages anyway in case of failure.
2306          */
2307         if (nr_pages == 1 && ret)
2308                 return CHARGE_RETRY;
2309
2310         /*
2311          * At task move, charge accounts can be doubly counted. So, it's
2312          * better to wait until the end of task_move if something is going on.
2313          */
2314         if (mem_cgroup_wait_acct_move(mem_over_limit))
2315                 return CHARGE_RETRY;
2316
2317         /* If we don't need to call oom-killer at el, return immediately */
2318         if (!oom_check)
2319                 return CHARGE_NOMEM;
2320         /* check OOM */
2321         if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2322                 return CHARGE_OOM_DIE;
2323
2324         return CHARGE_RETRY;
2325 }
2326
2327 /*
2328  * Unlike exported interface, "oom" parameter is added. if oom==true,
2329  * oom-killer can be invoked.
2330  */
2331 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2332                                    gfp_t gfp_mask,
2333                                    unsigned int nr_pages,
2334                                    struct mem_cgroup **ptr,
2335                                    bool oom)
2336 {
2337         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2338         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2339         struct mem_cgroup *memcg = NULL;
2340         int ret;
2341
2342         /*
2343          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2344          * in system level. So, allow to go ahead dying process in addition to
2345          * MEMDIE process.
2346          */
2347         if (unlikely(test_thread_flag(TIF_MEMDIE)
2348                      || fatal_signal_pending(current)))
2349                 goto bypass;
2350
2351         /*
2352          * We always charge the cgroup the mm_struct belongs to.
2353          * The mm_struct's mem_cgroup changes on task migration if the
2354          * thread group leader migrates. It's possible that mm is not
2355          * set, if so charge the init_mm (happens for pagecache usage).
2356          */
2357         if (!*ptr && !mm)
2358                 goto bypass;
2359 again:
2360         if (*ptr) { /* css should be a valid one */
2361                 memcg = *ptr;
2362                 VM_BUG_ON(css_is_removed(&memcg->css));
2363                 if (mem_cgroup_is_root(memcg))
2364                         goto done;
2365                 if (nr_pages == 1 && consume_stock(memcg))
2366                         goto done;
2367                 css_get(&memcg->css);
2368         } else {
2369                 struct task_struct *p;
2370
2371                 rcu_read_lock();
2372                 p = rcu_dereference(mm->owner);
2373                 /*
2374                  * Because we don't have task_lock(), "p" can exit.
2375                  * In that case, "memcg" can point to root or p can be NULL with
2376                  * race with swapoff. Then, we have small risk of mis-accouning.
2377                  * But such kind of mis-account by race always happens because
2378                  * we don't have cgroup_mutex(). It's overkill and we allo that
2379                  * small race, here.
2380                  * (*) swapoff at el will charge against mm-struct not against
2381                  * task-struct. So, mm->owner can be NULL.
2382                  */
2383                 memcg = mem_cgroup_from_task(p);
2384                 if (!memcg || mem_cgroup_is_root(memcg)) {
2385                         rcu_read_unlock();
2386                         goto done;
2387                 }
2388                 if (nr_pages == 1 && consume_stock(memcg)) {
2389                         /*
2390                          * It seems dagerous to access memcg without css_get().
2391                          * But considering how consume_stok works, it's not
2392                          * necessary. If consume_stock success, some charges
2393                          * from this memcg are cached on this cpu. So, we
2394                          * don't need to call css_get()/css_tryget() before
2395                          * calling consume_stock().
2396                          */
2397                         rcu_read_unlock();
2398                         goto done;
2399                 }
2400                 /* after here, we may be blocked. we need to get refcnt */
2401                 if (!css_tryget(&memcg->css)) {
2402                         rcu_read_unlock();
2403                         goto again;
2404                 }
2405                 rcu_read_unlock();
2406         }
2407
2408         do {
2409                 bool oom_check;
2410
2411                 /* If killed, bypass charge */
2412                 if (fatal_signal_pending(current)) {
2413                         css_put(&memcg->css);
2414                         goto bypass;
2415                 }
2416
2417                 oom_check = false;
2418                 if (oom && !nr_oom_retries) {
2419                         oom_check = true;
2420                         nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2421                 }
2422
2423                 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2424                 switch (ret) {
2425                 case CHARGE_OK:
2426                         break;
2427                 case CHARGE_RETRY: /* not in OOM situation but retry */
2428                         batch = nr_pages;
2429                         css_put(&memcg->css);
2430                         memcg = NULL;
2431                         goto again;
2432                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2433                         css_put(&memcg->css);
2434                         goto nomem;
2435                 case CHARGE_NOMEM: /* OOM routine works */
2436                         if (!oom) {
2437                                 css_put(&memcg->css);
2438                                 goto nomem;
2439                         }
2440                         /* If oom, we never return -ENOMEM */
2441                         nr_oom_retries--;
2442                         break;
2443                 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2444                         css_put(&memcg->css);
2445                         goto bypass;
2446                 }
2447         } while (ret != CHARGE_OK);
2448
2449         if (batch > nr_pages)
2450                 refill_stock(memcg, batch - nr_pages);
2451         css_put(&memcg->css);
2452 done:
2453         *ptr = memcg;
2454         return 0;
2455 nomem:
2456         *ptr = NULL;
2457         return -ENOMEM;
2458 bypass:
2459         *ptr = NULL;
2460         return 0;
2461 }
2462
2463 /*
2464  * Somemtimes we have to undo a charge we got by try_charge().
2465  * This function is for that and do uncharge, put css's refcnt.
2466  * gotten by try_charge().
2467  */
2468 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2469                                        unsigned int nr_pages)
2470 {
2471         if (!mem_cgroup_is_root(memcg)) {
2472                 unsigned long bytes = nr_pages * PAGE_SIZE;
2473
2474                 res_counter_uncharge(&memcg->res, bytes);
2475                 if (do_swap_account)
2476                         res_counter_uncharge(&memcg->memsw, bytes);
2477         }
2478 }
2479
2480 /*
2481  * A helper function to get mem_cgroup from ID. must be called under
2482  * rcu_read_lock(). The caller must check css_is_removed() or some if
2483  * it's concern. (dropping refcnt from swap can be called against removed
2484  * memcg.)
2485  */
2486 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2487 {
2488         struct cgroup_subsys_state *css;
2489
2490         /* ID 0 is unused ID */
2491         if (!id)
2492                 return NULL;
2493         css = css_lookup(&mem_cgroup_subsys, id);
2494         if (!css)
2495                 return NULL;
2496         return container_of(css, struct mem_cgroup, css);
2497 }
2498
2499 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2500 {
2501         struct mem_cgroup *memcg = NULL;
2502         struct page_cgroup *pc;
2503         unsigned short id;
2504         swp_entry_t ent;
2505
2506         VM_BUG_ON(!PageLocked(page));
2507
2508         pc = lookup_page_cgroup(page);
2509         lock_page_cgroup(pc);
2510         if (PageCgroupUsed(pc)) {
2511                 memcg = pc->mem_cgroup;
2512                 if (memcg && !css_tryget(&memcg->css))
2513                         memcg = NULL;
2514         } else if (PageSwapCache(page)) {
2515                 ent.val = page_private(page);
2516                 id = lookup_swap_cgroup(ent);
2517                 rcu_read_lock();
2518                 memcg = mem_cgroup_lookup(id);
2519                 if (memcg && !css_tryget(&memcg->css))
2520                         memcg = NULL;
2521                 rcu_read_unlock();
2522         }
2523         unlock_page_cgroup(pc);
2524         return memcg;
2525 }
2526
2527 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2528                                        struct page *page,
2529                                        unsigned int nr_pages,
2530                                        struct page_cgroup *pc,
2531                                        enum charge_type ctype)
2532 {
2533         lock_page_cgroup(pc);
2534         if (unlikely(PageCgroupUsed(pc))) {
2535                 unlock_page_cgroup(pc);
2536                 __mem_cgroup_cancel_charge(memcg, nr_pages);
2537                 return;
2538         }
2539         /*
2540          * we don't need page_cgroup_lock about tail pages, becase they are not
2541          * accessed by any other context at this point.
2542          */
2543         pc->mem_cgroup = memcg;
2544         /*
2545          * We access a page_cgroup asynchronously without lock_page_cgroup().
2546          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2547          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2548          * before USED bit, we need memory barrier here.
2549          * See mem_cgroup_add_lru_list(), etc.
2550          */
2551         smp_wmb();
2552         switch (ctype) {
2553         case MEM_CGROUP_CHARGE_TYPE_CACHE:
2554         case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2555                 SetPageCgroupCache(pc);
2556                 SetPageCgroupUsed(pc);
2557                 break;
2558         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2559                 ClearPageCgroupCache(pc);
2560                 SetPageCgroupUsed(pc);
2561                 break;
2562         default:
2563                 break;
2564         }
2565
2566         mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2567         unlock_page_cgroup(pc);
2568         /*
2569          * "charge_statistics" updated event counter. Then, check it.
2570          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2571          * if they exceeds softlimit.
2572          */
2573         memcg_check_events(memcg, page);
2574 }
2575
2576 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2577
2578 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2579                         (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2580 /*
2581  * Because tail pages are not marked as "used", set it. We're under
2582  * zone->lru_lock, 'splitting on pmd' and compund_lock.
2583  */
2584 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2585 {
2586         struct page_cgroup *head_pc = lookup_page_cgroup(head);
2587         struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2588         unsigned long flags;
2589
2590         if (mem_cgroup_disabled())
2591                 return;
2592         /*
2593          * We have no races with charge/uncharge but will have races with
2594          * page state accounting.
2595          */
2596         move_lock_page_cgroup(head_pc, &flags);
2597
2598         tail_pc->mem_cgroup = head_pc->mem_cgroup;
2599         smp_wmb(); /* see __commit_charge() */
2600         if (PageCgroupAcctLRU(head_pc)) {
2601                 enum lru_list lru;
2602                 struct mem_cgroup_per_zone *mz;
2603
2604                 /*
2605                  * LRU flags cannot be copied because we need to add tail
2606                  *.page to LRU by generic call and our hook will be called.
2607                  * We hold lru_lock, then, reduce counter directly.
2608                  */
2609                 lru = page_lru(head);
2610                 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2611                 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2612         }
2613         tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2614         move_unlock_page_cgroup(head_pc, &flags);
2615 }
2616 #endif
2617
2618 /**
2619  * mem_cgroup_move_account - move account of the page
2620  * @page: the page
2621  * @nr_pages: number of regular pages (>1 for huge pages)
2622  * @pc: page_cgroup of the page.
2623  * @from: mem_cgroup which the page is moved from.
2624  * @to: mem_cgroup which the page is moved to. @from != @to.
2625  * @uncharge: whether we should call uncharge and css_put against @from.
2626  *
2627  * The caller must confirm following.
2628  * - page is not on LRU (isolate_page() is useful.)
2629  * - compound_lock is held when nr_pages > 1
2630  *
2631  * This function doesn't do "charge" nor css_get to new cgroup. It should be
2632  * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2633  * true, this function does "uncharge" from old cgroup, but it doesn't if
2634  * @uncharge is false, so a caller should do "uncharge".
2635  */
2636 static int mem_cgroup_move_account(struct page *page,
2637                                    unsigned int nr_pages,
2638                                    struct page_cgroup *pc,
2639                                    struct mem_cgroup *from,
2640                                    struct mem_cgroup *to,
2641                                    bool uncharge)
2642 {
2643         unsigned long flags;
2644         int ret;
2645
2646         VM_BUG_ON(from == to);
2647         VM_BUG_ON(PageLRU(page));
2648         /*
2649          * The page is isolated from LRU. So, collapse function
2650          * will not handle this page. But page splitting can happen.
2651          * Do this check under compound_page_lock(). The caller should
2652          * hold it.
2653          */
2654         ret = -EBUSY;
2655         if (nr_pages > 1 && !PageTransHuge(page))
2656                 goto out;
2657
2658         lock_page_cgroup(pc);
2659
2660         ret = -EINVAL;
2661         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2662                 goto unlock;
2663
2664         move_lock_page_cgroup(pc, &flags);
2665
2666         if (PageCgroupFileMapped(pc)) {
2667                 /* Update mapped_file data for mem_cgroup */
2668                 preempt_disable();
2669                 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2670                 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2671                 preempt_enable();
2672         }
2673         mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2674         if (uncharge)
2675                 /* This is not "cancel", but cancel_charge does all we need. */
2676                 __mem_cgroup_cancel_charge(from, nr_pages);
2677
2678         /* caller should have done css_get */
2679         pc->mem_cgroup = to;
2680         mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2681         /*
2682          * We charges against "to" which may not have any tasks. Then, "to"
2683          * can be under rmdir(). But in current implementation, caller of
2684          * this function is just force_empty() and move charge, so it's
2685          * guaranteed that "to" is never removed. So, we don't check rmdir
2686          * status here.
2687          */
2688         move_unlock_page_cgroup(pc, &flags);
2689         ret = 0;
2690 unlock:
2691         unlock_page_cgroup(pc);
2692         /*
2693          * check events
2694          */
2695         memcg_check_events(to, page);
2696         memcg_check_events(from, page);
2697 out:
2698         return ret;
2699 }
2700
2701 /*
2702  * move charges to its parent.
2703  */
2704
2705 static int mem_cgroup_move_parent(struct page *page,
2706                                   struct page_cgroup *pc,
2707                                   struct mem_cgroup *child,
2708                                   gfp_t gfp_mask)
2709 {
2710         struct cgroup *cg = child->css.cgroup;
2711         struct cgroup *pcg = cg->parent;
2712         struct mem_cgroup *parent;
2713         unsigned int nr_pages;
2714         unsigned long uninitialized_var(flags);
2715         int ret;
2716
2717         /* Is ROOT ? */
2718         if (!pcg)
2719                 return -EINVAL;
2720
2721         ret = -EBUSY;
2722         if (!get_page_unless_zero(page))
2723                 goto out;
2724         if (isolate_lru_page(page))
2725                 goto put;
2726
2727         nr_pages = hpage_nr_pages(page);
2728
2729         parent = mem_cgroup_from_cont(pcg);
2730         ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2731         if (ret || !parent)
2732                 goto put_back;
2733
2734         if (nr_pages > 1)
2735                 flags = compound_lock_irqsave(page);
2736
2737         ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2738         if (ret)
2739                 __mem_cgroup_cancel_charge(parent, nr_pages);
2740
2741         if (nr_pages > 1)
2742                 compound_unlock_irqrestore(page, flags);
2743 put_back:
2744         putback_lru_page(page);
2745 put:
2746         put_page(page);
2747 out:
2748         return ret;
2749 }
2750
2751 /*
2752  * Charge the memory controller for page usage.
2753  * Return
2754  * 0 if the charge was successful
2755  * < 0 if the cgroup is over its limit
2756  */
2757 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2758                                 gfp_t gfp_mask, enum charge_type ctype)
2759 {
2760         struct mem_cgroup *memcg = NULL;
2761         unsigned int nr_pages = 1;
2762         struct page_cgroup *pc;
2763         bool oom = true;
2764         int ret;
2765
2766         if (PageTransHuge(page)) {
2767                 nr_pages <<= compound_order(page);
2768                 VM_BUG_ON(!PageTransHuge(page));
2769                 /*
2770                  * Never OOM-kill a process for a huge page.  The
2771                  * fault handler will fall back to regular pages.
2772                  */
2773                 oom = false;
2774         }
2775
2776         pc = lookup_page_cgroup(page);
2777         BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2778
2779         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2780         if (ret || !memcg)
2781                 return ret;
2782
2783         __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2784         return 0;
2785 }
2786
2787 int mem_cgroup_newpage_charge(struct page *page,
2788                               struct mm_struct *mm, gfp_t gfp_mask)
2789 {
2790         if (mem_cgroup_disabled())
2791                 return 0;
2792         /*
2793          * If already mapped, we don't have to account.
2794          * If page cache, page->mapping has address_space.
2795          * But page->mapping may have out-of-use anon_vma pointer,
2796          * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2797          * is NULL.
2798          */
2799         if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2800                 return 0;
2801         if (unlikely(!mm))
2802                 mm = &init_mm;
2803         return mem_cgroup_charge_common(page, mm, gfp_mask,
2804                                 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2805 }
2806
2807 static void
2808 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2809                                         enum charge_type ctype);
2810
2811 static void
2812 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2813                                         enum charge_type ctype)
2814 {
2815         struct page_cgroup *pc = lookup_page_cgroup(page);
2816         /*
2817          * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2818          * is already on LRU. It means the page may on some other page_cgroup's
2819          * LRU. Take care of it.
2820          */
2821         mem_cgroup_lru_del_before_commit(page);
2822         __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2823         mem_cgroup_lru_add_after_commit(page);
2824         return;
2825 }
2826
2827 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2828                                 gfp_t gfp_mask)
2829 {
2830         struct mem_cgroup *memcg = NULL;
2831         int ret;
2832
2833         if (mem_cgroup_disabled())
2834                 return 0;
2835         if (PageCompound(page))
2836                 return 0;
2837
2838         if (unlikely(!mm))
2839                 mm = &init_mm;
2840
2841         if (page_is_file_cache(page)) {
2842                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2843                 if (ret || !memcg)
2844                         return ret;
2845
2846                 /*
2847                  * FUSE reuses pages without going through the final
2848                  * put that would remove them from the LRU list, make
2849                  * sure that they get relinked properly.
2850                  */
2851                 __mem_cgroup_commit_charge_lrucare(page, memcg,
2852                                         MEM_CGROUP_CHARGE_TYPE_CACHE);
2853                 return ret;
2854         }
2855         /* shmem */
2856         if (PageSwapCache(page)) {
2857                 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2858                 if (!ret)
2859                         __mem_cgroup_commit_charge_swapin(page, memcg,
2860                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2861         } else
2862                 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2863                                         MEM_CGROUP_CHARGE_TYPE_SHMEM);
2864
2865         return ret;
2866 }
2867
2868 /*
2869  * While swap-in, try_charge -> commit or cancel, the page is locked.
2870  * And when try_charge() successfully returns, one refcnt to memcg without
2871  * struct page_cgroup is acquired. This refcnt will be consumed by
2872  * "commit()" or removed by "cancel()"
2873  */
2874 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2875                                  struct page *page,
2876                                  gfp_t mask, struct mem_cgroup **ptr)
2877 {
2878         struct mem_cgroup *memcg;
2879         int ret;
2880
2881         *ptr = NULL;
2882
2883         if (mem_cgroup_disabled())
2884                 return 0;
2885
2886         if (!do_swap_account)
2887                 goto charge_cur_mm;
2888         /*
2889          * A racing thread's fault, or swapoff, may have already updated
2890          * the pte, and even removed page from swap cache: in those cases
2891          * do_swap_page()'s pte_same() test will fail; but there's also a
2892          * KSM case which does need to charge the page.
2893          */
2894         if (!PageSwapCache(page))
2895                 goto charge_cur_mm;
2896         memcg = try_get_mem_cgroup_from_page(page);
2897         if (!memcg)
2898                 goto charge_cur_mm;
2899         *ptr = memcg;
2900         ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2901         css_put(&memcg->css);
2902         return ret;
2903 charge_cur_mm:
2904         if (unlikely(!mm))
2905                 mm = &init_mm;
2906         return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2907 }
2908
2909 static void
2910 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2911                                         enum charge_type ctype)
2912 {
2913         if (mem_cgroup_disabled())
2914                 return;
2915         if (!ptr)
2916                 return;
2917         cgroup_exclude_rmdir(&ptr->css);
2918
2919         __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2920         /*
2921          * Now swap is on-memory. This means this page may be
2922          * counted both as mem and swap....double count.
2923          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2924          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2925          * may call delete_from_swap_cache() before reach here.
2926          */
2927         if (do_swap_account && PageSwapCache(page)) {
2928                 swp_entry_t ent = {.val = page_private(page)};
2929                 unsigned short id;
2930                 struct mem_cgroup *memcg;
2931
2932                 id = swap_cgroup_record(ent, 0);
2933                 rcu_read_lock();
2934                 memcg = mem_cgroup_lookup(id);
2935                 if (memcg) {
2936                         /*
2937                          * This recorded memcg can be obsolete one. So, avoid
2938                          * calling css_tryget
2939                          */
2940                         if (!mem_cgroup_is_root(memcg))
2941                                 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2942                         mem_cgroup_swap_statistics(memcg, false);
2943                         mem_cgroup_put(memcg);
2944                 }
2945                 rcu_read_unlock();
2946         }
2947         /*
2948          * At swapin, we may charge account against cgroup which has no tasks.
2949          * So, rmdir()->pre_destroy() can be called while we do this charge.
2950          * In that case, we need to call pre_destroy() again. check it here.
2951          */
2952         cgroup_release_and_wakeup_rmdir(&ptr->css);
2953 }
2954
2955 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2956 {
2957         __mem_cgroup_commit_charge_swapin(page, ptr,
2958                                         MEM_CGROUP_CHARGE_TYPE_MAPPED);
2959 }
2960
2961 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2962 {
2963         if (mem_cgroup_disabled())
2964                 return;
2965         if (!memcg)
2966                 return;
2967         __mem_cgroup_cancel_charge(memcg, 1);
2968 }
2969
2970 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2971                                    unsigned int nr_pages,
2972                                    const enum charge_type ctype)
2973 {
2974         struct memcg_batch_info *batch = NULL;
2975         bool uncharge_memsw = true;
2976
2977         /* If swapout, usage of swap doesn't decrease */
2978         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2979                 uncharge_memsw = false;
2980
2981         batch = &current->memcg_batch;
2982         /*
2983          * In usual, we do css_get() when we remember memcg pointer.
2984          * But in this case, we keep res->usage until end of a series of
2985          * uncharges. Then, it's ok to ignore memcg's refcnt.
2986          */
2987         if (!batch->memcg)
2988                 batch->memcg = memcg;
2989         /*
2990          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2991          * In those cases, all pages freed continuously can be expected to be in
2992          * the same cgroup and we have chance to coalesce uncharges.
2993          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2994          * because we want to do uncharge as soon as possible.
2995          */
2996
2997         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2998                 goto direct_uncharge;
2999
3000         if (nr_pages > 1)
3001                 goto direct_uncharge;
3002
3003         /*
3004          * In typical case, batch->memcg == mem. This means we can
3005          * merge a series of uncharges to an uncharge of res_counter.
3006          * If not, we uncharge res_counter ony by one.
3007          */
3008         if (batch->memcg != memcg)
3009                 goto direct_uncharge;
3010         /* remember freed charge and uncharge it later */
3011         batch->nr_pages++;
3012         if (uncharge_memsw)
3013                 batch->memsw_nr_pages++;
3014         return;
3015 direct_uncharge:
3016         res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3017         if (uncharge_memsw)
3018                 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3019         if (unlikely(batch->memcg != memcg))
3020                 memcg_oom_recover(memcg);
3021         return;
3022 }
3023
3024 /*
3025  * uncharge if !page_mapped(page)
3026  */
3027 static struct mem_cgroup *
3028 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3029 {
3030         struct mem_cgroup *memcg = NULL;
3031         unsigned int nr_pages = 1;
3032         struct page_cgroup *pc;
3033
3034         if (mem_cgroup_disabled())
3035                 return NULL;
3036
3037         if (PageSwapCache(page))
3038                 return NULL;
3039
3040         if (PageTransHuge(page)) {
3041                 nr_pages <<= compound_order(page);
3042                 VM_BUG_ON(!PageTransHuge(page));
3043         }
3044         /*
3045          * Check if our page_cgroup is valid
3046          */
3047         pc = lookup_page_cgroup(page);
3048         if (unlikely(!pc || !PageCgroupUsed(pc)))
3049                 return NULL;
3050
3051         lock_page_cgroup(pc);
3052
3053         memcg = pc->mem_cgroup;
3054
3055         if (!PageCgroupUsed(pc))
3056                 goto unlock_out;
3057
3058         switch (ctype) {
3059         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3060         case MEM_CGROUP_CHARGE_TYPE_DROP:
3061                 /* See mem_cgroup_prepare_migration() */
3062                 if (page_mapped(page) || PageCgroupMigration(pc))
3063                         goto unlock_out;
3064                 break;
3065         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3066                 if (!PageAnon(page)) {  /* Shared memory */
3067                         if (page->mapping && !page_is_file_cache(page))
3068                                 goto unlock_out;
3069                 } else if (page_mapped(page)) /* Anon */
3070                                 goto unlock_out;
3071                 break;
3072         default:
3073                 break;
3074         }
3075
3076         mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3077
3078         ClearPageCgroupUsed(pc);
3079         /*
3080          * pc->mem_cgroup is not cleared here. It will be accessed when it's
3081          * freed from LRU. This is safe because uncharged page is expected not
3082          * to be reused (freed soon). Exception is SwapCache, it's handled by
3083          * special functions.
3084          */
3085
3086         unlock_page_cgroup(pc);
3087         /*
3088          * even after unlock, we have memcg->res.usage here and this memcg
3089          * will never be freed.
3090          */
3091         memcg_check_events(memcg, page);
3092         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3093                 mem_cgroup_swap_statistics(memcg, true);
3094                 mem_cgroup_get(memcg);
3095         }
3096         if (!mem_cgroup_is_root(memcg))
3097                 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3098
3099         return memcg;
3100
3101 unlock_out:
3102         unlock_page_cgroup(pc);
3103         return NULL;
3104 }
3105
3106 void mem_cgroup_uncharge_page(struct page *page)
3107 {
3108         /* early check. */
3109         if (page_mapped(page))
3110                 return;
3111         if (page->mapping && !PageAnon(page))
3112                 return;
3113         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3114 }
3115
3116 void mem_cgroup_uncharge_cache_page(struct page *page)
3117 {
3118         VM_BUG_ON(page_mapped(page));
3119         VM_BUG_ON(page->mapping);
3120         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3121 }
3122
3123 /*
3124  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3125  * In that cases, pages are freed continuously and we can expect pages
3126  * are in the same memcg. All these calls itself limits the number of
3127  * pages freed at once, then uncharge_start/end() is called properly.
3128  * This may be called prural(2) times in a context,
3129  */
3130
3131 void mem_cgroup_uncharge_start(void)
3132 {
3133         current->memcg_batch.do_batch++;
3134         /* We can do nest. */
3135         if (current->memcg_batch.do_batch == 1) {
3136                 current->memcg_batch.memcg = NULL;
3137                 current->memcg_batch.nr_pages = 0;
3138                 current->memcg_batch.memsw_nr_pages = 0;
3139         }
3140 }
3141
3142 void mem_cgroup_uncharge_end(void)
3143 {
3144         struct memcg_batch_info *batch = &current->memcg_batch;
3145
3146         if (!batch->do_batch)
3147                 return;
3148
3149         batch->do_batch--;
3150         if (batch->do_batch) /* If stacked, do nothing. */
3151                 return;
3152
3153         if (!batch->memcg)
3154                 return;
3155         /*
3156          * This "batch->memcg" is valid without any css_get/put etc...
3157          * bacause we hide charges behind us.
3158          */
3159         if (batch->nr_pages)
3160                 res_counter_uncharge(&batch->memcg->res,
3161                                      batch->nr_pages * PAGE_SIZE);
3162         if (batch->memsw_nr_pages)
3163                 res_counter_uncharge(&batch->memcg->memsw,
3164                                      batch->memsw_nr_pages * PAGE_SIZE);
3165         memcg_oom_recover(batch->memcg);
3166         /* forget this pointer (for sanity check) */
3167         batch->memcg = NULL;
3168 }
3169
3170 #ifdef CONFIG_SWAP
3171 /*
3172  * called after __delete_from_swap_cache() and drop "page" account.
3173  * memcg information is recorded to swap_cgroup of "ent"
3174  */
3175 void
3176 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3177 {
3178         struct mem_cgroup *memcg;
3179         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3180
3181         if (!swapout) /* this was a swap cache but the swap is unused ! */
3182                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3183
3184         memcg = __mem_cgroup_uncharge_common(page, ctype);
3185
3186         /*
3187          * record memcg information,  if swapout && memcg != NULL,
3188          * mem_cgroup_get() was called in uncharge().
3189          */
3190         if (do_swap_account && swapout && memcg)
3191                 swap_cgroup_record(ent, css_id(&memcg->css));
3192 }
3193 #endif
3194
3195 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3196 /*
3197  * called from swap_entry_free(). remove record in swap_cgroup and
3198  * uncharge "memsw" account.
3199  */
3200 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3201 {
3202         struct mem_cgroup *memcg;
3203         unsigned short id;
3204
3205         if (!do_swap_account)
3206                 return;
3207
3208         id = swap_cgroup_record(ent, 0);
3209         rcu_read_lock();
3210         memcg = mem_cgroup_lookup(id);
3211         if (memcg) {
3212                 /*
3213                  * We uncharge this because swap is freed.
3214                  * This memcg can be obsolete one. We avoid calling css_tryget
3215                  */
3216                 if (!mem_cgroup_is_root(memcg))
3217                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3218                 mem_cgroup_swap_statistics(memcg, false);
3219                 mem_cgroup_put(memcg);
3220         }
3221         rcu_read_unlock();
3222 }
3223
3224 /**
3225  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3226  * @entry: swap entry to be moved
3227  * @from:  mem_cgroup which the entry is moved from
3228  * @to:  mem_cgroup which the entry is moved to
3229  * @need_fixup: whether we should fixup res_counters and refcounts.
3230  *
3231  * It succeeds only when the swap_cgroup's record for this entry is the same
3232  * as the mem_cgroup's id of @from.
3233  *
3234  * Returns 0 on success, -EINVAL on failure.
3235  *
3236  * The caller must have charged to @to, IOW, called res_counter_charge() about
3237  * both res and memsw, and called css_get().
3238  */
3239 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3240                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3241 {
3242         unsigned short old_id, new_id;
3243
3244         old_id = css_id(&from->css);
3245         new_id = css_id(&to->css);
3246
3247         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3248                 mem_cgroup_swap_statistics(from, false);
3249                 mem_cgroup_swap_statistics(to, true);
3250                 /*
3251                  * This function is only called from task migration context now.
3252                  * It postpones res_counter and refcount handling till the end
3253                  * of task migration(mem_cgroup_clear_mc()) for performance
3254                  * improvement. But we cannot postpone mem_cgroup_get(to)
3255                  * because if the process that has been moved to @to does
3256                  * swap-in, the refcount of @to might be decreased to 0.
3257                  */
3258                 mem_cgroup_get(to);
3259                 if (need_fixup) {
3260                         if (!mem_cgroup_is_root(from))
3261                                 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3262                         mem_cgroup_put(from);
3263                         /*
3264                          * we charged both to->res and to->memsw, so we should
3265                          * uncharge to->res.
3266                          */
3267                         if (!mem_cgroup_is_root(to))
3268                                 res_counter_uncharge(&to->res, PAGE_SIZE);
3269                 }
3270                 return 0;
3271         }
3272         return -EINVAL;
3273 }
3274 #else
3275 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3276                 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3277 {
3278         return -EINVAL;
3279 }
3280 #endif
3281
3282 /*
3283  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3284  * page belongs to.
3285  */
3286 int mem_cgroup_prepare_migration(struct page *page,
3287         struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3288 {
3289         struct mem_cgroup *memcg = NULL;
3290         struct page_cgroup *pc;
3291         enum charge_type ctype;
3292         int ret = 0;
3293
3294         *ptr = NULL;
3295
3296         VM_BUG_ON(PageTransHuge(page));
3297         if (mem_cgroup_disabled())
3298                 return 0;
3299
3300         pc = lookup_page_cgroup(page);
3301         lock_page_cgroup(pc);
3302         if (PageCgroupUsed(pc)) {
3303                 memcg = pc->mem_cgroup;
3304                 css_get(&memcg->css);
3305                 /*
3306                  * At migrating an anonymous page, its mapcount goes down
3307                  * to 0 and uncharge() will be called. But, even if it's fully
3308                  * unmapped, migration may fail and this page has to be
3309                  * charged again. We set MIGRATION flag here and delay uncharge
3310                  * until end_migration() is called
3311                  *
3312                  * Corner Case Thinking
3313                  * A)
3314                  * When the old page was mapped as Anon and it's unmap-and-freed
3315                  * while migration was ongoing.
3316                  * If unmap finds the old page, uncharge() of it will be delayed
3317                  * until end_migration(). If unmap finds a new page, it's
3318                  * uncharged when it make mapcount to be 1->0. If unmap code
3319                  * finds swap_migration_entry, the new page will not be mapped
3320                  * and end_migration() will find it(mapcount==0).
3321                  *
3322                  * B)
3323                  * When the old page was mapped but migraion fails, the kernel
3324                  * remaps it. A charge for it is kept by MIGRATION flag even
3325                  * if mapcount goes down to 0. We can do remap successfully
3326                  * without charging it again.
3327                  *
3328                  * C)
3329                  * The "old" page is under lock_page() until the end of
3330                  * migration, so, the old page itself will not be swapped-out.
3331                  * If the new page is swapped out before end_migraton, our
3332                  * hook to usual swap-out path will catch the event.
3333                  */
3334                 if (PageAnon(page))
3335                         SetPageCgroupMigration(pc);
3336         }
3337         unlock_page_cgroup(pc);
3338         /*
3339          * If the page is not charged at this point,
3340          * we return here.
3341          */
3342         if (!memcg)
3343                 return 0;
3344
3345         *ptr = memcg;
3346         ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3347         css_put(&memcg->css);/* drop extra refcnt */
3348         if (ret || *ptr == NULL) {
3349                 if (PageAnon(page)) {
3350                         lock_page_cgroup(pc);
3351                         ClearPageCgroupMigration(pc);
3352                         unlock_page_cgroup(pc);
3353                         /*
3354                          * The old page may be fully unmapped while we kept it.
3355                          */
3356                         mem_cgroup_uncharge_page(page);
3357                 }
3358                 return -ENOMEM;
3359         }
3360         /*
3361          * We charge new page before it's used/mapped. So, even if unlock_page()
3362          * is called before end_migration, we can catch all events on this new
3363          * page. In the case new page is migrated but not remapped, new page's
3364          * mapcount will be finally 0 and we call uncharge in end_migration().
3365          */
3366         pc = lookup_page_cgroup(newpage);
3367         if (PageAnon(page))
3368                 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3369         else if (page_is_file_cache(page))
3370                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3371         else
3372                 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3373         __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3374         return ret;
3375 }
3376
3377 /* remove redundant charge if migration failed*/
3378 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3379         struct page *oldpage, struct page *newpage, bool migration_ok)
3380 {
3381         struct page *used, *unused;
3382         struct page_cgroup *pc;
3383
3384         if (!memcg)
3385                 return;
3386         /* blocks rmdir() */
3387         cgroup_exclude_rmdir(&memcg->css);
3388         if (!migration_ok) {
3389                 used = oldpage;
3390                 unused = newpage;
3391         } else {
3392                 used = newpage;
3393                 unused = oldpage;
3394         }
3395         /*
3396          * We disallowed uncharge of pages under migration because mapcount
3397          * of the page goes down to zero, temporarly.
3398          * Clear the flag and check the page should be charged.
3399          */
3400         pc = lookup_page_cgroup(oldpage);
3401         lock_page_cgroup(pc);
3402         ClearPageCgroupMigration(pc);
3403         unlock_page_cgroup(pc);
3404
3405         __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3406
3407         /*
3408          * If a page is a file cache, radix-tree replacement is very atomic
3409          * and we can skip this check. When it was an Anon page, its mapcount
3410          * goes down to 0. But because we added MIGRATION flage, it's not
3411          * uncharged yet. There are several case but page->mapcount check
3412          * and USED bit check in mem_cgroup_uncharge_page() will do enough
3413          * check. (see prepare_charge() also)
3414          */
3415         if (PageAnon(used))
3416                 mem_cgroup_uncharge_page(used);
3417         /*
3418          * At migration, we may charge account against cgroup which has no
3419          * tasks.
3420          * So, rmdir()->pre_destroy() can be called while we do this charge.
3421          * In that case, we need to call pre_destroy() again. check it here.
3422          */
3423         cgroup_release_and_wakeup_rmdir(&memcg->css);
3424 }
3425
3426 #ifdef CONFIG_DEBUG_VM
3427 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3428 {
3429         struct page_cgroup *pc;
3430
3431         pc = lookup_page_cgroup(page);
3432         if (likely(pc) && PageCgroupUsed(pc))
3433                 return pc;
3434         return NULL;
3435 }
3436
3437 bool mem_cgroup_bad_page_check(struct page *page)
3438 {
3439         if (mem_cgroup_disabled())
3440                 return false;
3441
3442         return lookup_page_cgroup_used(page) != NULL;
3443 }
3444
3445 void mem_cgroup_print_bad_page(struct page *page)
3446 {
3447         struct page_cgroup *pc;
3448
3449         pc = lookup_page_cgroup_used(page);
3450         if (pc) {
3451                 int ret = -1;
3452                 char *path;
3453
3454                 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3455                        pc, pc->flags, pc->mem_cgroup);
3456
3457                 path = kmalloc(PATH_MAX, GFP_KERNEL);
3458                 if (path) {
3459                         rcu_read_lock();
3460                         ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3461                                                         path, PATH_MAX);
3462                         rcu_read_unlock();
3463                 }
3464
3465                 printk(KERN_CONT "(%s)\n",
3466                                 (ret < 0) ? "cannot get the path" : path);
3467                 kfree(path);
3468         }
3469 }
3470 #endif
3471
3472 static DEFINE_MUTEX(set_limit_mutex);
3473
3474 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3475                                 unsigned long long val)
3476 {
3477         int retry_count;
3478         u64 memswlimit, memlimit;
3479         int ret = 0;
3480         int children = mem_cgroup_count_children(memcg);
3481         u64 curusage, oldusage;
3482         int enlarge;
3483
3484         /*
3485          * For keeping hierarchical_reclaim simple, how long we should retry
3486          * is depends on callers. We set our retry-count to be function
3487          * of # of children which we should visit in this loop.
3488          */
3489         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3490
3491         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3492
3493         enlarge = 0;
3494         while (retry_count) {
3495                 if (signal_pending(current)) {
3496                         ret = -EINTR;
3497                         break;
3498                 }
3499                 /*
3500                  * Rather than hide all in some function, I do this in
3501                  * open coded manner. You see what this really does.
3502                  * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3503                  */
3504                 mutex_lock(&set_limit_mutex);
3505                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3506                 if (memswlimit < val) {
3507                         ret = -EINVAL;
3508                         mutex_unlock(&set_limit_mutex);
3509                         break;
3510                 }
3511
3512                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3513                 if (memlimit < val)
3514                         enlarge = 1;
3515
3516                 ret = res_counter_set_limit(&memcg->res, val);
3517                 if (!ret) {
3518                         if (memswlimit == val)
3519                                 memcg->memsw_is_minimum = true;
3520                         else
3521                                 memcg->memsw_is_minimum = false;
3522                 }
3523                 mutex_unlock(&set_limit_mutex);
3524
3525                 if (!ret)
3526                         break;
3527
3528                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3529                                                 MEM_CGROUP_RECLAIM_SHRINK,
3530                                                 NULL);
3531                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3532                 /* Usage is reduced ? */
3533                 if (curusage >= oldusage)
3534                         retry_count--;
3535                 else
3536                         oldusage = curusage;
3537         }
3538         if (!ret && enlarge)
3539                 memcg_oom_recover(memcg);
3540
3541         return ret;
3542 }
3543
3544 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3545                                         unsigned long long val)
3546 {
3547         int retry_count;
3548         u64 memlimit, memswlimit, oldusage, curusage;
3549         int children = mem_cgroup_count_children(memcg);
3550         int ret = -EBUSY;
3551         int enlarge = 0;
3552
3553         /* see mem_cgroup_resize_res_limit */
3554         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3555         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3556         while (retry_count) {
3557                 if (signal_pending(current)) {
3558                         ret = -EINTR;
3559                         break;
3560                 }
3561                 /*
3562                  * Rather than hide all in some function, I do this in
3563                  * open coded manner. You see what this really does.
3564                  * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3565                  */
3566                 mutex_lock(&set_limit_mutex);
3567                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3568                 if (memlimit > val) {
3569                         ret = -EINVAL;
3570                         mutex_unlock(&set_limit_mutex);
3571                         break;
3572                 }
3573                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3574                 if (memswlimit < val)
3575                         enlarge = 1;
3576                 ret = res_counter_set_limit(&memcg->memsw, val);
3577                 if (!ret) {
3578                         if (memlimit == val)
3579                                 memcg->memsw_is_minimum = true;
3580                         else
3581                                 memcg->memsw_is_minimum = false;
3582                 }
3583                 mutex_unlock(&set_limit_mutex);
3584
3585                 if (!ret)
3586                         break;
3587
3588                 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3589                                                 MEM_CGROUP_RECLAIM_NOSWAP |
3590                                                 MEM_CGROUP_RECLAIM_SHRINK,
3591                                                 NULL);
3592                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3593                 /* Usage is reduced ? */
3594                 if (curusage >= oldusage)
3595                         retry_count--;
3596                 else
3597                         oldusage = curusage;
3598         }
3599         if (!ret && enlarge)
3600                 memcg_oom_recover(memcg);
3601         return ret;
3602 }
3603
3604 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3605                                             gfp_t gfp_mask,
3606                                             unsigned long *total_scanned)
3607 {
3608         unsigned long nr_reclaimed = 0;
3609         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3610         unsigned long reclaimed;
3611         int loop = 0;
3612         struct mem_cgroup_tree_per_zone *mctz;
3613         unsigned long long excess;
3614         unsigned long nr_scanned;
3615
3616         if (order > 0)
3617                 return 0;
3618
3619         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3620         /*
3621          * This loop can run a while, specially if mem_cgroup's continuously
3622          * keep exceeding their soft limit and putting the system under
3623          * pressure
3624          */
3625         do {
3626                 if (next_mz)
3627                         mz = next_mz;
3628                 else
3629                         mz = mem_cgroup_largest_soft_limit_node(mctz);
3630                 if (!mz)
3631                         break;
3632
3633                 nr_scanned = 0;
3634                 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3635                                                 gfp_mask,
3636                                                 MEM_CGROUP_RECLAIM_SOFT,
3637                                                 &nr_scanned);
3638                 nr_reclaimed += reclaimed;
3639                 *total_scanned += nr_scanned;
3640                 spin_lock(&mctz->lock);
3641
3642                 /*
3643                  * If we failed to reclaim anything from this memory cgroup
3644                  * it is time to move on to the next cgroup
3645                  */
3646                 next_mz = NULL;
3647                 if (!reclaimed) {
3648                         do {
3649                                 /*
3650                                  * Loop until we find yet another one.
3651                                  *
3652                                  * By the time we get the soft_limit lock
3653                                  * again, someone might have aded the
3654                                  * group back on the RB tree. Iterate to
3655                                  * make sure we get a different mem.
3656                                  * mem_cgroup_largest_soft_limit_node returns
3657                                  * NULL if no other cgroup is present on
3658                                  * the tree
3659                                  */
3660                                 next_mz =
3661                                 __mem_cgroup_largest_soft_limit_node(mctz);
3662                                 if (next_mz == mz)
3663                                         css_put(&next_mz->mem->css);
3664                                 else /* next_mz == NULL or other memcg */
3665                                         break;
3666                         } while (1);
3667                 }
3668                 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3669                 excess = res_counter_soft_limit_excess(&mz->mem->res);
3670                 /*
3671                  * One school of thought says that we should not add
3672                  * back the node to the tree if reclaim returns 0.
3673                  * But our reclaim could return 0, simply because due
3674                  * to priority we are exposing a smaller subset of
3675                  * memory to reclaim from. Consider this as a longer
3676                  * term TODO.
3677                  */
3678                 /* If excess == 0, no tree ops */
3679                 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3680                 spin_unlock(&mctz->lock);
3681                 css_put(&mz->mem->css);
3682                 loop++;
3683                 /*
3684                  * Could not reclaim anything and there are no more
3685                  * mem cgroups to try or we seem to be looping without
3686                  * reclaiming anything.
3687                  */
3688                 if (!nr_reclaimed &&
3689                         (next_mz == NULL ||
3690                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3691                         break;
3692         } while (!nr_reclaimed);
3693         if (next_mz)
3694                 css_put(&next_mz->mem->css);
3695         return nr_reclaimed;
3696 }
3697
3698 /*
3699  * This routine traverse page_cgroup in given list and drop them all.
3700  * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3701  */
3702 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3703                                 int node, int zid, enum lru_list lru)
3704 {
3705         struct zone *zone;
3706         struct mem_cgroup_per_zone *mz;
3707         struct page_cgroup *pc, *busy;
3708         unsigned long flags, loop;
3709         struct list_head *list;
3710         int ret = 0;
3711
3712         zone = &NODE_DATA(node)->node_zones[zid];
3713         mz = mem_cgroup_zoneinfo(memcg, node, zid);
3714         list = &mz->lists[lru];
3715
3716         loop = MEM_CGROUP_ZSTAT(mz, lru);
3717         /* give some margin against EBUSY etc...*/
3718         loop += 256;
3719         busy = NULL;
3720         while (loop--) {
3721                 struct page *page;
3722
3723                 ret = 0;
3724                 spin_lock_irqsave(&zone->lru_lock, flags);
3725                 if (list_empty(list)) {
3726                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3727                         break;
3728                 }
3729                 pc = list_entry(list->prev, struct page_cgroup, lru);
3730                 if (busy == pc) {
3731                         list_move(&pc->lru, list);
3732                         busy = NULL;
3733                         spin_unlock_irqrestore(&zone->lru_lock, flags);
3734                         continue;
3735                 }
3736                 spin_unlock_irqrestore(&zone->lru_lock, flags);
3737
3738                 page = lookup_cgroup_page(pc);
3739
3740                 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3741                 if (ret == -ENOMEM)
3742                         break;
3743
3744                 if (ret == -EBUSY || ret == -EINVAL) {
3745                         /* found lock contention or "pc" is obsolete. */
3746                         busy = pc;
3747                         cond_resched();
3748                 } else
3749                         busy = NULL;
3750         }
3751
3752         if (!ret && !list_empty(list))
3753                 return -EBUSY;
3754         return ret;
3755 }
3756
3757 /*
3758  * make mem_cgroup's charge to be 0 if there is no task.
3759  * This enables deleting this mem_cgroup.
3760  */
3761 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3762 {
3763         int ret;
3764         int node, zid, shrink;
3765         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3766         struct cgroup *cgrp = memcg->css.cgroup;
3767
3768         css_get(&memcg->css);
3769
3770         shrink = 0;
3771         /* should free all ? */
3772         if (free_all)
3773                 goto try_to_free;
3774 move_account:
3775         do {
3776                 ret = -EBUSY;
3777                 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3778                         goto out;
3779                 ret = -EINTR;
3780                 if (signal_pending(current))
3781                         goto out;
3782                 /* This is for making all *used* pages to be on LRU. */
3783                 lru_add_drain_all();
3784                 drain_all_stock_sync(memcg);
3785                 ret = 0;
3786                 mem_cgroup_start_move(memcg);
3787                 for_each_node_state(node, N_HIGH_MEMORY) {
3788                         for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3789                                 enum lru_list l;
3790                                 for_each_lru(l) {
3791                                         ret = mem_cgroup_force_empty_list(memcg,
3792                                                         node, zid, l);
3793                                         if (ret)
3794                                                 break;
3795                                 }
3796                         }
3797                         if (ret)
3798                                 break;
3799                 }
3800                 mem_cgroup_end_move(memcg);
3801                 memcg_oom_recover(memcg);
3802                 /* it seems parent cgroup doesn't have enough mem */
3803                 if (ret == -ENOMEM)
3804                         goto try_to_free;
3805                 cond_resched();
3806         /* "ret" should also be checked to ensure all lists are empty. */
3807         } while (memcg->res.usage > 0 || ret);
3808 out:
3809         css_put(&memcg->css);
3810         return ret;
3811
3812 try_to_free:
3813         /* returns EBUSY if there is a task or if we come here twice. */
3814         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3815                 ret = -EBUSY;
3816                 goto out;
3817         }
3818         /* we call try-to-free pages for make this cgroup empty */
3819         lru_add_drain_all();
3820         /* try to free all pages in this cgroup */
3821         shrink = 1;
3822         while (nr_retries && memcg->res.usage > 0) {
3823                 int progress;
3824
3825                 if (signal_pending(current)) {
3826                         ret = -EINTR;
3827                         goto out;
3828                 }
3829                 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3830                                                 false);
3831                 if (!progress) {
3832                         nr_retries--;
3833                         /* maybe some writeback is necessary */
3834                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3835                 }
3836
3837         }
3838         lru_add_drain();
3839         /* try move_account...there may be some *locked* pages. */
3840         goto move_account;
3841 }
3842
3843 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3844 {
3845         return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3846 }
3847
3848
3849 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3850 {
3851         return mem_cgroup_from_cont(cont)->use_hierarchy;
3852 }
3853
3854 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3855                                         u64 val)
3856 {
3857         int retval = 0;
3858         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3859         struct cgroup *parent = cont->parent;
3860         struct mem_cgroup *parent_memcg = NULL;
3861
3862         if (parent)
3863                 parent_memcg = mem_cgroup_from_cont(parent);
3864
3865         cgroup_lock();
3866         /*
3867          * If parent's use_hierarchy is set, we can't make any modifications
3868          * in the child subtrees. If it is unset, then the change can
3869          * occur, provided the current cgroup has no children.
3870          *
3871          * For the root cgroup, parent_mem is NULL, we allow value to be
3872          * set if there are no children.
3873          */
3874         if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3875                                 (val == 1 || val == 0)) {
3876                 if (list_empty(&cont->children))
3877                         memcg->use_hierarchy = val;
3878                 else
3879                         retval = -EBUSY;
3880         } else
3881                 retval = -EINVAL;
3882         cgroup_unlock();
3883
3884         return retval;
3885 }
3886
3887
3888 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3889                                                enum mem_cgroup_stat_index idx)
3890 {
3891         struct mem_cgroup *iter;
3892         long val = 0;
3893
3894         /* Per-cpu values can be negative, use a signed accumulator */
3895         for_each_mem_cgroup_tree(iter, memcg)
3896                 val += mem_cgroup_read_stat(iter, idx);
3897
3898         if (val < 0) /* race ? */
3899                 val = 0;
3900         return val;
3901 }
3902
3903 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3904 {
3905         u64 val;
3906
3907         if (!mem_cgroup_is_root(memcg)) {
3908                 if (!swap)
3909                         return res_counter_read_u64(&memcg->res, RES_USAGE);
3910                 else
3911                         return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3912         }
3913
3914         val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3915         val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3916
3917         if (swap)
3918                 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3919
3920         return val << PAGE_SHIFT;
3921 }
3922
3923 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3924 {
3925         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3926         u64 val;
3927         int type, name;
3928
3929         type = MEMFILE_TYPE(cft->private);
3930         name = MEMFILE_ATTR(cft->private);
3931         switch (type) {
3932         case _MEM:
3933                 if (name == RES_USAGE)
3934                         val = mem_cgroup_usage(memcg, false);
3935                 else
3936                         val = res_counter_read_u64(&memcg->res, name);
3937                 break;
3938         case _MEMSWAP:
3939                 if (name == RES_USAGE)
3940                         val = mem_cgroup_usage(memcg, true);
3941                 else
3942                         val = res_counter_read_u64(&memcg->memsw, name);
3943                 break;
3944         default:
3945                 BUG();
3946                 break;
3947         }
3948         return val;
3949 }
3950 /*
3951  * The user of this function is...
3952  * RES_LIMIT.
3953  */
3954 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3955                             const char *buffer)
3956 {
3957         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3958         int type, name;
3959         unsigned long long val;
3960         int ret;
3961
3962         type = MEMFILE_TYPE(cft->private);
3963         name = MEMFILE_ATTR(cft->private);
3964         switch (name) {
3965         case RES_LIMIT:
3966                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3967                         ret = -EINVAL;
3968                         break;
3969                 }
3970                 /* This function does all necessary parse...reuse it */
3971                 ret = res_counter_memparse_write_strategy(buffer, &val);
3972                 if (ret)
3973                         break;
3974                 if (type == _MEM)
3975                         ret = mem_cgroup_resize_limit(memcg, val);
3976                 else
3977                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
3978                 break;
3979         case RES_SOFT_LIMIT:
3980                 ret = res_counter_memparse_write_strategy(buffer, &val);
3981                 if (ret)
3982                         break;
3983                 /*
3984                  * For memsw, soft limits are hard to implement in terms
3985                  * of semantics, for now, we support soft limits for
3986                  * control without swap
3987                  */
3988                 if (type == _MEM)
3989                         ret = res_counter_set_soft_limit(&memcg->res, val);
3990                 else
3991                         ret = -EINVAL;
3992                 break;
3993         default:
3994                 ret = -EINVAL; /* should be BUG() ? */
3995                 break;
3996         }
3997         return ret;
3998 }
3999
4000 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4001                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4002 {
4003         struct cgroup *cgroup;
4004         unsigned long long min_limit, min_memsw_limit, tmp;
4005
4006         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4007         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4008         cgroup = memcg->css.cgroup;
4009         if (!memcg->use_hierarchy)
4010                 goto out;
4011
4012         while (cgroup->parent) {
4013                 cgroup = cgroup->parent;
4014                 memcg = mem_cgroup_from_cont(cgroup);
4015                 if (!memcg->use_hierarchy)
4016                         break;
4017                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4018                 min_limit = min(min_limit, tmp);
4019                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4020                 min_memsw_limit = min(min_memsw_limit, tmp);
4021         }
4022 out:
4023         *mem_limit = min_limit;
4024         *memsw_limit = min_memsw_limit;
4025         return;
4026 }
4027
4028 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4029 {
4030         struct mem_cgroup *memcg;
4031         int type, name;
4032
4033         memcg = mem_cgroup_from_cont(cont);
4034         type = MEMFILE_TYPE(event);
4035         name = MEMFILE_ATTR(event);
4036         switch (name) {
4037         case RES_MAX_USAGE:
4038                 if (type == _MEM)
4039                         res_counter_reset_max(&memcg->res);
4040                 else
4041                         res_counter_reset_max(&memcg->memsw);
4042                 break;
4043         case RES_FAILCNT:
4044                 if (type == _MEM)
4045                         res_counter_reset_failcnt(&memcg->res);
4046                 else
4047                         res_counter_reset_failcnt(&memcg->memsw);
4048                 break;
4049         }
4050
4051         return 0;
4052 }
4053
4054 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4055                                         struct cftype *cft)
4056 {
4057         return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4058 }
4059
4060 #ifdef CONFIG_MMU
4061 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4062                                         struct cftype *cft, u64 val)
4063 {
4064         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4065
4066         if (val >= (1 << NR_MOVE_TYPE))
4067                 return -EINVAL;
4068         /*
4069          * We check this value several times in both in can_attach() and
4070          * attach(), so we need cgroup lock to prevent this value from being
4071          * inconsistent.
4072          */
4073         cgroup_lock();
4074         memcg->move_charge_at_immigrate = val;
4075         cgroup_unlock();
4076
4077         return 0;
4078 }
4079 #else
4080 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4081                                         struct cftype *cft, u64 val)
4082 {
4083         return -ENOSYS;
4084 }
4085 #endif
4086
4087
4088 /* For read statistics */
4089 enum {
4090         MCS_CACHE,
4091         MCS_RSS,
4092         MCS_FILE_MAPPED,
4093         MCS_PGPGIN,
4094         MCS_PGPGOUT,
4095         MCS_SWAP,
4096         MCS_PGFAULT,
4097         MCS_PGMAJFAULT,
4098         MCS_INACTIVE_ANON,
4099         MCS_ACTIVE_ANON,
4100         MCS_INACTIVE_FILE,
4101         MCS_ACTIVE_FILE,
4102         MCS_UNEVICTABLE,
4103         NR_MCS_STAT,
4104 };
4105
4106 struct mcs_total_stat {
4107         s64 stat[NR_MCS_STAT];
4108 };
4109
4110 struct {
4111         char *local_name;
4112         char *total_name;
4113 } memcg_stat_strings[NR_MCS_STAT] = {
4114         {"cache", "total_cache"},
4115         {"rss", "total_rss"},
4116         {"mapped_file", "total_mapped_file"},
4117         {"pgpgin", "total_pgpgin"},
4118         {"pgpgout", "total_pgpgout"},
4119         {"swap", "total_swap"},
4120         {"pgfault", "total_pgfault"},
4121         {"pgmajfault", "total_pgmajfault"},
4122         {"inactive_anon", "total_inactive_anon"},
4123         {"active_anon", "total_active_anon"},
4124         {"inactive_file", "total_inactive_file"},
4125         {"active_file", "total_active_file"},
4126         {"unevictable", "total_unevictable"}
4127 };
4128
4129
4130 static void
4131 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4132 {
4133         s64 val;
4134
4135         /* per cpu stat */
4136         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4137         s->stat[MCS_CACHE] += val * PAGE_SIZE;
4138         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4139         s->stat[MCS_RSS] += val * PAGE_SIZE;
4140         val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4141         s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4142         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4143         s->stat[MCS_PGPGIN] += val;
4144         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4145         s->stat[MCS_PGPGOUT] += val;
4146         if (do_swap_account) {
4147                 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4148                 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4149         }
4150         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4151         s->stat[MCS_PGFAULT] += val;
4152         val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4153         s->stat[MCS_PGMAJFAULT] += val;
4154
4155         /* per zone stat */
4156         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4157         s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4158         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4159         s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4160         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4161         s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4162         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4163         s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4164         val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4165         s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4166 }
4167
4168 static void
4169 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4170 {
4171         struct mem_cgroup *iter;
4172
4173         for_each_mem_cgroup_tree(iter, memcg)
4174                 mem_cgroup_get_local_stat(iter, s);
4175 }
4176
4177 #ifdef CONFIG_NUMA
4178 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4179 {
4180         int nid;
4181         unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4182         unsigned long node_nr;
4183         struct cgroup *cont = m->private;
4184         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4185
4186         total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4187         seq_printf(m, "total=%lu", total_nr);
4188         for_each_node_state(nid, N_HIGH_MEMORY) {
4189                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4190                 seq_printf(m, " N%d=%lu", nid, node_nr);
4191         }
4192         seq_putc(m, '\n');
4193
4194         file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4195         seq_printf(m, "file=%lu", file_nr);
4196         for_each_node_state(nid, N_HIGH_MEMORY) {
4197                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4198                                 LRU_ALL_FILE);
4199                 seq_printf(m, " N%d=%lu", nid, node_nr);
4200         }
4201         seq_putc(m, '\n');
4202
4203         anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4204         seq_printf(m, "anon=%lu", anon_nr);
4205         for_each_node_state(nid, N_HIGH_MEMORY) {
4206                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4207                                 LRU_ALL_ANON);
4208                 seq_printf(m, " N%d=%lu", nid, node_nr);
4209         }
4210         seq_putc(m, '\n');
4211
4212         unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4213         seq_printf(m, "unevictable=%lu", unevictable_nr);
4214         for_each_node_state(nid, N_HIGH_MEMORY) {
4215                 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4216                                 BIT(LRU_UNEVICTABLE));
4217                 seq_printf(m, " N%d=%lu", nid, node_nr);
4218         }
4219         seq_putc(m, '\n');
4220         return 0;
4221 }
4222 #endif /* CONFIG_NUMA */
4223
4224 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4225                                  struct cgroup_map_cb *cb)
4226 {
4227         struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4228         struct mcs_total_stat mystat;
4229         int i;
4230
4231         memset(&mystat, 0, sizeof(mystat));
4232         mem_cgroup_get_local_stat(mem_cont, &mystat);
4233
4234
4235         for (i = 0; i < NR_MCS_STAT; i++) {
4236                 if (i == MCS_SWAP && !do_swap_account)
4237                         continue;
4238                 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4239         }
4240
4241         /* Hierarchical information */
4242         {
4243                 unsigned long long limit, memsw_limit;
4244                 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4245                 cb->fill(cb, "hierarchical_memory_limit", limit);
4246                 if (do_swap_account)
4247                         cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4248         }
4249
4250         memset(&mystat, 0, sizeof(mystat));
4251         mem_cgroup_get_total_stat(mem_cont, &mystat);
4252         for (i = 0; i < NR_MCS_STAT; i++) {
4253                 if (i == MCS_SWAP && !do_swap_account)
4254                         continue;
4255                 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4256         }
4257
4258 #ifdef CONFIG_DEBUG_VM
4259         {
4260                 int nid, zid;
4261                 struct mem_cgroup_per_zone *mz;
4262                 unsigned long recent_rotated[2] = {0, 0};
4263                 unsigned long recent_scanned[2] = {0, 0};
4264
4265                 for_each_online_node(nid)
4266                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4267                                 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4268
4269                                 recent_rotated[0] +=
4270                                         mz->reclaim_stat.recent_rotated[0];
4271                                 recent_rotated[1] +=
4272                                         mz->reclaim_stat.recent_rotated[1];
4273                                 recent_scanned[0] +=
4274                                         mz->reclaim_stat.recent_scanned[0];
4275                                 recent_scanned[1] +=
4276                                         mz->reclaim_stat.recent_scanned[1];
4277                         }
4278                 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4279                 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4280                 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4281                 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4282         }
4283 #endif
4284
4285         return 0;
4286 }
4287
4288 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4289 {
4290         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4291
4292         return mem_cgroup_swappiness(memcg);
4293 }
4294
4295 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4296                                        u64 val)
4297 {
4298         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4299         struct mem_cgroup *parent;
4300
4301         if (val > 100)
4302                 return -EINVAL;
4303
4304         if (cgrp->parent == NULL)
4305                 return -EINVAL;
4306
4307         parent = mem_cgroup_from_cont(cgrp->parent);
4308
4309         cgroup_lock();
4310
4311         /* If under hierarchy, only empty-root can set this value */
4312         if ((parent->use_hierarchy) ||
4313             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4314                 cgroup_unlock();
4315                 return -EINVAL;
4316         }
4317
4318         memcg->swappiness = val;
4319
4320         cgroup_unlock();
4321
4322         return 0;
4323 }
4324
4325 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4326 {
4327         struct mem_cgroup_threshold_ary *t;
4328         u64 usage;
4329         int i;
4330
4331         rcu_read_lock();
4332         if (!swap)
4333                 t = rcu_dereference(memcg->thresholds.primary);
4334         else
4335                 t = rcu_dereference(memcg->memsw_thresholds.primary);
4336
4337         if (!t)
4338                 goto unlock;
4339
4340         usage = mem_cgroup_usage(memcg, swap);
4341
4342         /*
4343          * current_threshold points to threshold just below usage.
4344          * If it's not true, a threshold was crossed after last
4345          * call of __mem_cgroup_threshold().
4346          */
4347         i = t->current_threshold;
4348
4349         /*
4350          * Iterate backward over array of thresholds starting from
4351          * current_threshold and check if a threshold is crossed.
4352          * If none of thresholds below usage is crossed, we read
4353          * only one element of the array here.
4354          */
4355         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4356                 eventfd_signal(t->entries[i].eventfd, 1);
4357
4358         /* i = current_threshold + 1 */
4359         i++;
4360
4361         /*
4362          * Iterate forward over array of thresholds starting from
4363          * current_threshold+1 and check if a threshold is crossed.
4364          * If none of thresholds above usage is crossed, we read
4365          * only one element of the array here.
4366          */
4367         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4368                 eventfd_signal(t->entries[i].eventfd, 1);
4369
4370         /* Update current_threshold */
4371         t->current_threshold = i - 1;
4372 unlock:
4373         rcu_read_unlock();
4374 }
4375
4376 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4377 {
4378         while (memcg) {
4379                 __mem_cgroup_threshold(memcg, false);
4380                 if (do_swap_account)
4381                         __mem_cgroup_threshold(memcg, true);
4382
4383                 memcg = parent_mem_cgroup(memcg);
4384         }
4385 }
4386
4387 static int compare_thresholds(const void *a, const void *b)
4388 {
4389         const struct mem_cgroup_threshold *_a = a;
4390         const struct mem_cgroup_threshold *_b = b;
4391
4392         return _a->threshold - _b->threshold;
4393 }
4394
4395 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4396 {
4397         struct mem_cgroup_eventfd_list *ev;
4398
4399         list_for_each_entry(ev, &memcg->oom_notify, list)
4400                 eventfd_signal(ev->eventfd, 1);
4401         return 0;
4402 }
4403
4404 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4405 {
4406         struct mem_cgroup *iter;
4407
4408         for_each_mem_cgroup_tree(iter, memcg)
4409                 mem_cgroup_oom_notify_cb(iter);
4410 }
4411
4412 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4413         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4414 {
4415         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4416         struct mem_cgroup_thresholds *thresholds;
4417         struct mem_cgroup_threshold_ary *new;
4418         int type = MEMFILE_TYPE(cft->private);
4419         u64 threshold, usage;
4420         int i, size, ret;
4421
4422         ret = res_counter_memparse_write_strategy(args, &threshold);
4423         if (ret)
4424                 return ret;
4425
4426         mutex_lock(&memcg->thresholds_lock);
4427
4428         if (type == _MEM)
4429                 thresholds = &memcg->thresholds;
4430         else if (type == _MEMSWAP)
4431                 thresholds = &memcg->memsw_thresholds;
4432         else
4433                 BUG();
4434
4435         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4436
4437         /* Check if a threshold crossed before adding a new one */
4438         if (thresholds->primary)
4439                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4440
4441         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4442
4443         /* Allocate memory for new array of thresholds */
4444         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4445                         GFP_KERNEL);
4446         if (!new) {
4447                 ret = -ENOMEM;
4448                 goto unlock;
4449         }
4450         new->size = size;
4451
4452         /* Copy thresholds (if any) to new array */
4453         if (thresholds->primary) {
4454                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4455                                 sizeof(struct mem_cgroup_threshold));
4456         }
4457
4458         /* Add new threshold */
4459         new->entries[size - 1].eventfd = eventfd;
4460         new->entries[size - 1].threshold = threshold;
4461
4462         /* Sort thresholds. Registering of new threshold isn't time-critical */
4463         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4464                         compare_thresholds, NULL);
4465
4466         /* Find current threshold */
4467         new->current_threshold = -1;
4468         for (i = 0; i < size; i++) {
4469                 if (new->entries[i].threshold < usage) {
4470                         /*
4471                          * new->current_threshold will not be used until
4472                          * rcu_assign_pointer(), so it's safe to increment
4473                          * it here.
4474                          */
4475                         ++new->current_threshold;
4476                 }
4477         }
4478
4479         /* Free old spare buffer and save old primary buffer as spare */
4480         kfree(thresholds->spare);
4481         thresholds->spare = thresholds->primary;
4482
4483         rcu_assign_pointer(thresholds->primary, new);
4484
4485         /* To be sure that nobody uses thresholds */
4486         synchronize_rcu();
4487
4488 unlock:
4489         mutex_unlock(&memcg->thresholds_lock);
4490
4491         return ret;
4492 }
4493
4494 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4495         struct cftype *cft, struct eventfd_ctx *eventfd)
4496 {
4497         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4498         struct mem_cgroup_thresholds *thresholds;
4499         struct mem_cgroup_threshold_ary *new;
4500         int type = MEMFILE_TYPE(cft->private);
4501         u64 usage;
4502         int i, j, size;
4503
4504         mutex_lock(&memcg->thresholds_lock);
4505         if (type == _MEM)
4506                 thresholds = &memcg->thresholds;
4507         else if (type == _MEMSWAP)
4508                 thresholds = &memcg->memsw_thresholds;
4509         else
4510                 BUG();
4511
4512         /*
4513          * Something went wrong if we trying to unregister a threshold
4514          * if we don't have thresholds
4515          */
4516         BUG_ON(!thresholds);
4517
4518         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4519
4520         /* Check if a threshold crossed before removing */
4521         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4522
4523         /* Calculate new number of threshold */
4524         size = 0;
4525         for (i = 0; i < thresholds->primary->size; i++) {
4526                 if (thresholds->primary->entries[i].eventfd != eventfd)
4527                         size++;
4528         }
4529
4530         new = thresholds->spare;
4531
4532         /* Set thresholds array to NULL if we don't have thresholds */
4533         if (!size) {
4534                 kfree(new);
4535                 new = NULL;
4536                 goto swap_buffers;
4537         }
4538
4539         new->size = size;
4540
4541         /* Copy thresholds and find current threshold */
4542         new->current_threshold = -1;
4543         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4544                 if (thresholds->primary->entries[i].eventfd == eventfd)
4545                         continue;
4546
4547                 new->entries[j] = thresholds->primary->entries[i];
4548                 if (new->entries[j].threshold < usage) {
4549                         /*
4550                          * new->current_threshold will not be used
4551                          * until rcu_assign_pointer(), so it's safe to increment
4552                          * it here.
4553                          */
4554                         ++new->current_threshold;
4555                 }
4556                 j++;
4557         }
4558
4559 swap_buffers:
4560         /* Swap primary and spare array */
4561         thresholds->spare = thresholds->primary;
4562         rcu_assign_pointer(thresholds->primary, new);
4563
4564         /* To be sure that nobody uses thresholds */
4565         synchronize_rcu();
4566
4567         mutex_unlock(&memcg->thresholds_lock);
4568 }
4569
4570 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4571         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4572 {
4573         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4574         struct mem_cgroup_eventfd_list *event;
4575         int type = MEMFILE_TYPE(cft->private);
4576
4577         BUG_ON(type != _OOM_TYPE);
4578         event = kmalloc(sizeof(*event), GFP_KERNEL);
4579         if (!event)
4580                 return -ENOMEM;
4581
4582         spin_lock(&memcg_oom_lock);
4583
4584         event->eventfd = eventfd;
4585         list_add(&event->list, &memcg->oom_notify);
4586
4587         /* already in OOM ? */
4588         if (atomic_read(&memcg->under_oom))
4589                 eventfd_signal(eventfd, 1);
4590         spin_unlock(&memcg_oom_lock);
4591
4592         return 0;
4593 }
4594
4595 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4596         struct cftype *cft, struct eventfd_ctx *eventfd)
4597 {
4598         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4599         struct mem_cgroup_eventfd_list *ev, *tmp;
4600         int type = MEMFILE_TYPE(cft->private);
4601
4602         BUG_ON(type != _OOM_TYPE);
4603
4604         spin_lock(&memcg_oom_lock);
4605
4606         list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4607                 if (ev->eventfd == eventfd) {
4608                         list_del(&ev->list);
4609                         kfree(ev);
4610                 }
4611         }
4612
4613         spin_unlock(&memcg_oom_lock);
4614 }
4615
4616 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4617         struct cftype *cft,  struct cgroup_map_cb *cb)
4618 {
4619         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4620
4621         cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4622
4623         if (atomic_read(&memcg->under_oom))
4624                 cb->fill(cb, "under_oom", 1);
4625         else
4626                 cb->fill(cb, "under_oom", 0);
4627         return 0;
4628 }
4629
4630 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4631         struct cftype *cft, u64 val)
4632 {
4633         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4634         struct mem_cgroup *parent;
4635
4636         /* cannot set to root cgroup and only 0 and 1 are allowed */
4637         if (!cgrp->parent || !((val == 0) || (val == 1)))
4638                 return -EINVAL;
4639
4640         parent = mem_cgroup_from_cont(cgrp->parent);
4641
4642         cgroup_lock();
4643         /* oom-kill-disable is a flag for subhierarchy. */
4644         if ((parent->use_hierarchy) ||
4645             (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4646                 cgroup_unlock();
4647                 return -EINVAL;
4648         }
4649         memcg->oom_kill_disable = val;
4650         if (!val)
4651                 memcg_oom_recover(memcg);
4652         cgroup_unlock();
4653         return 0;
4654 }
4655
4656 #ifdef CONFIG_NUMA
4657 static const struct file_operations mem_control_numa_stat_file_operations = {
4658         .read = seq_read,
4659         .llseek = seq_lseek,
4660         .release = single_release,
4661 };
4662
4663 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4664 {
4665         struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4666
4667         file->f_op = &mem_control_numa_stat_file_operations;
4668         return single_open(file, mem_control_numa_stat_show, cont);
4669 }
4670 #endif /* CONFIG_NUMA */
4671
4672 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4673 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4674 {
4675         /*
4676          * Part of this would be better living in a separate allocation
4677          * function, leaving us with just the cgroup tree population work.
4678          * We, however, depend on state such as network's proto_list that
4679          * is only initialized after cgroup creation. I found the less
4680          * cumbersome way to deal with it to defer it all to populate time
4681          */
4682         return mem_cgroup_sockets_init(cont, ss);
4683 };
4684
4685 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4686                                 struct cgroup *cont)
4687 {
4688         mem_cgroup_sockets_destroy(cont, ss);
4689 }
4690 #else
4691 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4692 {
4693         return 0;
4694 }
4695
4696 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4697                                 struct cgroup *cont)
4698 {
4699 }
4700 #endif
4701
4702 static struct cftype mem_cgroup_files[] = {
4703         {
4704                 .name = "usage_in_bytes",
4705                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4706                 .read_u64 = mem_cgroup_read,
4707                 .register_event = mem_cgroup_usage_register_event,
4708                 .unregister_event = mem_cgroup_usage_unregister_event,
4709         },
4710         {
4711                 .name = "max_usage_in_bytes",
4712                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4713                 .trigger = mem_cgroup_reset,
4714                 .read_u64 = mem_cgroup_read,
4715         },
4716         {
4717                 .name = "limit_in_bytes",
4718                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4719                 .write_string = mem_cgroup_write,
4720                 .read_u64 = mem_cgroup_read,
4721         },
4722         {
4723                 .name = "soft_limit_in_bytes",
4724                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4725                 .write_string = mem_cgroup_write,
4726                 .read_u64 = mem_cgroup_read,
4727         },
4728         {
4729                 .name = "failcnt",
4730                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4731                 .trigger = mem_cgroup_reset,
4732                 .read_u64 = mem_cgroup_read,
4733         },
4734         {
4735                 .name = "stat",
4736                 .read_map = mem_control_stat_show,
4737         },
4738         {
4739                 .name = "force_empty",
4740                 .trigger = mem_cgroup_force_empty_write,
4741         },
4742         {
4743                 .name = "use_hierarchy",
4744                 .write_u64 = mem_cgroup_hierarchy_write,
4745                 .read_u64 = mem_cgroup_hierarchy_read,
4746         },
4747         {
4748                 .name = "swappiness",
4749                 .read_u64 = mem_cgroup_swappiness_read,
4750                 .write_u64 = mem_cgroup_swappiness_write,
4751         },
4752         {
4753                 .name = "move_charge_at_immigrate",
4754                 .read_u64 = mem_cgroup_move_charge_read,
4755                 .write_u64 = mem_cgroup_move_charge_write,
4756         },
4757         {
4758                 .name = "oom_control",
4759                 .read_map = mem_cgroup_oom_control_read,
4760                 .write_u64 = mem_cgroup_oom_control_write,
4761                 .register_event = mem_cgroup_oom_register_event,
4762                 .unregister_event = mem_cgroup_oom_unregister_event,
4763                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4764         },
4765 #ifdef CONFIG_NUMA
4766         {
4767                 .name = "numa_stat",
4768                 .open = mem_control_numa_stat_open,
4769                 .mode = S_IRUGO,
4770         },
4771 #endif
4772 };
4773
4774 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4775 static struct cftype memsw_cgroup_files[] = {
4776         {
4777                 .name = "memsw.usage_in_bytes",
4778                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4779                 .read_u64 = mem_cgroup_read,
4780                 .register_event = mem_cgroup_usage_register_event,
4781                 .unregister_event = mem_cgroup_usage_unregister_event,
4782         },
4783         {
4784                 .name = "memsw.max_usage_in_bytes",
4785                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4786                 .trigger = mem_cgroup_reset,
4787                 .read_u64 = mem_cgroup_read,
4788         },
4789         {
4790                 .name = "memsw.limit_in_bytes",
4791                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4792                 .write_string = mem_cgroup_write,
4793                 .read_u64 = mem_cgroup_read,
4794         },
4795         {
4796                 .name = "memsw.failcnt",
4797                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4798                 .trigger = mem_cgroup_reset,
4799                 .read_u64 = mem_cgroup_read,
4800         },
4801 };
4802
4803 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4804 {
4805         if (!do_swap_account)
4806                 return 0;
4807         return cgroup_add_files(cont, ss, memsw_cgroup_files,
4808                                 ARRAY_SIZE(memsw_cgroup_files));
4809 };
4810 #else
4811 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4812 {
4813         return 0;
4814 }
4815 #endif
4816
4817 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4818 {
4819         struct mem_cgroup_per_node *pn;
4820         struct mem_cgroup_per_zone *mz;
4821         enum lru_list l;
4822         int zone, tmp = node;
4823         /*
4824          * This routine is called against possible nodes.
4825          * But it's BUG to call kmalloc() against offline node.
4826          *
4827          * TODO: this routine can waste much memory for nodes which will
4828          *       never be onlined. It's better to use memory hotplug callback
4829          *       function.
4830          */
4831         if (!node_state(node, N_NORMAL_MEMORY))
4832                 tmp = -1;
4833         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4834         if (!pn)
4835                 return 1;
4836
4837         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4838                 mz = &pn->zoneinfo[zone];
4839                 for_each_lru(l)
4840                         INIT_LIST_HEAD(&mz->lists[l]);
4841                 mz->usage_in_excess = 0;
4842                 mz->on_tree = false;
4843                 mz->mem = memcg;
4844         }
4845         memcg->info.nodeinfo[node] = pn;
4846         return 0;
4847 }
4848
4849 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4850 {
4851         kfree(memcg->info.nodeinfo[node]);
4852 }
4853
4854 static struct mem_cgroup *mem_cgroup_alloc(void)
4855 {
4856         struct mem_cgroup *mem;
4857         int size = sizeof(struct mem_cgroup);
4858
4859         /* Can be very big if MAX_NUMNODES is very big */
4860         if (size < PAGE_SIZE)
4861                 mem = kzalloc(size, GFP_KERNEL);
4862         else
4863                 mem = vzalloc(size);
4864
4865         if (!mem)
4866                 return NULL;
4867
4868         mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4869         if (!mem->stat)
4870                 goto out_free;
4871         spin_lock_init(&mem->pcp_counter_lock);
4872         return mem;
4873
4874 out_free:
4875         if (size < PAGE_SIZE)
4876                 kfree(mem);
4877         else
4878                 vfree(mem);
4879         return NULL;
4880 }
4881
4882 /*
4883  * At destroying mem_cgroup, references from swap_cgroup can remain.
4884  * (scanning all at force_empty is too costly...)
4885  *
4886  * Instead of clearing all references at force_empty, we remember
4887  * the number of reference from swap_cgroup and free mem_cgroup when
4888  * it goes down to 0.
4889  *
4890  * Removal of cgroup itself succeeds regardless of refs from swap.
4891  */
4892
4893 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4894 {
4895         int node;
4896
4897         mem_cgroup_remove_from_trees(memcg);
4898         free_css_id(&mem_cgroup_subsys, &memcg->css);
4899
4900         for_each_node_state(node, N_POSSIBLE)
4901                 free_mem_cgroup_per_zone_info(memcg, node);
4902
4903         free_percpu(memcg->stat);
4904         if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4905                 kfree(memcg);
4906         else
4907                 vfree(memcg);
4908 }
4909
4910 static void mem_cgroup_get(struct mem_cgroup *memcg)
4911 {
4912         atomic_inc(&memcg->refcnt);
4913 }
4914
4915 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4916 {
4917         if (atomic_sub_and_test(count, &memcg->refcnt)) {
4918                 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4919                 __mem_cgroup_free(memcg);
4920                 if (parent)
4921                         mem_cgroup_put(parent);
4922         }
4923 }
4924
4925 static void mem_cgroup_put(struct mem_cgroup *memcg)
4926 {
4927         __mem_cgroup_put(memcg, 1);
4928 }
4929
4930 /*
4931  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4932  */
4933 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4934 {
4935         if (!memcg->res.parent)
4936                 return NULL;
4937         return mem_cgroup_from_res_counter(memcg->res.parent, res);
4938 }
4939 EXPORT_SYMBOL(parent_mem_cgroup);
4940
4941 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4942 static void __init enable_swap_cgroup(void)
4943 {
4944         if (!mem_cgroup_disabled() && really_do_swap_account)
4945                 do_swap_account = 1;
4946 }
4947 #else
4948 static void __init enable_swap_cgroup(void)
4949 {
4950 }
4951 #endif
4952
4953 static int mem_cgroup_soft_limit_tree_init(void)
4954 {
4955         struct mem_cgroup_tree_per_node *rtpn;
4956         struct mem_cgroup_tree_per_zone *rtpz;
4957         int tmp, node, zone;
4958
4959         for_each_node_state(node, N_POSSIBLE) {
4960                 tmp = node;
4961                 if (!node_state(node, N_NORMAL_MEMORY))
4962                         tmp = -1;
4963                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4964                 if (!rtpn)
4965                         return 1;
4966
4967                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4968
4969                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4970                         rtpz = &rtpn->rb_tree_per_zone[zone];
4971                         rtpz->rb_root = RB_ROOT;
4972                         spin_lock_init(&rtpz->lock);
4973                 }
4974         }
4975         return 0;
4976 }
4977
4978 static struct cgroup_subsys_state * __ref
4979 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4980 {
4981         struct mem_cgroup *memcg, *parent;
4982         long error = -ENOMEM;
4983         int node;
4984
4985         memcg = mem_cgroup_alloc();
4986         if (!memcg)
4987                 return ERR_PTR(error);
4988
4989         for_each_node_state(node, N_POSSIBLE)
4990                 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4991                         goto free_out;
4992
4993         /* root ? */
4994         if (cont->parent == NULL) {
4995                 int cpu;
4996                 enable_swap_cgroup();
4997                 parent = NULL;
4998                 if (mem_cgroup_soft_limit_tree_init())
4999                         goto free_out;
5000                 root_mem_cgroup = memcg;
5001                 for_each_possible_cpu(cpu) {
5002                         struct memcg_stock_pcp *stock =
5003                                                 &per_cpu(memcg_stock, cpu);
5004                         INIT_WORK(&stock->work, drain_local_stock);
5005                 }
5006                 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5007         } else {
5008                 parent = mem_cgroup_from_cont(cont->parent);
5009                 memcg->use_hierarchy = parent->use_hierarchy;
5010                 memcg->oom_kill_disable = parent->oom_kill_disable;
5011         }
5012
5013         if (parent && parent->use_hierarchy) {
5014                 res_counter_init(&memcg->res, &parent->res);
5015                 res_counter_init(&memcg->memsw, &parent->memsw);
5016                 /*
5017                  * We increment refcnt of the parent to ensure that we can
5018                  * safely access it on res_counter_charge/uncharge.
5019                  * This refcnt will be decremented when freeing this
5020                  * mem_cgroup(see mem_cgroup_put).
5021                  */
5022                 mem_cgroup_get(parent);
5023         } else {
5024                 res_counter_init(&memcg->res, NULL);
5025                 res_counter_init(&memcg->memsw, NULL);
5026         }
5027         memcg->last_scanned_child = 0;
5028         memcg->last_scanned_node = MAX_NUMNODES;
5029         INIT_LIST_HEAD(&memcg->oom_notify);
5030
5031         if (parent)
5032                 memcg->swappiness = mem_cgroup_swappiness(parent);
5033         atomic_set(&memcg->refcnt, 1);
5034         memcg->move_charge_at_immigrate = 0;
5035         mutex_init(&memcg->thresholds_lock);
5036         return &memcg->css;
5037 free_out:
5038         __mem_cgroup_free(memcg);
5039         return ERR_PTR(error);
5040 }
5041
5042 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5043                                         struct cgroup *cont)
5044 {
5045         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5046
5047         return mem_cgroup_force_empty(memcg, false);
5048 }
5049
5050 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5051                                 struct cgroup *cont)
5052 {
5053         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5054
5055         kmem_cgroup_destroy(ss, cont);
5056
5057         mem_cgroup_put(memcg);
5058 }
5059
5060 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5061                                 struct cgroup *cont)
5062 {
5063         int ret;
5064
5065         ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5066                                 ARRAY_SIZE(mem_cgroup_files));
5067
5068         if (!ret)
5069                 ret = register_memsw_files(cont, ss);
5070
5071         if (!ret)
5072                 ret = register_kmem_files(cont, ss);
5073
5074         return ret;
5075 }
5076
5077 #ifdef CONFIG_MMU
5078 /* Handlers for move charge at task migration. */
5079 #define PRECHARGE_COUNT_AT_ONCE 256
5080 static int mem_cgroup_do_precharge(unsigned long count)
5081 {
5082         int ret = 0;
5083         int batch_count = PRECHARGE_COUNT_AT_ONCE;
5084         struct mem_cgroup *memcg = mc.to;
5085
5086         if (mem_cgroup_is_root(memcg)) {
5087                 mc.precharge += count;
5088                 /* we don't need css_get for root */
5089                 return ret;
5090         }
5091         /* try to charge at once */
5092         if (count > 1) {
5093                 struct res_counter *dummy;
5094                 /*
5095                  * "memcg" cannot be under rmdir() because we've already checked
5096                  * by cgroup_lock_live_cgroup() that it is not removed and we
5097                  * are still under the same cgroup_mutex. So we can postpone
5098                  * css_get().
5099                  */
5100                 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5101                         goto one_by_one;
5102                 if (do_swap_account && res_counter_charge(&memcg->memsw,
5103                                                 PAGE_SIZE * count, &dummy)) {
5104                         res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5105                         goto one_by_one;
5106                 }
5107                 mc.precharge += count;
5108                 return ret;
5109         }
5110 one_by_one:
5111         /* fall back to one by one charge */
5112         while (count--) {
5113                 if (signal_pending(current)) {
5114                         ret = -EINTR;
5115                         break;
5116                 }
5117                 if (!batch_count--) {
5118                         batch_count = PRECHARGE_COUNT_AT_ONCE;
5119                         cond_resched();
5120                 }
5121                 ret = __mem_cgroup_try_charge(NULL,
5122                                         GFP_KERNEL, 1, &memcg, false);
5123                 if (ret || !memcg)
5124                         /* mem_cgroup_clear_mc() will do uncharge later */
5125                         return -ENOMEM;
5126                 mc.precharge++;
5127         }
5128         return ret;
5129 }
5130
5131 /**
5132  * is_target_pte_for_mc - check a pte whether it is valid for move charge
5133  * @vma: the vma the pte to be checked belongs
5134  * @addr: the address corresponding to the pte to be checked
5135  * @ptent: the pte to be checked
5136  * @target: the pointer the target page or swap ent will be stored(can be NULL)
5137  *
5138  * Returns
5139  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
5140  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5141  *     move charge. if @target is not NULL, the page is stored in target->page
5142  *     with extra refcnt got(Callers should handle it).
5143  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5144  *     target for charge migration. if @target is not NULL, the entry is stored
5145  *     in target->ent.
5146  *
5147  * Called with pte lock held.
5148  */
5149 union mc_target {
5150         struct page     *page;
5151         swp_entry_t     ent;
5152 };
5153
5154 enum mc_target_type {
5155         MC_TARGET_NONE, /* not used */
5156         MC_TARGET_PAGE,
5157         MC_TARGET_SWAP,
5158 };
5159
5160 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5161                                                 unsigned long addr, pte_t ptent)
5162 {
5163         struct page *page = vm_normal_page(vma, addr, ptent);
5164
5165         if (!page || !page_mapped(page))
5166                 return NULL;
5167         if (PageAnon(page)) {
5168                 /* we don't move shared anon */
5169                 if (!move_anon() || page_mapcount(page) > 2)
5170                         return NULL;
5171         } else if (!move_file())
5172                 /* we ignore mapcount for file pages */
5173                 return NULL;
5174         if (!get_page_unless_zero(page))
5175                 return NULL;
5176
5177         return page;
5178 }
5179
5180 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5181                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5182 {
5183         int usage_count;
5184         struct page *page = NULL;
5185         swp_entry_t ent = pte_to_swp_entry(ptent);
5186
5187         if (!move_anon() || non_swap_entry(ent))
5188                 return NULL;
5189         usage_count = mem_cgroup_count_swap_user(ent, &page);
5190         if (usage_count > 1) { /* we don't move shared anon */
5191                 if (page)
5192                         put_page(page);
5193                 return NULL;
5194         }
5195         if (do_swap_account)
5196                 entry->val = ent.val;
5197
5198         return page;
5199 }
5200
5201 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5202                         unsigned long addr, pte_t ptent, swp_entry_t *entry)
5203 {
5204         struct page *page = NULL;
5205         struct inode *inode;
5206         struct address_space *mapping;
5207         pgoff_t pgoff;
5208
5209         if (!vma->vm_file) /* anonymous vma */
5210                 return NULL;
5211         if (!move_file())
5212                 return NULL;
5213
5214         inode = vma->vm_file->f_path.dentry->d_inode;
5215         mapping = vma->vm_file->f_mapping;
5216         if (pte_none(ptent))
5217                 pgoff = linear_page_index(vma, addr);
5218         else /* pte_file(ptent) is true */
5219                 pgoff = pte_to_pgoff(ptent);
5220
5221         /* page is moved even if it's not RSS of this task(page-faulted). */
5222         page = find_get_page(mapping, pgoff);
5223
5224 #ifdef CONFIG_SWAP
5225         /* shmem/tmpfs may report page out on swap: account for that too. */
5226         if (radix_tree_exceptional_entry(page)) {
5227                 swp_entry_t swap = radix_to_swp_entry(page);
5228                 if (do_swap_account)
5229                         *entry = swap;
5230                 page = find_get_page(&swapper_space, swap.val);
5231         }
5232 #endif
5233         return page;
5234 }
5235
5236 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5237                 unsigned long addr, pte_t ptent, union mc_target *target)
5238 {
5239         struct page *page = NULL;
5240         struct page_cgroup *pc;
5241         int ret = 0;
5242         swp_entry_t ent = { .val = 0 };
5243
5244         if (pte_present(ptent))
5245                 page = mc_handle_present_pte(vma, addr, ptent);
5246         else if (is_swap_pte(ptent))
5247                 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5248         else if (pte_none(ptent) || pte_file(ptent))
5249                 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5250
5251         if (!page && !ent.val)
5252                 return 0;
5253         if (page) {
5254                 pc = lookup_page_cgroup(page);
5255                 /*
5256                  * Do only loose check w/o page_cgroup lock.
5257                  * mem_cgroup_move_account() checks the pc is valid or not under
5258                  * the lock.
5259                  */
5260                 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5261                         ret = MC_TARGET_PAGE;
5262                         if (target)
5263                                 target->page = page;
5264                 }
5265                 if (!ret || !target)
5266                         put_page(page);
5267         }
5268         /* There is a swap entry and a page doesn't exist or isn't charged */
5269         if (ent.val && !ret &&
5270                         css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5271                 ret = MC_TARGET_SWAP;
5272                 if (target)
5273                         target->ent = ent;
5274         }
5275         return ret;
5276 }
5277
5278 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5279                                         unsigned long addr, unsigned long end,
5280                                         struct mm_walk *walk)
5281 {
5282         struct vm_area_struct *vma = walk->private;
5283         pte_t *pte;
5284         spinlock_t *ptl;
5285
5286         split_huge_page_pmd(walk->mm, pmd);
5287
5288         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5289         for (; addr != end; pte++, addr += PAGE_SIZE)
5290                 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5291                         mc.precharge++; /* increment precharge temporarily */
5292         pte_unmap_unlock(pte - 1, ptl);
5293         cond_resched();
5294
5295         return 0;
5296 }
5297
5298 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5299 {
5300         unsigned long precharge;
5301         struct vm_area_struct *vma;
5302
5303         down_read(&mm->mmap_sem);
5304         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5305                 struct mm_walk mem_cgroup_count_precharge_walk = {
5306                         .pmd_entry = mem_cgroup_count_precharge_pte_range,
5307                         .mm = mm,
5308                         .private = vma,
5309                 };
5310                 if (is_vm_hugetlb_page(vma))
5311                         continue;
5312                 walk_page_range(vma->vm_start, vma->vm_end,
5313                                         &mem_cgroup_count_precharge_walk);
5314         }
5315         up_read(&mm->mmap_sem);
5316
5317         precharge = mc.precharge;
5318         mc.precharge = 0;
5319
5320         return precharge;
5321 }
5322
5323 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5324 {
5325         unsigned long precharge = mem_cgroup_count_precharge(mm);
5326
5327         VM_BUG_ON(mc.moving_task);
5328         mc.moving_task = current;
5329         return mem_cgroup_do_precharge(precharge);
5330 }
5331
5332 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5333 static void __mem_cgroup_clear_mc(void)
5334 {
5335         struct mem_cgroup *from = mc.from;
5336         struct mem_cgroup *to = mc.to;
5337
5338         /* we must uncharge all the leftover precharges from mc.to */
5339         if (mc.precharge) {
5340                 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5341                 mc.precharge = 0;
5342         }
5343         /*
5344          * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5345          * we must uncharge here.
5346          */
5347         if (mc.moved_charge) {
5348                 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5349                 mc.moved_charge = 0;
5350         }
5351         /* we must fixup refcnts and charges */
5352         if (mc.moved_swap) {
5353                 /* uncharge swap account from the old cgroup */
5354                 if (!mem_cgroup_is_root(mc.from))
5355                         res_counter_uncharge(&mc.from->memsw,
5356                                                 PAGE_SIZE * mc.moved_swap);
5357                 __mem_cgroup_put(mc.from, mc.moved_swap);
5358
5359                 if (!mem_cgroup_is_root(mc.to)) {
5360                         /*
5361                          * we charged both to->res and to->memsw, so we should
5362                          * uncharge to->res.
5363                          */
5364                         res_counter_uncharge(&mc.to->res,
5365                                                 PAGE_SIZE * mc.moved_swap);
5366                 }
5367                 /* we've already done mem_cgroup_get(mc.to) */
5368                 mc.moved_swap = 0;
5369         }
5370         memcg_oom_recover(from);
5371         memcg_oom_recover(to);
5372         wake_up_all(&mc.waitq);
5373 }
5374
5375 static void mem_cgroup_clear_mc(void)
5376 {
5377         struct mem_cgroup *from = mc.from;
5378
5379         /*
5380          * we must clear moving_task before waking up waiters at the end of
5381          * task migration.
5382          */
5383         mc.moving_task = NULL;
5384         __mem_cgroup_clear_mc();
5385         spin_lock(&mc.lock);
5386         mc.from = NULL;
5387         mc.to = NULL;
5388         spin_unlock(&mc.lock);
5389         mem_cgroup_end_move(from);
5390 }
5391
5392 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5393                                 struct cgroup *cgroup,
5394                                 struct cgroup_taskset *tset)
5395 {
5396         struct task_struct *p = cgroup_taskset_first(tset);
5397         int ret = 0;
5398         struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5399
5400         if (memcg->move_charge_at_immigrate) {
5401                 struct mm_struct *mm;
5402                 struct mem_cgroup *from = mem_cgroup_from_task(p);
5403
5404                 VM_BUG_ON(from == memcg);
5405
5406                 mm = get_task_mm(p);
5407                 if (!mm)
5408                         return 0;
5409                 /* We move charges only when we move a owner of the mm */
5410                 if (mm->owner == p) {
5411                         VM_BUG_ON(mc.from);
5412                         VM_BUG_ON(mc.to);
5413                         VM_BUG_ON(mc.precharge);
5414                         VM_BUG_ON(mc.moved_charge);
5415                         VM_BUG_ON(mc.moved_swap);
5416                         mem_cgroup_start_move(from);
5417                         spin_lock(&mc.lock);
5418                         mc.from = from;
5419                         mc.to = memcg;
5420                         spin_unlock(&mc.lock);
5421                         /* We set mc.moving_task later */
5422
5423                         ret = mem_cgroup_precharge_mc(mm);
5424                         if (ret)
5425                                 mem_cgroup_clear_mc();
5426                 }
5427                 mmput(mm);
5428         }
5429         return ret;
5430 }
5431
5432 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5433                                 struct cgroup *cgroup,
5434                                 struct cgroup_taskset *tset)
5435 {
5436         mem_cgroup_clear_mc();
5437 }
5438
5439 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5440                                 unsigned long addr, unsigned long end,
5441                                 struct mm_walk *walk)
5442 {
5443         int ret = 0;
5444         struct vm_area_struct *vma = walk->private;
5445         pte_t *pte;
5446         spinlock_t *ptl;
5447
5448         split_huge_page_pmd(walk->mm, pmd);
5449 retry:
5450         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5451         for (; addr != end; addr += PAGE_SIZE) {
5452                 pte_t ptent = *(pte++);
5453                 union mc_target target;
5454                 int type;
5455                 struct page *page;
5456                 struct page_cgroup *pc;
5457                 swp_entry_t ent;
5458
5459                 if (!mc.precharge)
5460                         break;
5461
5462                 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5463                 switch (type) {
5464                 case MC_TARGET_PAGE:
5465                         page = target.page;
5466                         if (isolate_lru_page(page))
5467                                 goto put;
5468                         pc = lookup_page_cgroup(page);
5469                         if (!mem_cgroup_move_account(page, 1, pc,
5470                                                      mc.from, mc.to, false)) {
5471                                 mc.precharge--;
5472                                 /* we uncharge from mc.from later. */
5473                                 mc.moved_charge++;
5474                         }
5475                         putback_lru_page(page);
5476 put:                    /* is_target_pte_for_mc() gets the page */
5477                         put_page(page);
5478                         break;
5479                 case MC_TARGET_SWAP:
5480                         ent = target.ent;
5481                         if (!mem_cgroup_move_swap_account(ent,
5482                                                 mc.from, mc.to, false)) {
5483                                 mc.precharge--;
5484                                 /* we fixup refcnts and charges later. */
5485                                 mc.moved_swap++;
5486                         }
5487                         break;
5488                 default:
5489                         break;
5490                 }
5491         }
5492         pte_unmap_unlock(pte - 1, ptl);
5493         cond_resched();
5494
5495         if (addr != end) {
5496                 /*
5497                  * We have consumed all precharges we got in can_attach().
5498                  * We try charge one by one, but don't do any additional
5499                  * charges to mc.to if we have failed in charge once in attach()
5500                  * phase.
5501                  */
5502                 ret = mem_cgroup_do_precharge(1);
5503                 if (!ret)
5504                         goto retry;
5505         }
5506
5507         return ret;
5508 }
5509
5510 static void mem_cgroup_move_charge(struct mm_struct *mm)
5511 {
5512         struct vm_area_struct *vma;
5513
5514         lru_add_drain_all();
5515 retry:
5516         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5517                 /*
5518                  * Someone who are holding the mmap_sem might be waiting in
5519                  * waitq. So we cancel all extra charges, wake up all waiters,
5520                  * and retry. Because we cancel precharges, we might not be able
5521                  * to move enough charges, but moving charge is a best-effort
5522                  * feature anyway, so it wouldn't be a big problem.
5523                  */
5524                 __mem_cgroup_clear_mc();
5525                 cond_resched();
5526                 goto retry;
5527         }
5528         for (vma = mm->mmap; vma; vma = vma->vm_next) {
5529                 int ret;
5530                 struct mm_walk mem_cgroup_move_charge_walk = {
5531                         .pmd_entry = mem_cgroup_move_charge_pte_range,
5532                         .mm = mm,
5533                         .private = vma,
5534                 };
5535                 if (is_vm_hugetlb_page(vma))
5536                         continue;
5537                 ret = walk_page_range(vma->vm_start, vma->vm_end,
5538                                                 &mem_cgroup_move_charge_walk);
5539                 if (ret)
5540                         /*
5541                          * means we have consumed all precharges and failed in
5542                          * doing additional charge. Just abandon here.
5543                          */
5544                         break;
5545         }
5546         up_read(&mm->mmap_sem);
5547 }
5548
5549 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5550                                 struct cgroup *cont,
5551                                 struct cgroup_taskset *tset)
5552 {
5553         struct task_struct *p = cgroup_taskset_first(tset);
5554         struct mm_struct *mm = get_task_mm(p);
5555
5556         if (mm) {
5557                 if (mc.to)
5558                         mem_cgroup_move_charge(mm);
5559                 put_swap_token(mm);
5560                 mmput(mm);
5561         }
5562         if (mc.to)
5563                 mem_cgroup_clear_mc();
5564 }
5565 #else   /* !CONFIG_MMU */
5566 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5567                                 struct cgroup *cgroup,
5568                                 struct cgroup_taskset *tset)
5569 {
5570         return 0;
5571 }
5572 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5573                                 struct cgroup *cgroup,
5574                                 struct cgroup_taskset *tset)
5575 {
5576 }
5577 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5578                                 struct cgroup *cont,
5579                                 struct cgroup_taskset *tset)
5580 {
5581 }
5582 #endif
5583
5584 struct cgroup_subsys mem_cgroup_subsys = {
5585         .name = "memory",
5586         .subsys_id = mem_cgroup_subsys_id,
5587         .create = mem_cgroup_create,
5588         .pre_destroy = mem_cgroup_pre_destroy,
5589         .destroy = mem_cgroup_destroy,
5590         .populate = mem_cgroup_populate,
5591         .can_attach = mem_cgroup_can_attach,
5592         .cancel_attach = mem_cgroup_cancel_attach,
5593         .attach = mem_cgroup_move_task,
5594         .early_init = 0,
5595         .use_id = 1,
5596 };
5597
5598 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5599 static int __init enable_swap_account(char *s)
5600 {
5601         /* consider enabled if no parameter or 1 is given */
5602         if (!strcmp(s, "1"))
5603                 really_do_swap_account = 1;
5604         else if (!strcmp(s, "0"))
5605                 really_do_swap_account = 0;
5606         return 1;
5607 }
5608 __setup("swapaccount=", enable_swap_account);
5609
5610 #endif